🎧🍌 Inside Spellbook: Canada’s Fastest Growing AI Company | Scott Stevenson, Co-Founder & CEO

Why legal AI is so hot today, why fine-tuning models was mistake, building network effects in vertical AI, launching 100 products in 3 years, and top-down vs bottoms-up adoption in AI

Spellbook is an AI copilot for contract review and drafting. Essentially, “Cursor for lawyers.” They have 4,000 customers in 80 countries and are the fastest growing AI company in Canada.

It also might be the largest company in the world built on a Microsoft Word plugin.

Scott has been building in legal AI longer than almost anyone (since 2018). We talk about how legal software was untouched before LLM’s, why legal AI is so hot right now, if the hype is sustainable, how vertical AI tools should navigate product differentiation vs ChatGPT and Claude, and why Spellbook uses a bottoms up go-to-market motion when most AI legal software has gone top down.

We talk about why fine-tuning your own models was the biggest early mistake AI companies made, building a network effect as a vertical AI product, how $30 trillion per year flows through contracts, and Spellbook’s philosophy of “Don’t sharpen your axe when the chainsaw is coming out tomorrow”.

Spellbook spent a few years finding PMF before really taking off in 2022. Scott shares their playbook for launching over 100 product experiments in 3 years, how they knew when to lean in, scaling Spellbook post-PMF, and what he’s learned working with Keith Rabois after raising a $50m Series B from Khosla Ventures in 2025.

---

Support this Episode’s Sponsor

Flex: Sign-up for Flex Elite with code TURNER, get $1,000. Apply here.

To inquire about sponsoring future episodes, click here.

---

👉 Stream on Spotify and Apple

---

Timestamps to jump in:

• 0:30 Spellbook: “Cursor for Contracts”

• 3:08 Building the world’s largest Microsoft Word plugin

• 14:06 Why legal software was untouched before LLMs

• 18:32 $30 trillion moves through contracts annually

• 20:51 Why ChatGPT won’t replace vertical tools

• 25:15 Fine-tuning was the biggest mistake in AI

• 30:00 Differences between pro and amateur gamers

• 37:38 Top-down vs. bottoms-up in legal AI

• 42:27 The long-tail of legal AI software

• 47:24 Building for models that don’t exist yet

• 51:20 Skating where the puck is going

• 1:01:35 The legal bill that cost 50% of his bank account

• 1:09:33 Testing 100 landing pages in 3 years

• 1:14:06 The moment Spellbook hit PMF

• 1:19:17 Building new brands for each product experiment

• 1:23:10 Raising a Series B with a tweet

• 1:27:41 What Scott learned from Keith Rabois

• 1:31:16 Scott's favorite new AI tool

Referenced:

• Check out Spellbook

• Careers at Spellbook

• Playing to Win by David Sirlin

• Find the Fast Moving Water by James Currier at NFX

• Spellbook’s Case Study with Replit

• Twin (Scott’s favorite new AI tool)

Find Scott:

• X / Twitter

• LinkedIn

• Blog

---

Related Episode

---

👉 Stream on YouTube, Spotify, and Apple

Transcript

Find transcripts of all prior episodes here.

Turner Novak:

Scott, welcome to the show.

Scott Stevenson:

Thanks for having me, Turner. Great to be here.

Turner Novak:

Yeah, I’m excited. So I heard that you are the fastest growing AI company in Canada. Is this true?

Scott Stevenson:

We have been told this by a couple investors who have a very good, I would say, visibility of the Canadian market.

Turner Novak:

Interesting. Okay, so for people who don’t know Spellbook, because I feel like not a lot of people even heard of you before, so what do you guys do?

Scott Stevenson:

Yes, we’re basically Cursor for contracts, so an AI copilot for contract review and drafting. Yeah, we have 4,000 customers in 80 countries and we go very deep on this problem of commercial legal work. So if you are building a company or hiring employees, launching a coffee shop, anything you do in the world economically often is tied to a contract if it’s any substantial kind of transactions.

Turner Novak:

So this could be like signing a lease, hiring someone, doing a business deal of like, “We’ll pay you this and you’ll give me this much back for you to deliver this value to me in these products or services”?

Scott Stevenson:

Exactly, yeah. So we laser focus on that part of, I guess, the legal market. And we sell both to law firms and to in-house legal teams and in-house contract management teams and so on as well.

So our software will do things like catch mistakes or risks in contracts, help you standardize your contracts to, say, your company’s standards, help you draft more easily or doing something like a venture capital financing transaction. You could take a term sheet and then use our agent kind of like Claude Code to draft the 10 agreements you would need to do that transaction.

Turner Novak:

And you mentioned you have 4,000 customers. There’s a couple other big players, Harvey, Lagora. I think Harvey has like 1,000. Lagora has almost 1,000. So you have like double both of them combined.

Scott Stevenson:

Yeah. Yeah, we have quite a few customers. Yeah.

Turner Novak:

So then why has nobody really talked about Spellbook? What’s going on?

Scott Stevenson:

Well, people do talk about us. We’ve definitely taken a different approach to the market, and actually we were the first company in the world to bring a generative AI product to lawyers back in the summer of 2022, so it was a little before ChatGPT. I think we’ve had a little bit more of a heads down approach and we’ve had a bit more of a bottoms up approach in building our products. So rather than doing these big top down sales to like Am Law 100 law firms, we really sell bottom up to the lawyers and the contract managers who are using the software and kind of organically expand upwards from there. So we’re really focused on sort of like the end user versus just trying to get these very large top down deals pushed down to super large firms, yeah. So it’s a slower, I think, build of our customer base, but definitely compounding and snowballing.

Turner Novak:

Yeah. And the product is literally a Word plugin, like a Microsoft Word plugin. That’s essentially the product. I may be distilling this down, I’ll make it a little simpler. So how does it work exactly?

Scott Stevenson:

Yeah, so it’s a lot like Cursor or GitHub Copilot was our original inspiration. And the core of the product sits on top of Microsoft Word, which is where most lawyers are doing their drafting and reviewing work. The vast majority of contracts all go through Microsoft Word and we sit on top as sort of this intelligence layer. Now, we do have another separate desktop app as well that’s a little bit more something like Claude Code where it can do these kind of like complex multi-document projects, but the original core of the app is kind of based on top of where lawyers work.

And yeah, our idea of a great product for lawyers is that it should be like an electric bicycle. So lawyers know how to ride a bike already, they’re already drafting by hand. We want it to be an electric bike. So they’re still steering, they’re still pedaling. They’re in the same environment that they were before. It’s not like they got a cyber truck and now it’s like auto self-driving them around town or a plane. They’re still just driving their bike, but now they can get up over the hills a lot easier. And I think that’s like, I come from an engineering background and that’s what I liked about a lot of the coding tools is that I’m still very much in the driver’s seat, still in control, not completely doing something completely different than what I was before.

Turner Novak:

This episode is brought to you by Flex. It’s the AI native private bank for business owners. I use Flex personally and I love it because I use AI to underwrite the cashflow of your business, giving you a real credit line. The best part is 60 days of float, double the industry standard. Flex has all the features you’d expect from a modern financial platform, like unlimited cards, expense management, bill pay that syncs with your credit line, and their new consumer card, Flex Elite.

Flex Elite is a brand new Ramp-like experience for your personal life, a credit card with points, premium perks, concierge services, personal banking, cards and expense management for your family, net worth tracking across public and private assets, and a whole lot more fully integrated with your business spend. One card for your businesses, one card for your personal life, one card for everything. To skip the wait list, head to flex.one and use my code Turner to get an additional 100,000 points worth $1,000 after spending your first $10,000 with Flex Elite. That’s Flex.one and code Turner for $1,000 on your first $10,000 of spend. Thank you, Flex. And now let’s jump in.

So for somebody who is not a lawyer, and maybe this actually might be helpful for any lawyers listening, but what is an example of something you can do with AI here beyond ... I guess I’m thinking when you say Cursor or GitHub Copilot, I’m thinking I’m typing something and then it just fills out the line for me and it starts to write for me. Can you just kind of explain just how the product kind of works?

And then I think the Word plugin’s kind of an interesting dynamic where there’s not very many products that are Word plugins that have gotten ... You probably are the biggest Microsoft Word plugin ever.

Scott Stevenson:

That should be our claim to fame, like biggest Microsoft Word plugin ever. Yeah.

Turner Novak:

Oh, man. So then instead of the VC saying, “I don’t invest in ChatGPT wrappers.” You’re like, “A Word plugin.”

So I’m interested just like what are the things you do with it? And then how does a Word plugin work as a product?

Scott Stevenson:

Yeah, the first thing we had is what you mentioned, so like this sort of auto complete functionality where you start typing and it continues. And that was like what GitHub Copilot was. We do a lot more than that today.

The biggest thing that we do and the most popular thing is contract review. So you can take any contracts, say like a lease, a sales agreement, and you can instantly kind of review it for risks and issues and it will learn over time what you tend to flag and what you don’t so it gets better and better.

And I think a misconception people have about legal AI and contract reviews that there’s some right answer, but it’s actually contract review is completely subjective. It’s almost more like a YouTube recommendation algorithm of like, what do I think this lawyer is going to care about in this contract? So they can run it against a contract, get sort of a sorted list of things we think ... Changes to the contract that we think they’ll care about. Maybe they really care about payment terms, maybe they really care about data security and data privacy. We bubble those to the top and then we make suggested edits to the contract.

So a lawyer using the product can kind of go through all of these suggestions and accept or reject them. And it will automatically apply that to the contract with track changes on and everything. So it makes it really easy to do these.

Turner Novak:

This is like the red line thing. If you ever got in a legal doc, there’s always a redline version of it where you-

Scott Stevenson:

Yes. Yep.

