Latticework · Y Combinator
Solving the Blank Canvas Problem
Eddie Kim built Gusto's AI co-founder himself — and in doing so, surfaced the most important UX problem in the agentic era.
500kGusto customers
$1BAnnual revenue
1/5New US businesses
64mRuntime
Eddie Kim, Gusto co-founder & head of technology, in conversation with Y Combinator.
I · The Frame
The 99.9 Percent Problem
Gusto crossed a billion dollars in annual revenue serving over half a million small businesses — one in five new companies in the United States starts on the platform. That scale makes Eddie Kim's thesis about the blank canvas problem more than a product opinion. It is empirical evidence about how most people actually experience AI.
The cutting edge of the tech world has persuaded itself that the agentic era is already here. In practice, the vast majority of users still treat AI as a glorified search engine — they type a question, wait for a response, type again. The conversational ping-pong loop is comfortable but shallow. The truly powerful move — directing an AI to take sustained, multi-step action on your behalf — requires a creative leap that most people never make.
Kim's insight is that this is not a failure of imagination on the user's part. It is a product failure. The interface offers infinite degrees of freedom at exactly the moment the user most needs constraint and suggestion. The latticework this conversation adds is not just about Gusto's product; it is a frame for thinking about every AI product's onboarding, default state, and activation arc.
the agentic world that we've been promised has never really materialized for most people
have this like thing I have in mind and then I'm expecting a response from it. And this like sort of like agentic world that we've been promised has never really materialized for most people out there. I think there's like the tech world that's on the cutting edge. Folks that are like creative enough to come up with all these ways to automate, you know, all parts of their lives, but the remaining 99 99.9% um are still kind of like using this as a glorified search engine. And so what we wanted to do with Gusto co-founder is bring all of these like agentic promises that we were receive that we received u and turn them into into reality. And so one of the ways that we've done that with Gusso co-founder is instead of starting with like a open claw type experience where uh it has a lot of powerful capabilities um but it has a main has a pro has a problem which I call the blank canvas problem. We actually start with all of
II · The Reinforced
Models the Source Amplifies
Activation energy is the minimum input required to start a reaction. In chemistry the catalyst's job is to lower that threshold. In AI product design the blank canvas — a blinking cursor and infinite possibility — raises it. Kim's solution is to pre-populate the interface with automations drawn from what the customer already does on Gusto: running payroll, approving timesheets, reminding employees. The canvas is no longer blank; the first move is already half-made. This is the textbook activation-energy intervention applied to software onboarding.
Inversion is Charlie Munger's prescription to solve a problem by starting from its reverse. Rather than asking "what should we build?" Kim asked "what are customers already doing every single week?" The massage spa that exports MindBody data, reformats it in Google Sheets, and finally enters it into Gusto to run payroll — that workflow is the product spec. The answer was already in the data.
powerful capabilities but it has a problem which I call the blank canvas problem
powerful capabilities um but it has a main has a pro has a problem which I call the blank canvas problem. We actually start with all of the things that Gusto is already solving for our customers. You know, those are things in payroll, HR, time, scheduling, and then we sort of suggest ways that we can automate their running payroll. They're approving a time sheets. These are things that they're doing sort of on a regular basis. You know, payrolls run once a week usually for hourly if you have hourly employees or twice a month if you have salaried employees. Um, and we we will suggest ways that we can actually take it end to end without the user even having to like log into Gusto and and run their process. >> Um, something I'm really curious about because you guys are already at like massive scale. Um, how what does it from the inside? What does it look like trying to get a product like this is essentially new product like trying to get it out there like um >> how did you come up with the idea and then just how do you sort of at your
Skin in the game is visible throughout: Eddie built the initial prototype himself, using it to automate his own workflows before showing anyone else. The founder who ships is a different animal from the founder who directs. The prototype revealed what the roadmap could not — where the friction actually lives, and what "good enough" feels like in practice.
Leverage shows up in the team composition story. A small squad — including a designer who had never written production code — shipped a complex AI product in a compressed timeline. The reason was not heroics; it was that AI dramatically lowered the cost of crossing role boundaries. One engineer with Claude Code was doing the work of three. The leverage ratio on technical output had changed fundamentally.
a massage spa business that exports data from MindBody, they put into Google Sheets
there's a business like a massage uh spa business that um I talked to and they export data from MindBody, they put into Google Sheets, they have all these calculations on how that data from MindBody gets translated into uh commission tips and um and hours and and and things like that. And then and only then they put it into gusto and they run payroll. That last part is obviously like very very simple and straightforward, but it's all like kind of like this work before the work that takes them a lot of time and they do like literally every single week, right? And so when you like actually talk about this to a customer, they actually instantly get it because they have so many of these like I do this every week. like it's on my calendar and it takes me like one hour every single week and if Gusto co-founder can just like do this for me and it just texts me like a summary with a final approval where I just say yes or no um over either SMS or Slack like they actually really get that it doesn't really take much like selling at all. >> Yeah, that makes a lot of sense. And actually I guess now you're hearing you
III · The Contradicted
Models the Source Bends
The classical SaaS product wisdom is to do customer development first, build nothing until you have a validated insight, and never let engineering run ahead of product understanding. Kim violated this deliberately. He built the prototype himself, on his own time, and only then began sharing it. The prototype was the discovery instrument, not the output of one. In the agentic era, building fast enough to show rather than tell may be the only reliable path to shared understanding.
