Cofounder, Gusto's AI teammate for small business owners, went from idea to closed beta in 11 weeks. That pace raised an obvious question for our research team: how do you learn from customers fast enough to keep up with a team shipping that quickly, without losing the depth that makes research worth doing?
Raj and I sat down with Amy Thibodeau, Gusto’s Chief Design Officer, to talk through how we answered that question. Watch the conversation below, then read on for the fuller writeup of what we built and what we'd tell another team facing the same tradeoff.
Cofounder, Gusto's AI teammate for small business owners, was built by a very small cross functional team in 11 weeks. That's fast enough to force a hard question for research: how do you incorporate learning from customers in a process that moves that quickly?
That question isn't unique to Cofounder. Research teams across the industry are running into the same problem as product development accelerates.
To get started, we built a panel of customers to conduct research over the 11 weeks, to avoid recruiting becoming a bottleneck down the road and help us run faster rounds of learning. We ran surveys and interviews, but scoped them more narrowly than usual, tying each one directly to what the team was about to ship. If we were shipping a new onboarding feature, the next survey or interview would be adapted to collect feedback on it.
AI was an important tool that helped us execute at this speed. It handled creating clear, consistent communications for recruiting and managing the beta group, gave us a fast starting point for spinning up new survey and interview guides, and generated session summaries so anyone could follow along as research unfolded. For this last part, we were able to use MCP servers and Claude skills to help pull session transcripts from our research tool and feed them directly into Claude Code where a research skill we’ve created helps handle the summary creation.
We also kept the broader team close to the research itself. We flagged issues in Slack threads the same day they surfaced, and invited teammates to sit in on live user sessions, so they could learn about unclear terminology, unexpected AI responses, or resonant use cases as patterns emerged, not after a report was written.
This approach helped us move faster on decisions like tuning onboarding for users who struggled to grasp the product's breadth, and identifying which tasks and automations people gravitated toward first. The engineers we worked with also built a system to generate GitHub tickets directly from UXR insights, turning customer learning into action faster.
The narrowly scoped studies meant leaving some questions on the table over the course of those 11 weeks. The team wasn’t running the kind of discovery work we might have done previously. What we gained in speed, we gave up in breadth, and we made that tradeoff deliberately. We could confidently make that call because the risk of shipping Cofounder without evaluating with customers was too high.
This model of rapid, narrowly scoped studies was possible because we had deeper customer insight to draw on. Our colleague, Courtney Bussing, led an ongoing foundational workstream into how small business owners think about and use AI to help run their businesses. With that foundation we didn't have to build an understanding from scratch. Our team drew on work that came before the project even started, and we’re feeding what we learned back towards the broader team as well.
That's the model we're building: foundational research that builds understanding of our customers, and rapid research that informs decisions at the pace those decisions demand. AI helps us make both possible.
How we’re using AI for UXR at Gusto
With product teams moving so much faster, we've needed to adapt how we learn from and share customer understanding, or risk a disconnect between research and the products being built. As Constantine Papas put it, “AI made it cheap to propose. It did not make it any cheaper to validate. [...] The more options generated, the more bets place–and the more important it becomes to have a fast, credible way to check at least some of them against reality.”
A research team that uses AI well can move faster across the whole process, and help both people and AI agents act on what they find. Here's what that has looked like for us so far.
Research Claude Skills for end-to-end studies. Anyone on the R&D team, researcher or not, can build a solid research plan, discussion guide, survey, summarize sessions, and conduct analysis in Claude Code by leveraging Claude Skills the research team builds and maintains. Tasks that once took a few days now take an hour or two. At Gusto, we want any PM or Designer to be able to lead a study on their own, and the Skill enforces some of what a researcher would otherwise catch by hand: it flags leading questions, checks that a plan has enough participants to answer what it's asking, and will help build a screener question and customer list that will help team include the right people in the research. It's not a substitute for a researcher's judgment, but it catches the mistakes that used to slip through when someone was moving fast and building a plan from scratch.
Better prototyping tools for researchers. Work that used to take at least a week can now happen in a day or two, using MCP servers connected to our design systems that let Claude Code build functional prototypes directly. That capability is no longer limited to designers; researchers and PMs can now build their own concepts to test.
Insights explorers get teams their findings faster. We're experimenting with AI tools that help teams get more out of past research. Our research planning skill pulls relevant findings from our research repository to sharpen a new study's questions. We've also built what we call Session Explorers: using session transcripts, researchers can generate clickable internal documents showing the key elements of a session almost instantly, filterable around whatever questions matter most to the team. These aren't a replacement for a full research report, but they let a researcher surface findings fast enough for a team to start thinking about next steps immediately.
AI-moderated interviews. We're using AI-moderated interviews for running narrowly scoped evaluative work, where the questions are simple and the risk is low. We're not using them anywhere depth matters, and we’re building guardrails to help maintain quality and connection to the data that’s produced and interpreted.
Where we’re going from here
If there's one thing we'd tell another research team starting down a similar sprint cycle, it's this: don't wait for the product team to slow down to fit research in. Start by understanding how and where decisions are being made, then figure out what it takes to move at the product team's pace to inform those decisions with customer insight. Identify your bottlenecks and go after them directly, whether that's tooling, context, or process, but know which problem you're actually solving before you tackle it.
“There's a lovely irony in that the best way to make good use of AI to help us build our understanding of customers is through human connection and trust.” — Amy Thibodeau, Chief Design Officer
Relationships are foundational. For us that meant operating with skin in the game: showing up in the daily Slack threads and adjusting the next interview around whatever had just shipped. Have a plan for how you'll run rapid research on what needs evaluating, without losing sight of the deeper customer understanding that rapid research alone can't give you.
Now that Cofounder has expanded out of its closed beta group, we have more opportunities to learn from new and existing users, so our research isn't done. We're also collaborating with Design to develop best practices for designing in-product AI experiences based on what we’ve been learning from customer research. The goal is to help the whole company figure out how to craft AI experiences for our customers in ways that make sense and provide real value, without feeling forced. This is a challenge many in the technology space are grappling with.
We're also ramping up rapid research with AI-moderation and the team is learning a lot about how we can best leverage it and when, and what space that makes for other types of learning. We're also rolling out an end-to-end AI-assisted discovery loop for product teams that don't have a researcher embedded in them. The other half of what we’re doing with AI is building new tools that help humans and agents use our existing understanding of our customers, and bringing customer understanding into decisions that would otherwise move without it.
We're still learning what this looks like at scale, beyond one product. But Cofounder gave us a model to build on, and we're testing it now across other high-stakes bets.