Turner Novak:

People who don’t know, it’s basically the same document, but there is a second version that has a red line where everything got deleted and then like bolded things that were added. And it makes it really easy if you get something back, you can look at like the three changes or whatever, the red line version.

Scott Stevenson:

Exactly. Yeah. So yeah, that’s how lawyers operate, always with red lines or track changes turned on. So we do that. And then we have another version of that. So we’ve been growing very quickly in the enterprise in-house segments. So that’s been a huge, huge focus of us for the past year.

At the beginning of the last year, we had about almost no revenue from enterprise legal teams like eBay and Dropbox use us. And now it’s almost 60% of our revenue, so it’s growing very quickly. And what they love is our Playbooks feature. So Playbooks is like a review, except the legal team can set up a set of rules. Maybe they have 30, 40, 20 rules that dictate how they negotiate contracts, what they allow, what they don’t, what they’ll bend on. So if you’re a company reviewing thousands of contracts a year, thousands of NDAs, thousands of sales agreements, you can run them all through kind of your set of standards and your negotiation playbook and it will kind of automatically do that negotiation.

Turner Novak:

So these are almost like skills in Claude or something or like an artifact where you make your playbook basically.

Scott Stevenson:

Yeah. Yeah, yeah.

Turner Novak:

So what are some things a lawyer might do? What might be in a pretty standard playbook that someone might have?

Scott Stevenson:

It may be data residency. So if you’re a large company that really cares about your data security, maybe you mandate data residency needs to be in the US or Canada or something like that. So you could flag that on every sales agreement, make sure no one signs an agreement with data residency in another place or something like that. So that’s just one example.

There’s payment terms are really popular, making auto-renewals and all of these sorts of like commercial terms. Limitation of liability is definitely another big negotiated term that, depending on your negotiating power, you’re going to have different stances on what you will allow there and what you want. Yeah.

Turner Novak:

And how do you build a Word plugin? I’m super curious just like how that even works?

Scott Stevenson:

You got to go to Microsoft University. Yeah.

Turner Novak:

Yeah, so how do you actually build a Word plugin?

Scott Stevenson:

It’s pretty simple. It’s actually just a webpage. What actually is shown in the plugin is basically a web app that connects to the Word API to be able to do certain things. So it’s pretty straightforward.

Turner Novak:

Is it pretty simple to build? Would I have to go back and learn something or if I know like Java, TypeScript-

Scott Stevenson:

Yeah. Yeah, it’s based on JavaScript, so yeah, yeah.

Turner Novak:

Okay.

Scott Stevenson:

Yeah, a JavaScript, Typescript. So yeah, you can use that. Yeah, it’s pretty somewhat straightforward. The hardest part though is dealing with the actual manipulation of the Word document. And this is a file format that’s been around since the ‘90s at least. And there’s so many nuances, when you actually look under the hood, how these documents are represented. It’s incredibly complex. You can have like embedded software inside a contract. You can have embedded spreadsheets. There’s all sorts of weird hidden features, so-

Turner Novak:

People do that a lot?

Scott Stevenson:

No. But like every now and then there’s a lawyer or a law firm who just has this really weird file that they’ve kind of been adding onto since like 1997 and they throw it into Spellbook and some error will come up because there’s something in it that we’ve never seen before, but we’ve hammered ... We’ve been around since 2022, so we’ve hammered all those issues out.

And then the formatting is really nuanced. Lawyers really care about, is the formatting pristine? Are the sections labeled correctly? And it’s anyone who’s done complex formatting in Word knows it can be pretty challenging to deal with. So we’ve spent a lot of time on those. That’s the hardest part of integrating with Word.

Turner Novak:

Okay. Actually, my first job out of college, I worked in a bank as a credit analyst, we’re like lending money to businesses and we had to write a memo on each company like, “Here’s what they do, here’s their cashflow profile. Can they pay back a loan?” We did a collateral analysis. And all these things are pretty simple, but we did have actually embedded spreadsheets. Our template memo that you’re supposed to use was literally like a spreadsheet, like a cashflow model that was embedded into Word. And then we did the same thing with the collateral and it was just basically to make sure everyone just used the same standard. We’re all on the same page, which I always thought was like, it was super frustrating because if I ever had to do anything that was not standard, which is pretty much every time you have a separate spreadsheet and then you’re like figuring out how to get this thing into the Word memo and it was weird but-

Scott Stevenson:

There’s deep, deep features hidden in that format. Yeah. It’s almost like a programming language of its own, but yeah.

And then we do have what we call Spellbook Associate too. So as I mentioned, we have like a separate surface area that’s a little bit more of like a ChatGPT or Claude Code kind of shape, but really geared towards working on legal documents in Word. So it’s like using like Cursor’s agent or using Claude Code. And yeah, you can take like a term sheet, ask it to draft 10 other docs, or you could even throw in 1,000 docs for something like a data room review and have it build a table for you, extracting all the data, surfacing anything that’s concerning and so on.

Turner Novak:

Interesting. And I guess this kind of leads into some other stuff I want to talk about. Legal AI is probably one of the hotter areas of AI. There’s just a lot of momentum around it. It seems like it’s useful and like the adoption is there, but I’m just interested, as somebody who’s in it, so what is kind of going on right now?

Scott Stevenson:

Yeah, yeah. Yeah. Yeah, I think if you’re outside it, you might wonder like, yeah, is the hype real? It is probably one of the hottest verticals. Besides AI for coding, it’s maybe the hottest and most talked about vertical for AI right now. And I think there’s a reason why it has taken off so quickly.

One analogy I use is that large language models launching in like 2022 or with GPT-2 were kind of like the spreadsheet moment for lawyers. Accountants back in the ‘80s when spreadsheets were introduced, they started to be able to automate a lot of the work of like running a financial model. Before spreadsheets, it used to take like basically a building full of people to run a complex financial model. After spreadsheets and databases, we were able, using computers, to be able to automate a lot of the basic rudimentary math and formulas of kind of financial models.

And finance, maybe back in the ‘70s and ‘80s was run by like an army of people, humans. And now today, it’s maybe like actually 95% automated. If you think about the volume of transactions, how much bookkeeping is like semi-automated, software like Stripe and all of the tools we have to automate finance. We’ve automated a lot over many decades.

That has not happened in law or has not really started to happen until large language models in 2022. So that’s like, most verticals have seen decades of software adoption and automation, whereas law, basically up until 2022, still just ran 100% on an army of humans. The biggest advancements we had were like the word processor and email, and that made things go a little bit faster. But still, the core problem was software could not deal with unstructured text, so it couldn’t read unstructured text and it couldn’t write unstructured text. And that’s all lawyers do is they deal with 60-page documents of unstructured text and we just had no way to ingest it to understand it.

Turner Novak:

So you would just kind of have to just read it and/or you had 10 years of experience and just know these are the things I care about and I kind of know where to find them and it might only take me 20 minutes or something or five minutes, but it’s still-

Scott Stevenson:

Yeah. But you still have to read every word. If you’re reviewing a contract for a client, 60-page contract, you have to read every word basically. There could be something in there that you’re missing. So I think there’s just been radical inefficiency in the practice of law, and now it’s like the dam is breaking. There’s all this pent-up demand for legal efficiency, just like in every other vertical that has been able to adopt software. And now it’s finally able to be met because large language models are finally allowing us to actually help with the actual work that lawyers do.

Turner Novak:

So what was the software stack of a lawyer, I don’t know, five years ago, like pre-LLMs?

Scott Stevenson:

Word, Outlook. That’s pretty much it.

Turner Novak:

Did they have like a phone, like maybe Zoom or something?

Scott Stevenson:

Yeah, Zoom, maybe. I’d say five years ago, they’re still doing a lot of calls by phone and it’s like the phone bridges.

Turner Novak:

So basically, then they were doing all these, they were producing documents basically and they were reviewing and writing written documents with basically Word and email and then talking on the phone to communicate of what would be changed in the document essentially.

Scott Stevenson:

Yeah. Yeah.

Turner Novak:

Okay.

Scott Stevenson:

That was essentially the lawyer stack.

Turner Novak:

Okay.

Scott Stevenson:

It’s some deeply specialized software for things like entity management. If you have a complex entity structure with parent and children orgs, you might have like a chart of that.

Turner Novak:

Yeah, to like a design the org structure or something. Yeah.

Scott Stevenson:

But beyond that, especially for like commercial lawyers, which is where we’re focused, there really hasn’t ... E-signing or DocuSign. I forgot about that. DocuSign.

Turner Novak:

Oh, that’s fair.

Scott Stevenson:

Yeah.

Turner Novak:

Okay.

Scott Stevenson:

Yeah, yeah.

Turner Novak:

And I think, I don’t know if you mentioned it, you might have talked about it earlier when we got lunch, but there’s about 30 trillion in contracts that are signed per year.

Scott Stevenson:

Yeah, about $30 trillion moves through contracts every year.

Turner Novak:

So this is like economic value that is under a contractual agreement of some kind?

Scott Stevenson:

Exactly. Yeah, that’s right. Yeah. So there’s just this massive money flow moving through these contracts. And what inspired us to start the company is if you think about how inefficiently it’s happening, these contracts are probably taking 10 times longer than they should to be drafted and reviewed. And then things are still being missed because just imagine just being a human reviewing a 100-page contract and it’s like 8:00 PM, you have a deadline. It’s an almost hilariously impossible task to actually review 100 pages with a fine tooth comb on a tight timeline.

Turner Novak:

So did they not do it or did they do ... Well, how did they do this back before AI existed?

Scott Stevenson:

They would try their best. Yeah, yeah. Lawyers would try their best and contract managers would try their best.