Specialization and division of labor is one of Adam Smith's most durable ideas. It predicts that teams perform best when individuals go deep in narrow domains. Gusto's build story contradicts the prediction at the margin: a designer contributing production code and engineers prototyping Figma designs, with AI absorbing the translation cost. Division of labor held at the macro level, but the micro-boundaries dissolved. The lesson is not that specialization is dead — it is that AI has raised the minimum viable competence threshold in adjacent roles.
a designer actually able to write production grade code and engineers comfortable with designing
contributing as much code as any of the engineers, >> right? >> And it was really really cool to see a designer actually be able to like write like production grade code and then also like engineers really be comfortable with designing, right? And what happened is the engineers would kind of just like prototype something like do their best at design. they wouldn't be afraid of they wouldn't get their you know wrist slapped because they didn't you know design something perfectly or use Figma in the beginning um and the designer would know how to use cloud code enough to sort of like refine it and make it look really really good right and at the same time designer would like she wants some functionality she would write the code um and it was pretty good but then the engineers would then take that and like sort of like get it you know the rest of the way there and so you kind of have this like merging like designers are going more in engineering and engineers is going more into design and so the craft really became secondary and everybody's like shared like
The idea that interface neutrality is a virtue — that a general-purpose prompt is the ultimate flexible tool — is undermined here. Blank-canvas flexibility is the problem, not the goal. For the 99.9%, the right interface is highly opinionated, pre-loaded, and context-aware. Freedom is only useful once the user knows what they want. Before that, suggestion beats optionality every time.
IV · The New
Models Worth Adding
The blank canvas problem deserves its own entry in the latticework. It is distinct from activation energy because the problem is not just effort — it is the cognitive paralysis of open-ended choice. A general prompt interface offers infinite moves; most users make none, or default to the simplest. The solution is not a better prompt but a different paradigm: start from the customer's existing context and surface the moves that are already latent in their data.
The automation heartbeat is a new primitive worth naming. Kim's co-founder product runs not on user commands but on a scheduled cadence — a cron job that fires an LLM every week, every day, at the moment the customer's recurring task would normally begin. This is a fundamentally different architecture from chat: the system initiates, not the user. The heartbeat converts a reactive assistant into a proactive co-worker.
automations run on a heartbeat that gets triggered on a regular basis — a cron job that runs an LLM
leverage a prompt plus the data that we have about our customers uh to build them an automation for a business process that they're doing every single week. And so that ultimately became the the that's the main concept of Gustoa co-founder which is centered around an automation and Guso co-founder has many automations for a customer and they run on sort of a heartbeat that you know gets triggered on a regular basis. Um that code sync >> yeah and so like I mean part of that was also like when I was setting up open call I was like curious how it worked. So it, you know, it's open source. So I just read the source code of like, how is it actually like running this like automation for me and it was like surprisingly simple, right? This is a cron job that runs an LLM every 30 minutes. >> And that's exactly how it works in Gusto co-founder, right? So it wasn't even just like some of the concepts of like you know being able to text it through SMS or talk to it through Slack but it was al literally some of the like
The pre-workflow tax is the hidden cost embedded in the massage-spa story. The nominally simple step — entering data into Gusto — sits downstream of a manual gauntlet: export, reformat, calculate, import. The automation that matters is not the last step but the full chain. Products that optimize only the final step miss most of the value. Mapping the pre-workflow tax is a diagnostic technique: interview customers not about what they do in your product, but about what they do before they open it.
in the era of vibe coding, you're able to build the prototype yourself
in the era of vibe coding, you're able to like like you could build the prototype yourself versus like trying to get a a team to rally around it. >> That to me was even more mind-blowing than like kind of the what we built is like kind of how we built it. Um so obviously started with this like prototype. Um, I just kind of like showed a few of our engineers and designers and ultimately just kind of like organically roped them into working on this thing for me. So, one big change was like I was this was actually built by five people, five AI builders over the course of 10 weeks u from start to finish just a whiteboarding session and 10 weeks later like we went through through a full-blown like tier one launch in the company. Wow. that would never have been possible I think uh without AI and it was and the fact that it was just done with five people was also like equally and I was one of the five um to to uh in this team uh was
VI · Coda
The most important insight in this conversation is not about AI at all — it is about defaults. Every product makes a choice about what state the user finds when they arrive. Most AI products default to possibility. Gusto defaulted to context. The customer arrives to find their payroll half-automated, their recurring tasks surfaced, their weekly rhythm already mapped. They are not asked to imagine what AI could do for them. They are shown what it is already doing.
That shift — from invitation to demonstration — is the design pattern of the next decade of software. Not "what do you want to build?" but "here is what we noticed you do every week."
"The agentic world that we've been promised has never really materialized for most people out there… what we wanted to do with Gusto co-founder is bring all of these agentic promises into reality." — Eddie Kim, co-founder, Gusto