One approach is standardization. Before AI came around, I think the hope was things would standardize more and more so you could kind of see, okay, here’s the standard SaaS agreement. YC has like SaaS agreement that’s pretty common that most startups use. And you can kind of do a diff between like, “Oh, what is different about this agreement compared to the YC standard agreement?” So that was one method, like shortcut you could use if there was a standard template. And with venture capital financing transactions, there’s a standard set of templates. You can easily see what’s changed. So that was one approach that would make these reviews easier, especially with complex transactions. But I think most of law has been surprisingly resistant to standardization because the deals people want to make are all kind of unique and bespoke, yeah. Yeah.

That actually connects to like where we started with Spellbook. So Spellbook was the second product we launched in, or one of the, or kind of the last product we launched in 2022. But we initially thought we were going to drive standardization with templates and that was kind of where we had started.

Turner Novak:

Yeah, I definitely wanted to ask you about that, but I think maybe while we’re still talking about just general kind of like AI, non-Spellbook specific stuff, when I think of a couple months ago, I made a contract, I just go to ChatGPT and just say like, “Make me a contract, make no mistakes,” et cetera. Couldn’t lawyers kind of do that? Is there some sort of thing where sort of the generic AI products trip up on legal stuff?

Scott Stevenson:

Yeah. Good question. So the first thing I would say is lawyers don’t actually draft anything from scratch for the most part, especially contracts because they want to start with a trusted precedent that they understand inside out. Because if they go to ChatGPT and ChatGPT outputs the whole contract, then they have to review every single word of that and make sure they understand it completely. That’s very difficult.

So ChatGPT is not great at modifying existing work or building on kind of your existing library. So we have a few features in Spellbook, like one, you can start with the precedents that you’re familiar with and we’ll kind of modify those. You can start with a sales agreement and be like, “Make this GDPR-compliant,” and we’ll kind of surgically make those edits for you. We also have a feature called Library where we can kind of have your whole history of all the deals you’ve ever worked on and use that to kind of influence the output of Spellbook as well. So these are some of the ...

One of the things I would say is working off of your existing corpus of docs as a lawyer is really, really important and ChatGPT doesn’t do that super well. But two, I think lawyers want things built into their existing workflow. I think like the chat interface is great, but it’s still like the terminal UI of AI. I don’t think chat is the be all and end all. And we just have a lot of unique user experiences that would just never fit inside the shape of ChatGPT.

For instance, one thing you can do in Spellbook is compare it to the market. So if you’re, say, signing a commercial lease in Manhattan, you can say, “Compare this to the average commercial lease in Manhattan and tell me what’s not normal.” And then you can actually dig into the data, into the charts and actually look at the data that we’ve collected in real time from millions of contracts and explore that through this visual interface that has nothing to do with chat. It’s very, very distanced from that. So I think there’s a huge number of experiences that people want that don’t fit in a chat box.

Turner Novak:

Yeah. So this is all within the Word interface?

Scott Stevenson:

A lot of this is in the Word interface. Some of it is in our Spellbook Associate product as well. Yeah.

Turner Novak:

Okay. And this market data, it’s basically you take everything that’s run on Spellbook of every customer and it’s anonymized and you can see what dates of comps or something like that? Or it’s like some kind of database of data where you can compare it to the contract or-

Scott Stevenson:

Yeah, so it’s like an opt-in model and most of our customers have opted in. And the way it works is, we take anonymous aggregate statistics. We only capture things like what’s the average, I don’t know, price per square foot in a commercial lease in Manhattan, what’s the average late payment interest rate for SaaS agreements? The only thing we end up capturing is these very high level statistical pieces of data, and that’s what gets exposed so it allows it to be really privacy-friendly, and yet it’s an alternative to the approach of fine-tuning. This idea of fine-tuning was really hyped for a while. I think every founder wanted to sell VCs on fine-tuning because it sounds very complex and defensible and you’re going to have this great moat, but it actually works pretty terribly for a whole bunch of reasons, and I can talk about that if you want.

Turner Novak:

Yeah. I think it’d be interesting just because that was the meme or the meta was if you are an AI company, you must build your own model because there’s no defensibility and you probably need to buy a bunch of GPUs and you need to train them all. And it’s like if you’re just a ChatGPT wrapper, it was like this derogatory slur basically to call someone a ChatGPT wrapper.

Scott Stevenson:

Yeah, so I think that was very wrong. I think this is an idea where there are narratives that founders learn investors are hungry for, and then they pitch them because it’s very legible and easy to understand for an investor. This idea of training models are like, “Oh, it’s going to be like OpenAI.” OpenAI trained their model, and it was expensive and cash was like a moat for, basically became a moat, but that did not really pan out in really many other areas for a bunch of reasons. You saw Bloomberg made Bloomberg GPT, that was one of the early ones and they spent I think millions training it, and then GPT-4 came out and just completely beat it at finance tasks so it was a waste.

Similar things have happened in legal AI where a number of companies have tried to train their own legal specific models and fine tune them. I do not know of a single one that is still in use today at any of the major application providers so it ended up being this big waste of time. I think the much better approach is to build value around the models. And I think there’s a lot of really great ways to do that. I think RAG is actually really, really good and actually superior to fine-tuning. Are you familiar with RAG, like retrieval augmented generation?

Turner Novak:

Yeah, it’s basically when you take the model plus just the internet or external sources essentially.

Scott Stevenson:

Yeah, exactly. The way I think of it is fine-tuning and training is like injecting things into the long-term memory of a model or almost putting into the evolutionary fiber of the model, giving evolutionary instincts as well. But if you’re asking a model to cite case law for a litigation case, you don’t want it looking at its long-term memory or its evolutionary instincts, you want it to actually look up the information and make it hard citation that you can actually cite. And it’s a much actually less hallucination prone method of getting legal specific data to work in these systems.

One, relying on RAG, which we did from the early days, you actually get citations that you can trust and inspect. Whereas if you’re fine-tuning models, you’re still going to hallucinate and you have no way to inspect the data. Two, when you use RAG, you can filter. We can filter. If you’re a lawyer in London, UK, and yeah, you work for a healthcare company, you can actually filter down the data to say, “I want to compare my contract to only other healthcare related contracts in the UK.” You can filter the sources down, whereas when you train a model, you end up with this one size fits all model. And a lot of lawyers will complain, “Well, ChatGPT is too biased towards the US,” or it’s too biased towards public company contracts because those are the only ones that are available to train on.

Which gets to my third point is no one wants to ingest private legal data into these proprietary models because there’s always a chance it could be spit out again. So by using our approach with the statistics, people are actually comfortable allowing us to ingest private data because it’s fully anonymized whereas people are not comfortable with training models on their private legal data because there’s the chance that it could be spit back out again, and that’s not something they can accept.

For those three reasons, I think RAG and what we’re doing with this real time data with market comparison in Spellbook is just much, much more superior or very superior to fine-tuning for the most part. I think it was this case where the herd just ran in completely the wrong direction. And I tweet about this all the time, things that are hyperlegible, the story just sounded right like data is the new oil and fine-tuning these models, cash as a moat, it sounded right. But the reality I think is just much more nuanced and complex.

And now we see a lot of these companies from 2022, 2023 who went did that fine-tuning approach, a lot of them are shutting down and they’re probably every two to three weeks, I’ll hear from one of these companies who’s now, they’re now looking to get acquired because the market’s matured so fast. They spent a lot of time doing this deeper R&D that they actually wasn’t that effective. And that’s very core to our culture at Spellbook is... I wrote a blog post about this back in 2022 when we launched and it’s called like, is GPT-3 too easy? We used the foundation models, and back when we launched, people would say, “Well, is this too easy? Where’s your team of machine learning engineers? Shouldn’t you be training your own models?”

And I cited this book, have you ever read or seen this, called Playing to Win? It’s by David Sirlin, he’s a professional Street Fighter player. Have you seen this before?

Turner Novak:

I have read your posts, but I’ve not seen the book.

Scott Stevenson:

Okay, it’s an amazing book where this guy professionally played Street Fighter at the highest level, and he talks about what is different about the mindset of a professional player versus what he calls a “scrub” or an intermediate or a bad player of the game. He said the scrub basically loses the game before it even starts because they have this totally wrong mindset, and I’m paraphrasing, but they basically have this romantic vision of the game that if they do play the game this super proper way, this romantic version of the game that they’ll come out on top in the long run, whereas the pros basically relentlessly exploit whatever they can to win, even if it looks cheap, even if it’s easy, they don’t do things because they look hard. If you’ve ever played Street Fighter, have you ever called someone cheap that you were playing against in one of these games?

Turner Novak:

I haven’t really played Street Fighter competitively, but it reminds me if you ever played Halo 2, you could do this thing called double shotting where you could basically take a shot, but you would shoot two bullets and if you-

Scott Stevenson:

Okay, I never learned that trick.

Turner Novak:

So all the best pro players in Halo 2, it’s like you could literally do twice the damage of one shot so everyone got good at double shotting. And if you were a purist and you’re like, “I’m not doing that,” you can’t beat people who do it.

Scott Stevenson:

Exactly, yeah. I think there’s this thing in AI where it’s like, people almost, I think engineers in particular who love complexity almost can’t accept how simple these systems can be to add value to customers. And there’s this attraction to complexity, fine-tuning, R&D. I think it’s starting to die out now finally and people are realizing, okay, building on top of foundation models is probably the best approach a lot of the time. But yeah, I think a lot of the herd ran in the wrong direction and it’s pretty fascinating.

Turner Novak:

Do you think part of it is this dynamic of what if OpenAI builds this or what if Anthropic builds this? Because if you’re not building your own model that has any kind of differentiation, they could just tweak ChatGPT to work better for lawyers, or something like that, is that a part of this? And how do you then navigate that as a founder of building a product that’s not in the strike zone?

Scott Stevenson:

Sure, yeah. I mean, I do think that that is a reason why this narrative took off, but I think the pendulum swung too far in the direction of differentiation rather than customer value, so you had so many companies and investors focused so much on how do we differentiate and doing really complex and hard things that weren’t very useful. And seriously, so many of these companies are shutting down and selling off now, but that’s the reason it happened, but how do you pragmatically deal with it? Yeah, I mean, that threat is there. I think the vertical AI providers have to work very hard every day to continue to add unique value to these customers.

And I tell our team, we have to be two years ahead at all times in terms of delivering state-of-the-art experiences to lawyers, two years ahead. We should be shipping things today that other competitors or other companies will be shipping in two years. I think you have to have this ruthlessly fast culture continuously adding unique value. I think the way you add the value is one through the data. We have real time data from millions of contracts that we can use to deliver better results to our customers. We also have preference data, so we learn from each of our customers what they care about, and ChatGPT and Claude are not really doing that for contracts specifically.

The data is super important, and the features are important, but then it’s like, how do you fit into the workflow of your customers? Lawyers are so busy, they have so much going on. If you don’t fit super neatly into the workflow, they’re just not going to use it. And the reality is Claude’s not in Word, it’s not designed out of the gate to give a lawyer value, and there’s a million little friction points because of that and I think if you... We always say our goal is to build a toaster, a toaster product. We are really good at doing one thing, toasting contracts I guess.

Turner Novak:

With a $30 trillion size, market size or whatever.

Scott Stevenson:

Exactly, yeah. If we can just do that one thing well, and if you optimize your product for that purpose, you just make so, so, so many decisions differently that would never make sense for ChatGPT to make. There’s so many little nuanced decisions that make that toasting experience very easy, simple, and effective for our customers.

Turner Novak:

And then there’s, because I feel like there’s the seven powers of just how a business has competitive advantage. I feel like we’ve maybe almost forgot about them, but when you were describing some of the different features, when I think of network effects as a pretty powerful business and it’s just the more customers you have, the more market data you have, the more useful that feature becomes. Then there’s maybe a point where you have so much data that one additional point doesn’t matter, but to get to that point, there is some strength to that, some positioning strength and then-

Scott Stevenson:

I mean, it scales more than you would think because it’s like, oh, what does it matter whether you have one million data points or two million data points? Well, it’s do you have data in Manhattan? Do you have data in London? Do you have data in SF? Do you have data in the healthcare industry? Do you have data in the aviation industry? Do you have data in manufacturing? Do you have data in energy? When you think of it that way, these are all industries that you have to conquer to deliver the best product to all of the lawyers in those industries.

It might sound like, well, again, what’s the difference between having a million data points and two million data points? Well, it’s less about that. It’s like, how many industries are you in? How much geography are you in? And do you have a statistically significant sample where you can provide useful insights? And yeah, I think it is legitimately a really great data network effect that we have.

Turner Novak:

Yeah. I feel like we almost forgot some of these rules for a while and just, I feel like almost economies of scale took over in the sense of the ability to train the models and having the capital on the balance sheet that you could utilize, which all this stuff is important, but the other stuff still matters too, I guess.

Scott Stevenson:

Yes. Yeah, I think so.

Turner Novak:

One thing you mentioned that we jumped past it, but I want to talk about, you mentioned this difference between top down and bottoms up sales cycles in legal AI. Can you just talk a little bit more about that and how that’s played out in the industry?

Scott Stevenson:

Yeah. I think there’s been a divergence of products built in legal AI. There’s products like Harvey and Legora which, great companies, we don’t actually encounter them that much because we’re so specialized and we service a different customer base, but they’ve had to optimize to sell to the innovation teams at the AmLaw 100 firms. Back with a previous product, we’ve done that kind of sales cycle before, and it’s very different because these innovation teams are going to push top down across a very large firm and mandate usage, and they’re usually going to bring you this long list of 50 things that they need in order to move ahead and it’s a decision by committee sort of thing.

We had an early experience before we hit PMF with Spellbook of working with these committees, and they send you in very strange directions and ones that I think are maybe not best for the product. For example, at a lot of these large law firms, they operate on an hourly billing model and just decreasing all their billable hours is not a positive incentive. The incentive structure is really misaligned with AI.

Turner Novak:

So you don’t want them to get more work done almost?

Scott Stevenson:

Yeah. I mean, lawyers make a lot of money, especially in these large firms from billable hours. And so what you found was a lot of the time what the committee’s most concerned about is how do we advertise this to our clients? How do we do a press release to show that we’re innovative? How do we just constantly shove into our client’s face that we’re an innovative law firm so that we don’t look bad compared to the firm across the street? And we noticed that that was happening and the committees would ask for things like client portals. Well, we want our clients to be able to log in and see the innovation firsthand and-

Turner Novak:

What is that? What is a client portal?

Scott Stevenson:

It’s a place where a client of a law firm can log in and interact with the lawyer or do their work there. We actually built one of these in an earlier product we had called Rally because we also had this request and clients hated it. They’re like, “Why can’t I talk to my lawyer in email? I don’t want another login to another website.”

Turner Novak:

So is this thing tracking what each piece of work that’s done and it automatically puts it in? I can log in and see what you did or something like that?

Scott Stevenson:

Yeah. It’s like you can log in and see the documents together or collaborate on the documents together. And then the lawyers didn’t really like it when we launched it because they didn’t want to the client to see all their messy, how the sausage was made kind of stuff. I’ve seen this feature request a ton, a lot from the really large firms with the AmLaw 100 firms that want to show the innovation to clients, but I’ve mostly just seen it end up failing.

And so with that experience, we decided to take a really different approach where we’re going to sell bottom up to the actual end users, the lawyers. And the vast majority of our customers, we don’t even do a sales call or a demo call. It’s like, “Welcome aboard, Turner. You are going to use Spellbook today, and in five minutes, you’re going to be set up and actually using Spellbook for some real work or demo work and clicking around and getting value from it.”

And so that evolutionary pressure has enabled us to, I think, build a very different product that is much more Cursor-like in how it’s baked into the user’s existing workflow. It’s not this grand design thing that you’re rolling out across a massive firm, it’s like a really practical tool that is always within arm’s reach, that’s a win at the back of the lawyer. And because of that, we have amazing retention metrics like our net revenue retention of 130% plus. We’re doing more of a land and expand motion rather than top down, but we have a lot of customers and they’re growing their usage, expanding their seats. And yeah, we really like that way of building a product because it subjects you to different influences, the influences of the actual end user who’s going to be using this thing to get work done.

Turner Novak:

There is quite a few different legal AI software products that have gotten a lot of revenue. People are using it, whatever. What is the market leaders in some of these different legal AI categories look like? Because I think we talked about too, Harvey and Legora, but I think there’s like a lot more. I don’t know. Is there an easy way to educate people for a couple minutes on what’s working in all these different subcategories of legal?

Scott Stevenson:

So yeah, Harvey and Legora, fairly similar, started with the law firms and the AmLaw 100 types of customers, what I’ve been talking about.

Turner Novak:

What is their product? What do you use when you’re using-

Scott Stevenson:

I would say it’s a very broad platform that’s broadly like ChatGPT for law. They have a number of different things they do, but it’s quite broad because they’re rolling it out to a whole legal team that might include litigation teams and transactional teams and so on so it’s like if you imagine tuning ChatGPT or Claude for a legal use case.

Turner Novak:

Is it like a whole operating system to run your law firm on?

Scott Stevenson:

The work, yeah. I think that’s more what they’re building is this all encompassing operating system kind of thing for a firm. But very tuned towards the law firms, whereas we’ve had really amazing product market fit with the in-house legal teams who don’t care about the billable hour who... This is the other type of customers is the in-house legal team. They’re starting to sell to that customer base too, but it’s very different because they don’t care about the billable hour, they don’t care about showing the clients, their clients the legal innovation they’re doing. They really just want tools that they can switch on and deal with this hair on fire problem of, “I have too many contracts to deal with. I need to clear my queue. I need some way out.” And so that’s where we’ve really been shining is in that segment. In terms of other companies-

Turner Novak:

It’s probably personal litigation type of stuff?

Scott Stevenson:

Oh yeah, there’s litigation, there’s like EvenUp’s done really well for instance.

Turner Novak:

That’s personal litigation?

Scott Stevenson:

Personal injury.

Turner Novak:

Oh, personal injury. Okay. There’s that one, EvenUp, people that know EvenUp listening to this will, though they can write in the comments what EvenUp does. I’ve definitely heard of that one before.

Scott Stevenson:

Yeah. EvenUp is AI for personal injury cases.

Turner Novak:

Okay. And then they’re corporate advisory type stuff. Isn’t there a company called, it’s called Hebbia?

Scott Stevenson:

Oh, yeah. Hebbia was pretty broad at first, but from what I see, they’re really trying to take the position of being for finance now.

Turner Novak:

Oh, interesting. Yeah.

Scott Stevenson:

They’ve gone very deep down that angle. I don’t know if they’re still running their legal arm anymore.

Turner Novak:

And is there a couple others or there’s maybe a longer tail?

Scott Stevenson:

There is a very, very long tail. I would say other smaller companies that have launched. Sandstone is one that does in-house enterprise legal that they’ve launched pretty recently. And then there’s a really long tail of other startups doing similar things that I think a lot... This vertical has matured. This AI vertical has matured so, so fast that it’s been, I think, really, really hard for this long tail of companies to catch up to the point where it’s almost like every three weeks now, one of these small companies comes to us looking for maybe an acquisition or something like that. They’ve built decent customer bases, but it’s shocking how fast this vertical has moved.

Turner Novak:

Oh, so have you guys done any acquisitions or exploring some or?

Scott Stevenson:

We are actively looking at two now, and yeah, part of our strategy this year is definitely to roll up some of these smaller companies that couldn’t quite get a foothold, the market moved a little bit too fast. They’ve built maybe something similar or more lightweight. They have a little bit of a customer base. We look at the math and it’s like, we could spend money on Google Ads or we could just acquire a bunch of these small companies so I think there is consolidation happening for sure. Legal AI has been very hyped.

Turner Novak:

That’s fascinating because you think there would probably not be as much consolidation this early into a hyper growth market, but it’s probably that there’s... It’s just these massive fluctuations in product capabilities, adoption.

Scott Stevenson:

Yeah. I mean, it’s so fast. I mean, the speed you have to move to keep up with the market is really, really fast.

Turner Novak:

Have you guys found, were there certain times where the models would maybe like OpenAI or Anthropic would release a new model and just suddenly Spellbook worked so much better? I know a lot of people have had those.

Scott Stevenson:

Yeah, I mean, that definitely has happened. We started building Spellbook on GPT-2, so that was tough. GPT-3 was a little bit better. And yeah, it was funny when ChatGPT came out, everyone was like, “Oh my God, what’s going to happen to Spellbook?”

Turner Novak:

Oh, really?

Scott Stevenson:

This was 2022, everyone was worried about it. And when ChatGPT came out, our growth just exploded because the models started getting better that we were using, and lawyers were getting their feet wet in these generalized AI experiences and then searching on Google, “I want ChatGPT for lawyers.” That’s literally what they search and then they would find Spellbooks.

So every time the generalized models and platforms have launched new things or gotten better, it’s generally been very good for us, both from a capabilities’ perspective. And in terms of just getting lawyers interested in AI enough to look a step deeper. Yeah. One mantra we have at Spellbook is that I think a lot of other companies have maybe gotten wrong, and I think it’s important for everyone to think about when they’re adopting AI is we say. “It’s time to chop down trees with a blunt ax,” There’s the Abe Lincoln quote, “If I had six hours to chop down a tree, I’d spend five hours sharpening the ax.”

And I think a lot of engineers and a lot of knowledge workers, we’re used to the idea of mastering a tool and then getting dividends from that mastery. But the reality in AI now is there’s no time to master anything. Every six months a new tool or a new model comes out and we’ve had to teach our engineers like, “Look, we can’t sit around optimizing around GPT-3 because GPT-4 is going to come out in six months. And we can’t try to master this thing. No one is going to have time to master this thing.” So a really important part of our culture is there’s no point in sharpening the ax when the chainsaw is coming out tomorrow. And so what we teach our team to do is drop the ax, pick up the chainsaw, stop and keep moving on, marching forward, implementing new models, new techniques, delete old code very quickly when it’s not needed.

And as an engineer, I think it’s like an un-intuitive culture. I think one of our advantages is also just the culture that we’ve been building these products since 2022. And our team has kind of learned how to do this, which I think the natural instinct of a lot of experienced engineers is kind of in the opposite direction. I’m going to build a really complex, robust system around this model, but then the next model comes out in six months and then you just have to delete all that code. So yeah, it’s an interesting time to build software.

Turner Novak:

Yeah. Because I mean, I guess isn’t there this risk though that the models don’t get better? The chainsaw doesn’t come out, right?

Scott Stevenson:

Yeah.

Turner Novak:

And then it’s like your ax isn’t sharp and you’re screwed. I mean, I guess they can kind of go both ways.

Scott Stevenson:

Yeah. I mean, there is that risk, but yeah, I mean, I don’t think we’re there yet. I don’t think we’re seeing this sort of plateau yet.

Turner Novak:

That’s fair. And is there a reason that you have that specific viewpoint? What are you seeing to make you so confident? And then maybe how does that relate into how the AI software is going to change? When you look five years in the future, is there still just so much more room to use current capabilities to make the products in the future so much better or?

Scott Stevenson:

Yeah. I think you just look at the trajectory. It’s like, I’m Canadian, you’re skating where the puck is going. And the thing is, the puck, if you just draw a trendline, the puck is moving way faster than it ever has before. We’ve never seen technology advance at this pace in our lifetimes. So a lot of people, and just trying to do the math of skating where the puck is going, they’re thinking about the puck speed that you might’ve had in 2015, but it’s actually this really accelerated speed. And I think it’s just like the math of what angle do you want to go at, how ambitious do you want to be? If you are looking at the trend of where things are going and you point your angle to meet the puck at the right place, you’re going to be a lot more ambitious. And we’ve done that again and again.

What we do is, when we start building a feature or product, we aim to build things that are not doable today. And that’s like a really, really important feature that is a hard thing to tell your engineers. It’s like, “Your goal is to start building something now that will not work today. It will work in six months when the models get better.” That’s how you actually time the building of these features. Because if you build something that’s achievable today, it’s not going to be that impressive in six months.

Turner Novak:

How do you know what’s like an okay degree of like not quite yet possible, but will be possible soon?

Scott Stevenson:

Yeah. I mean, it’s intuition. It’s like shooting a basketball or something. You get a feel for just watching the technology. I think being plugged into X is really good for just sensing the velocity. I think X is generally like two years ahead of like LinkedIn on this stuff. So if you’re plugged in there, you’re going to be seeing what researchers are talking about. You’re going to be seeing what engineers are hacking on. And if you understand how the tech works, you’ll see that there’s a sequence of advancements that will inevitably be made that are going to make things easier.

And I think a lot of the advancements now are not even at the model level necessarily, but it’s just in terms of figuring out the right techniques for like, how do you schedule an agent to operate in the background rather than needing to be prompted? What techniques do you use for planning and how do you implement planning for long range tasks? These are things that are rapidly being iterated on and you can pull those into your software very easily.

Turner Novak:

Are those things you guys have thought about at Spellbook?

Scott Stevenson:

Yeah. A lot. Yeah, quite a lot.

Turner Novak:

Is it there yet? Is it in the product right now?

Scott Stevenson:

Like planning for long range tasks or like the scheduling of?

Turner Novak:

Yeah, like agents doing stuff.

Scott Stevenson:

Agents doing stuff. Yeah. I mean, so we definitely have agents that can do very long-running like drafting tasks today.

Turner Novak:

So what’s an example of that?

Scott Stevenson:

Yeah, one example would be like the financing transaction, like VC financing transaction, example I gave. So you have a term sheet and you need to draft a full set of NVCA docs. That could be like 1,000 edits. It’s actually quite detailed. I don’t know if you’ve seen like the full template set, but like before a lawyer has edited it, but it’s like there’s tons of optional language, there’s math you have to calculate. It’s a very deep problem. And so that’s something that we can do quite accurately and as like kind of a long range task. And it’s not just like filling in the blanks. It’s like you’re cross referencing these documents, making sure they’re consistent, doing math, making sure that the math adds up, like you’re calculating share prices and things like that. So that’s kind of I think where the state of the art is.

But that product, so Spellbook Associate is our agent product, we started working on that, like I said, before it was possible. We launched the first version of that product in Alpha almost two years ago now. So it was the first long-running agent for multi-document legal work ever launched and it didn’t really work when we launched it, but then the models got better and better and we got feedback and then today it works really, really well. And then the next thing that we’re really excited about is like agents working in the background. I think when people think about like, is AI overhyped or not, the biggest thing on my mind is the way most people use AI today is you put in a prompt and it works for like five minutes and then you get an answer back.

Turner Novak:

It’s just kind of like better Google.

Scott Stevenson:

Yeah, better Google. But the thing I think about is imagine having an employee like that. You go to the employee, you ask them a question, they work for five minutes and they give you an answer and then they do nothing. That’s like the worst employee ever. That’s what we have today. It’s like the worst employee ever, who if you go over their shoulder and ask them a question, they’ll work hard and give you something, but after that they do nothing. They just kind of sit there and there’s such an easy gap for us to jump to say, “Well, how do we make these agents work in the background all the time, like an actual employee pushing the boulder forward without us?” I think that is going to be like a 10X for AI and agents.

So I don’t think people understand how impactful this technology is going to be because when you have like a AI coworker in your Slack who’s like, I know phishing through your emails, finding work to do, looking at what your clients are asking for, say if you’re a lawyer, that’s going to be just this massive leap forward in productivity. So that we’re working on that now for like basically getting to the point where your Spellbook agent can be in Slack as like this artificial legally competent coworker that you can delegate stuff to.

Turner Novak:

Yeah. Because I can think of from the VC perspective, it’s like you almost create this thing where you figure out the first engineer at Spellbook leaves and their LinkedIn says they’re working on a self-startup and my tool automatically messages them and gets on a call and even it’s like an AI agent that’s doing the call and then I get all this information and I just get an email, it’s like, “Would you like to invest or not?” Or whatever. That’d be pretty incredible if it does that.

Scott Stevenson:

I don’t think we’re far. I don’t think we’re that far from that. Yeah. I don’t know how we’re going to deal with the noise of AI agents calling and emailing everyone all the time.

Turner Novak:

And I get that a lot. You get all these like spam.

Scott Stevenson:

I do. Yeah. Yeah. They’re not very good.

Turner Novak:

And I don’t think, if I’m that founder, am I going to do a call with a VCAI associate? I probably wouldn’t though, but I might do a call with the guy who is making the investment decision or whatever. But in a sense, you do maybe skip through some pieces of the process and just more efficiently get to like, it’s like humans making decisions ultimately based on information from the AI. So maybe you do skip some, save some time. I’m not sure. I do go back and forth of like personally trying to wait out what are the ways that I just completely lean into AI and what are the ways that I just completely just choose to not do it at all and just lean more into like, podcasts is interesting, meeting in person, hanging out for two hours.

There’s like no technology really, aside from like cameras and mics and we’re just talking and communicating about this thing and that’s probably a good way to just get to know you better and build a relationship versus like, I don’t know, we could have like had our AIs exchanging information or something, but like...

Scott Stevenson:

That’s not interesting at all. Yeah.

Turner Novak:

Yeah. This is like you almost transcend and go above the technology in a way. I don’t know.

Scott Stevenson:

That’s true. That’s true. Yeah. I think about that so much with writing and tweeting. To me, AI writing, like creative or informational writing is so obvious still today, like the GPT-isms that everyone makes fun of. It’s not this, it’s that thing. You know what I’m talking about?

Turner Novak:

Kind of. Yeah. I honestly don’t even catch it because I don’t do enough AI writing. I don’t use it enough for writing.

Scott Stevenson:

Yeah. There’s like all these tropes that you just pick up on and the minute I see them, I feel like, “Oh, if it wasn’t worth the time for this person to write this, then it’s not worth the time for me to read it.” Because in a way, I think the fact that we’re taking the time to sit here and have this conversation or the fact that someone’s willing to actually sit down and write something themselves indicates that they thought it was important enough to invest that time. And that means for the reader or for the viewer, while that’s an indication that this might be worth my time as well. And it’s kind of like a proof of work with Bitcoin and stuff like that. I think if it’s like writing is sort of like this proof of work and the minute someone like at Spellbook sends me a recommendation of something we should do and the whole proposal is obviously written with AI.

I can’t trust that you actually thought about this enough. So to your point, we’re to not use AI. I think writing is an area, I’m like creative writing, writing recommendations, something I’m careful with. Luckily, contracts are very like, they’re not meant to be creative whatsoever. They’re very formulaic. Lawyers are not trying to be original. And so that’s one reason I think also why legal AI has taken off so much, especially in transactional work, is like there is no desire for really originality in contracts. People want to use standard language.

Turner Novak:

Yeah, because if you think about it, a business contract is almost the, it’s like the programming language of business, I guess, or something like that. If you want to really get philosophical about this stuff.

Scott Stevenson:

Oh, exactly. Yeah. It’s just like code. And so AI for code and AI for legal, I think in both of these areas, you’re not trying to write creative original code or creative original contracts. You’re trying to make functional documents. And that’s why I think AI works really well in those areas.

Turner Novak:

Yeah. And I want to talk a little bit about maybe early Spellbook stuff because you have some interesting history of the company, but going back even further than that, your first company that you started, you made an instrument, like invented an instrument. So what was that and how did it go?

Scott Stevenson:

Yeah. So that was my first kind of like ill-advised startup. I was super into electronic music and electronic... I studied computer engineering, but I grew up making electronic music, DJing, things like that. And I met this composer who was composing these awesome pieces of classical music, but incorporating these electronic elements. And he was like, “It’s really frustrating that there’s no electronic instrument that fits into that atmosphere that the audience will really appreciate.” If you go with a DJ turntable to a classical concert, people are like, “I don’t really understand, is this person just hitting buttons or whatever?”

Turner Novak:

They press play, but then they’re tweaking things.

Scott Stevenson:

Yeah. It’s like, “Are they actually doing anything? Are they not?” And so we conspired to build this instrument that would allow an electronic performer to really show the cause and effect of what they’re doing for these sorts of electronic performances and shape kind of like guitar or something. It has this beautiful wooden frame, kind of like an acoustic instrument.

Turner Novak:

And you almost hold it like you would an accordion in front of you like this.

Scott Stevenson:

You hold it on your lap or whatever. So you can face towards the audience. You can use it in a desktop mode as well, but it has all these lights and things, so it makes it obvious.

Turner Novak:

And there’s buttons and sliders kind of or something.

Scott Stevenson:

Yeah, button, sliders has like a synthesis engine. It has like a drum machine. You can use it in all these sorts of ways. And that was like my first very naive startup when I was fresh out of college, very little money.

Turner Novak:

So you like created this thing, made it, and were like kind of mass-producing them, but not...

Scott Stevenson:

We started producing them, selling some of them. We ran a Kickstarter. A couple of things happened, three things happened that set it off, of course. One was I got a really big legal bill. So one of my first bosses, this guy, Wally Haas, who ran this company, Avalon Microelectronics that I worked for was one of my first internships. He invested 20K in the company and that was a lot of money for me as like a broke student. And he was like, “This will get you to your next milestone.” And then one day we got like a 10K legal bill by surprise that took half that cash out of the bank account. And for me, that was an enormous amount of money. It was half the angel check we had and the amount of value we had gotten from that seemed like very, very little.

So at that point I started thinking about, “Okay, I think there’s like a way bigger problem to solve than electronic music instruments.” So that is where the idea for Spellbook came from was that like, frustration. But some other things that happened through that experience, like a hero of mine is this guy, Roger Linn, he built one of the first digital drum machines in the world. So if you listen to like ‘80s music and hear that like snare drum with like the big echo, that might be like a LinnDrum. And I got to go to NAMM, which is this big trade show for musical instruments. And I got to meet Roger Linn, this hero of mine and awesome guy, super friendly, but his company was still only two people, and he had dedicated his life to building these instruments. And I was like, I’m really glad he did that and I really appreciate it, but I don’t know if electronic instruments are like what I’m going to dedicate my life to. I don’t know if it’s a great market. And the TAM for niche electronic instruments is pretty small.

Turner Novak:

Maybe a million dollars, maybe a little more or less.

Scott Stevenson:

Yeah. These are like, they’re expensive to build. And I learned a ton about producing hardware and it was really fun, but maybe we’ll get back to it as a hobby someday. But like the legal market, having that pain myself and just seeing the size of the TAM of how many people touch contracts every day, it’s gigantic. Companies like Harvey and Legora really focus on the lawyers and the lawyer market. And I’ve talked a lot about lawyers, maybe 20 million lawyers in the world, but if you think about how many people touch contracts, it’s way, way, way beyond that 20 million number. So I was really inspired through that experience to start Spellbook.

Turner Novak:

Because like every salesperson, when you close a deal, there’s a contract related to it that you probably touch.

Scott Stevenson:

Yeah. Exactly.

Turner Novak:

I mean, it’s really any kind of business transaction that happens. There’s some handshake deals. Maybe you don’t sign a contract, but like most people do. Even the handshake deals that I do will do like a one-page contract, which is just like, we won’t screw each other, but we’re just making sure that we can’t screw each other basically with our very rough contract.

Scott Stevenson:

And even like an email can be considered a contract legally as well.

Turner Novak:

And so you were like, “Holy cow, I paid half of my bank account for this legal bill, I’m going to try to fix the legal market.” What happened from there?

Scott Stevenson:

Yeah. So I actually, I stewed on the idea for a while. I worked at a network monitoring company and was the director of engineering there, building out that product for a while. And then I was working on this in the background, trying to figure it out. And we went through a ton of different iterations. First in my mind, smart contracts were big and I was like, “Oh, maybe Ethereum smart contracts will be this automated type of contract we can all use.” That really obviously wasn’t going to work for a bunch of reasons. We showed blockchain smart contracts to a lawyer and they’re like, “I will never use this.” So we threw that away pretty quickly. But where we landed was, we had this product called Rally. It was a template-based product. So there was no AI at the time.

We actually launched in 2018 originally and we sold that to about 100 law firms. And basically what it let them do is build these really advanced legal templates. So if you’re doing a bunch of NDAs or sales agreements, you can build a template on our platform, which does a lot of things that you wouldn’t be able to do in a normal templateing engine. It’s built on Word. It can ingest legal data and then you could kind of spit out contracts much more efficiently.

But for a bunch of reasons, there wasn’t real PMF there for a long time. We were able to do 100 sales. We had raised some money. Our board was kind of like, “Let’s get on with it and scale this thing.” And we’re like, “No, we don’t think we have PMF,” a product market fit. Our view of product market fit is basically the customer is pulling the product out of your hands faster than you can keep up with. And until we hit that, we’re not scaling the company. So we kept the company super lean for a really long time as we built that out and we actually launched like over 100 landing pages.

Turner Novak:

This was like a three-year period, right?

Scott Stevenson:

Yeah.

Turner Novak:

I think you posted this one chart where it was just like the revenue I think of the company and there’s a point, it was maybe 2020 where you’re like, “We lost half our revenue or something.” What happened there?

Scott Stevenson:

Yeah. So we built out the platform. We did sell to the big law innovation committees at first and we had some very lucrative kind of early customers. And one of those customers churn was half our revenue and like we literally lost half our revenue overnight, but we felt that that wasn’t bringing our product in the right direction. Again, what I talked about earlier of like working with these innovation committees, they’re not the actual users. And so we’re like, “You know what? We’re going to start selling to these small firms solo lawyers to start and kind of snowball our way from there.”

Turner Novak:

So this is around when you started to test like a landing page. So you launched 100 landing pages in three years. What does that mean on practical? You were basically doing one every two weeks roughly.

Scott Stevenson:

Yeah, exactly. So we launched one every two weeks. Sometimes we would actually launch a product variation with the landing page. So at one point, our view was like, if we roll the dice enough times, eventually we’ll figure this out and find product market fit. And we’re optimizing for the number of app ads we could do. At one point we launched Shopify for law firms. So we took our templating engine and we put a store on top of it. So like a law firm could stand up like a Shopify store, like need an employment agreement, like click here and then it’ll use the template to like spit one out the backend. We tried an absurd number. We launched the client portal and we built landing pages for a ton of these different angles and things that we were trying.

Turner Novak:

So what’s the importance of launching a landing page for someone who’s like, “What is this even? What is the landing page?”

Scott Stevenson:

Yeah, a landing page is a single webpage that usually has no links to anything else that has a single message, an image that is trying to get you to sign up for a product or sign up for a wait list or something like that. And often you drive traffic to these through advertising or through social media through campaigns. So it’s not like people are landing on your homepage and finding it, you’re finding a way to drive traffic to it. So what we would do is launch these like little ad campaigns with maybe 1,000 bucks or something like that. And then we would drive traffic to the landing page and we would see what’s the cost per conversion, like how many dollars do we have to spend in ads to get a lawyer to sign up for the product from this landing page? And we literally tracked that over 100 landing pages where we could see, “Okay, we ran this experiment with client portal and that cost us $100 per lawyer and then we ran this experiment and that costs us $500 per lawyer.”

And you really get a sense of like what’s resonating and what’s not. And you really learn like how do you get a message straight into someone’s, past their like blood-brain barrier into their brain really, really fast. And yeah, then eventually we launched, the AI product was just another landing page. We were like, “Okay, GPT2 was around,” I had used GitHub Copilot for coding and we were like, “Oh, this is cool. We’re going to try launching GitHub Copilot for lawyers. We’re going to basically just launch another landing page with this.”

Turner Novak:

What did you call it? What was the buzzword because ChatGPT had not launched yet, right?

Scott Stevenson:

No, it had not. Yeah.

Turner Novak:

So how did you describe it in a couple words?

Scott Stevenson:

Well, we did call it Spellbook. The big thing we had was an image. The thing that we would do on a landing page is there’s a headline and there’s an image, and we would design the image or GIF or video. Our thesis is like, your image plus headline has to deliver this visceral sense of value in five seconds flat. And I think our headline was just like, “Draft and review contracts 10 times faster. Spellbook uses GPT-3.” People kind of knew what GPT-3 was on LinkedIn and stuff. It was a little bit of a buzzword. Even before ChatGPT, people were curious about what is this GPT-3 thing, and, “Spellbook uses GPT-3 to surgically redline your documents,” or something like that. But the important part was the image. We had an image of the word window and someone just clicking draft and it just drafts a clause instantly. And that was the sort of magic moment that once a lawyer saw hitting a button and drafting a clause from a headline.

Turner Novak:

Versus having to hand type or copy paste from somewhere.

Scott Stevenson:

It was just cuts through. And there’s this... Have you ever read this blog post that’s Find the Fast Moving Water on NFX? Have you ever seen it?

Turner Novak:

I don’t think so. What is it?

Scott Stevenson:

It’s a really good blog post. I forget who wrote it at NFX, but the author talks about their first experience of seeing Cabulus, which was like a precursor to Uber. And seeing this app for the first time, I think someone else had it, like his friend had it or something, and he was looking over their shoulder. And he was seeing this app for the first time and was like, “I had a neurochemical response to it. My pupils dilated. Blood rushed to my head and I just saw that the future ahead of me was going to look very different than what it has been when it comes to transportation.”

And I think that was the magic moment we were trying to hit with Spellbook, and we hit it. And within three months, we had 30,000 wait list signups. Within three months, we had more revenue from that product than the other three years we had selling everything else. When you hit that kind of resonance, you feel it. You see it in the numbers. It’s also unmistakable. I don’t know if you need to actually measure it as much as we did because when you really hit it, it’s like PMF resonance. Yeah, it’s unmistakable. You will know it. Yeah.

Turner Novak:

It’s like something that you can’t quantify because it’s just vastly order of magnitude or more of just way more resonance and usage of the product.

Scott Stevenson:

Yeah, signups. I mean, yeah. I mean, we couldn’t measure it too. The other things we were selling, the landing page might have cost us 100 or $200 in ads just for a signup. Now, when we first launched this, it was like $5, $10 signups, just almost like an order of magnitude cheaper in advertising to get someone to sign up.

Turner Novak:

So you basically pay $5 to get someone to sign up, and then if they convert and use it, they may pay 10 bucks a month or something and you might get like a 20% conversion rate from free signups to paid, and then you could do the math of saying like, “Okay, we need to acquire a fully converted user costs us about $25 and they pay us 10 bucks a month, so within three months, we are making money off of that customer.”

Scott Stevenson:

Yeah, yeah. That would be like your CAC payback. Yeah. And I think our original price was like 49 bucks a month. Today it’s more like 500 bucks a month. So VCs also talk about like, “Oh, our price is going to go down.” Actually, our price has only gone up as we’ve added more and more value to our customers. But yeah, that’s how you can do the math. And it is measurable. And when we hit that moment, our salespeople’s calendars were just completely blocked. Every day of the week, they would have eight sales meetings or onboardings actually. We didn’t do sales meetings. It’s just onboardings.

Turner Novak:

What were some of the other biggest things you think then you learned over that launch or landing page period and then just this thing worked and you just had to start going? Any lessons?

Scott Stevenson:

Yeah. I think the biggest lesson for us, or what I would tell other founders is yeah, if you keep beating... Resilience is incredibly important, obviously. If you believe in something enough and keep beating your head against the wall long enough, you will probably figure out something eventually. And if you keep your burn rate low, you can make money last a very long time if you’re scrappy. But the secret is you have to really believe in the problem. If you don’t believe in the problem, if we didn’t believe in the problem that we were solving, we never would’ve gone through a hundred landing pages for years of grinding with this super small team with very low salaries.

The reason we were able to do that is because we believe that legal efficiency was unsolved, that software still had not really changed how law was practiced and that a lot of people needed legal services, less expensive, a lot of in house teams needed more efficiency on their contracting, and we just really believed both, one, this problem is huge when you think about the scale of it, and two, that it is unsolved. And if you believe those things, you can be very resilient.

Turner Novak:

So then when this thing really started to work, you did not have a product yet, and were you just like, “Oh, we got to actually build this”?

Scott Stevenson:

Oh no, we did. I built the product, a really crappy version of it that our eng team tore apart. First we thought this was going to be a lead magnet. We didn’t even think this would be a product. We thought it’d be a cool marketing splash like, “Oh, first company to bring GPT-3 to lawyers. Wouldn’t that be a cool press headline? And then people will convert to our other product.” That’s what we thought would happen.

Turner Novak:

Really? Okay. And then eventually this became the product, right?

Scott Stevenson:

It is the product. Yeah. Yeah. It is 99.9% of our revenue is from Spellbook today, a very small amount from our previous products. But yeah, we actually built the prototype on Replit. This was before Replit had AI features. I built it in a couple weeks of evenings and weekends. It wasn’t a board goal. No one really knew it was being worked on. It was just a fun little side project.

And I built a really crappy version of it on Replit. And Replit ended up using us as a case study after, because the thing I said is it was such an amazing platform because I had this bolt of inspiration, and it would have perished. If it weren’t for platform like Replit, I would have just not found time to work on it and it would have just died in the shower where I had the idea. But because ideas are perishable and you have to rapidly chase them down because I was able to deploy it really fast, yeah, we actually did have a product on day one. It wasn’t great. The engineering team ended up having to basically rebuild it from the ground up.

Turner Novak:

Well, and I think the other interesting thing that I’ve heard you say before is that you were never attaching these experiments to kind of a legacy product. This was called Spellbook. It was a different sort of product.

Scott Stevenson:

Exactly. Yeah. I think that’s one of the things I would have done even more differently. We learned this towards the later phase. I think there’s this instinct for founders pre-PMF to just keep stacking on features and the list and the website gets longer and longer and more complex. That does not make it easier to pitch your product, especially in the earlier days. What you want is a pointed, easy to understand thing. So with Spellbook, our goal was like, “We’re not going to talk about any of the other things we built for the last three years. We’re only going to talk about this in a really pointed way to make it super simple for people to comprehend.” And that was part of the success. Yeah. So yeah, my advice to founders launching these landing pages would be just focus it on one thing. Don’t list off everything you’ve built because chances are if you’re pre-PMF, like 80% of what you’ve built doesn’t matter and you can delete it and there will be one or two things that matter.

Turner Novak:

Yeah. I think the other thing too is you think there’s all this historical context and history of the product, but literally 99% of the people, just they’re seeing you for the first time and they don’t give a shit about the history.

Scott Stevenson:

People are so busy, they are not paying attention.

Turner Novak:

Yeah. Yeah. And one thing I think I need to do more is just generally tweeting like, “Hey, by the way, I’m also invested in startups, whatever,” because I don’t do that enough. And there’s so many people that they’ve literally, they’ve followed me for years and they’re like, “Oh, I thought you were just kind of a meme account or something. I didn’t know you actually were an investor.” They’re like, “Oh man, I got to... “ So sometimes it’s just saying the message over and over again, reminding people the thing that you do almost.

Scott Stevenson:

Yeah. This is one of the things I’ve been learning from Keith since KV’s come on board. I think that’s something we were going to chat about.

Turner Novak:

Well, I wanted to ask you, so you guys really took off. You started growing really quickly. So what happened? Was it instantly you’re like, “Okay, we have PMF, we’re leaning into this”? Was there a debate of like, “How do we need to figure out this legacy old thing?” How did you manage that?

Scott Stevenson:

It was funny. No one knew we were working on this thing. The board had no idea, but by the time we got to our board meeting, it was almost like a mic drop drop moment. It was like, “Surprise, we launched a new product. Here’s the growth chart.” And it was just immediate consensus. It was like, “Wow.” Everyone’s like, “You need to go chase this thing. Don’t worry about the other stuff.”

Turner Novak:

And you’d been holding off, right? Weren’t your investors kind of like, “We have PMF. We need to start scaling”?

Scott Stevenson:

Yeah, yeah.

Turner Novak:

But at this point it was like, “We truly do now.”

Scott Stevenson:

Yeah, yeah, yeah. I mean, yeah, it was a pretty amazing moment for the team because we were so... We went through the 2020 era of the Zurp era, pure Zurp era where people were scaling way too fast pre-PMF and we really, really resisted it and sometimes we felt insane for doing it.

Turner Novak:

Yeah. Are you missing out because you didn’t do a remote video calling tool or something?

Scott Stevenson:

Yeah, yeah, yeah, and for not scaling the company. All the companies around us were just scaling, scaling, scaling, whether they had PMF or not. There was a lot of capital sloshing around and our investors were feeling the pressure also, I think to some extent to like, “Okay, let’s get the show on the road.” And it was a very validating moment for the whole team to be like, “This is what we’ve been talking about. This is real product market fit.” And the board was like, “Yeah, you’re right. I’m glad we waited for this moment.” And then it was just very fast consensus. It was like, “We have to scale this now.”

Turner Novak:

I think when I first came across you, you tweeted, it was a graph of your ARR growth or something and you were going to San Francisco to fundraise or something, and you just posted the growth. And I remember seeing it, I was like, “Oh, nice. That’s pretty cool.” And I just retweeted it because I was like-

Scott Stevenson:

Thank you.

Turner Novak:

... “I hope that you have a great fundraising round,” or whatever. And so how did that process kind of go of raising this... I think that one specifically was the series B.

Scott Stevenson:

That was our recent series B. Yep.

Turner Novak:

Okay. So just take us beginning to end, just how that process went.

Scott Stevenson:

So I mean, I will say it was the easiest raise we’ve ever had because it’s just numbers and metrics at this stage.

Turner Novak:

Yeah. Were you just showing the spreadsheet?

Scott Stevenson:

Basically showing the numbers, showing the growth, and it’s been really good. And it’s much easier than the pre-PMF phase where you’re convincing of just on pure vision. At least for me, it’s easier when you can kind of point to the numbers. But yeah, we started the raise. I made a tweet and I was like, “My goal of raising is usually to compress it into as tight a time as possible because I want to be working on the product. I want to be working with our customers.” No offense to... I like hanging out with VCs, but to move the company forward, I have to be constantly working with our team and our customers. And so you want to compress it, and you also compress it in order to create deal tension. If you dilute the deal tension across six months or whatever, no one moves, everything is sluggish.

And so I made this tweet. I tweeted our growth chart and I was like, “Yeah, raising our series B, I’m going to be in New York this week and SF the week after.” And that was it. And so it went viral maybe because of the chart, and we got bombarded my email. I still have emails I’ve not responded to from that moment from funds who reached out. And yeah, I set up in a hotel in New York for like a week and most of the investors all came to us. So one of the strategies I learned from another founder is try to, if you can sequence your meetings, we rented a boardroom in the hotel and we just said, “Hey, come here. We’re taking meetings this week, can dictate the schedule.”

So we had a lot of investors come by there, visited a couple offices and then did the same thing in SF basically. And then we got in touch with Keith and he was like, “I only take in person pitches.” And he was in New York and I was in SF and I had to fly all the way back and cut the SF time short.

Turner Novak:

And wasn’t he the only one who didn’t come to you basically? Yeah.

Scott Stevenson:

Yeah, basically. Yeah. I also went to the KV office as well, so I met Kanu there as well.

Turner Novak:

And what was kind of the difference? I know you said there was a pretty big mindset difference almost between East and West Coast investors, but what did you kind of experience there?

Scott Stevenson:

Yeah. For me, it was very night and day. For the most part, besides Keith, who I think is a very... He’s smart on the numbers, but he’s unique in how qualitatively he assesses the world. He’s very willing to bet on qualitative things. But the New York investors are extraordinarily quantitative and it seems like they almost have all converged on the exact same spreadsheet that they use, the exact same metrics, exact same benchmarks, exact same spreadsheets, to the point where it’s kind of a little bit absurd because if everyone’s looking at the same spreadsheet, then where’s the alpha? If everyone’s looking at the same thing, they’re going to pay the same price, everyone’s going to bet on the same companies.

I found there’s very little emphasis on the qualitative side with a lot of the New York investors we pitched, whereas in SF, there was a much deeper focus on the qualitative vision of the company and how is the future going to play out and why, kind of like the idea maze, kind of exploring the idea maze and understanding where things are going to go. Very different. Yeah, very different.

Turner Novak:

And then you end up, Keith and KV, Khosla Ventures led the round. What has it been like working with Keith just over the past couple months?

Scott Stevenson:

Yeah, incredible. So yeah, we’ve had a few board meetings and Keith’s an incredibly sharp investor. He’s been in the weeds. He really understands how things work on a deep, deep CEO level. And yeah, we just learned a ton. He obsesses about performance and what best in class people look like, and you just learn so much from someone like that.

One of the biggest things I’ve learned from Keith and that I’m learning is how to communicate really well. This guy is like a nuclear grade communicator, how incredibly concise he is, how he’s able to... I think he’s really good at counterpositioning companies and opportunities to cut through the noise in a way that other people really struggle with. And I mean, and he thinks about it deeply. It’s not by chance or he just happens to have this talent. It’s like he’s consciously very good at thinking about how to get the message to listeners and how to create a movement with a message in a way that... I think it’s one of the hardest skills for anyone to learn.

It’s a marketing skill. How do you get your message out into the world spreading? To do it, you have to have a really simple message and you have to repeat it a lot, like you were saying, people not knowing you’re an investor. You have to find a way to repeat that message or get that message in front of people because people are so busy, so distracted, they have no time. Getting someone to read more than five words is just an extremely hard challenge for the most part. Getting someone to watch more than a five-second video clip is pretty difficult a lot of the time.

So I mean, he’s just so aware of that and so good at dealing with it. An example I’ll give you, we did a series B announcement video and I think we booked 45 minutes for Keith to come down and talk about his view of the company and the opportunity. And we’re sitting down in front of cameras kind of like this and Keith sits down next to me and he just hits his lines. He’s just like, boom, boom, boom. He has five bullet points about why this opportunity is incredible, why the contract opportunity in particular is really special and how the market data that we’re collecting is going to change how contracting is done. I mean, he communicated everything in like five minutes and then he was like, “Okay, I think we’re done.” He couldn’t think of anything else to say. That was it.

And that’s the pattern that you notice. In the board meetings too, it’s like you’ll ask for feedback and he’ll say it in one sentence and then he’ll be quiet. And it’s like he’s so good at getting to the heart of the matter and then letting the core message breathe and be received. Yeah.

Turner Novak:

Are there things you’ve changed about marketing or messaging over the past couple months then?

Scott Stevenson:

Yeah, definitely. I think focusing more on delivering our core message, what we’re all about and repeating it to the point where it can become kind of boring for the speaker, I think just doing that more and doing it better.

Turner Novak:

One last thing I wanted to ask you about. So I feel like you’re maybe like bleeding edge of AI, quote unquote. Personally, what does your personal AI stack look like? What are you using? What kind of products and things are you taking advantage of?

Scott Stevenson:

I’ve tried a ton of stuff. Obviously I use Cursor and Claude Code on the engineering side for mainly building prototypes and things like that. The product I’m loving right now is Twin. Have you tried twin.so? Have you ever seen it?

Turner Novak:

No, I’ve never.

Scott Stevenson:

I mean, it’s on the surface very simple, but I think they just got like this generalized agent formula really right. So how it works is you can go in there, you can say, “I want to build an agent.” You build an agent by prompting. You don’t have to write code or connect together boxes or anything like that.

Turner Novak:

That’s always the most frustrating... There’s one called like N8N, I think.

Scott Stevenson:

Yeah, N8N. Yeah.

Turner Novak:

I never ended up getting it to work because I was like, “I don’t care enough to figure this out.”

Scott Stevenson:

Yeah. The way this works is you’re almost vibe coding these agents with a couple prompts and it’s really good at scheduling the work in the background, like what I was talking about earlier is you don’t want an agent that you have to prompt to do work. You want it to work on its own. And so I built probably about five agents with Twin now that I use daily.

Yeah. One of them is my Canada recruiting scanner, and what it does is every morning at 8:00 AM, it scans Twitter. Generally, we hire in the US, but most of our team is in Canada. We basically scan all of tech Twitter and find Canadians who are tweeting about AI and looking for engineers, designers and interesting people saying interesting things. And that fills our queue for recruiting, and that’s been a huge help.

Turner Novak:

Do you reach out to them or is it the team? What’s the process then for that?

Scott Stevenson:

Yeah. I mean, we manually kind of evaluate. Myself and our hiring managers will then look at the candidates and then we will do the reach outs ourselves. We’re not at the point where the agent goes and reaches out yet. We’re kind of tuning the quality and the filtering and stuff like that. So yeah, that’s been really useful.

We have another agent that does a similar thing where it just digests all the feedback from every channel, from Slack, from email, from HubSpot, and will basically summarize all of our product feedback every day in Slack. Why are people churning? Why are people expanding? Things like that.

Turner Novak:

So that’s probably the biggest one, twin.so?

Scott Stevenson:

For me, that’s the thing I’m loving the most right now. It’s just so easy. It’s so fast. I think you also, you want something that’s so easy that you want to make it so that if you’re dealing with a problem like, “Oh, I need to find my next podcast guest,” that it’s so fast to set up an agent to do that, that it almost takes you no extra time. So it’s like, “Oh, I’m going to go look for podcast guests or search Twitter,” or something.

Turner Novak:

I was going to say, you almost want the process of creating the agent is actually faster than just going on and doing the thing.

Scott Stevenson:

Actually, yeah, that’s right. Yeah. It’s actually faster. And Twin is the first thing that’s actually hit that level for me where it’s like, “I might as well create an agent,” and it actually works.

Turner Novak:

Yeah.

Scott Stevenson:

And the other thing I love, it has so many integrations, so it can suck in from so many things and then it can pump into Slack. I think the thing I always think about with products, the hardest thing is getting people in the habit of actually using them and getting the products in people’s faces and getting our team to go to a new agent product and changing their habits is really tough. But if we can pump the agent output into Slack, into channels that people are in, the usage is much better. Yeah.

Turner Novak:

That’s pretty cool. Well, I’ll throw a link in the show notes, people can check it out. I’m going to try it. I’ll see. I’ll let you know what I do with it. Yeah. But this is a lot of fun. Thanks for coming on the show.

Scott Stevenson:

Thanks for having me, Tuner.

---

Stream the full episode on YouTube, Spotify, or Apple.

Find transcripts of all other episodes here.