a16zHow AI Is Turning Figma’s Finance Team Into Builders
a16z
Finance as operations 0:00
Pier Milwani says finance at Figma should act like an operational leader. Early on, he built SQL tables by hand, joined product data with Stripe records, and pushed those numbers into systems like Salesforce and Marketo. That work gave him a single set of facts and let the team stop arguing about what was true. He sees AI as a bigger version of that same force, but he still expects people to think clearly and not outsource judgment.
Hiring for curiosity 8:30
For hiring, he looks for people who are intellectually curious, fluent with data, and able to ask sharp questions. In interviews, he uses a case study to see how candidates use the tools available, how well they understand the model they build, and whether they can explain it plainly. He values AI as a helper, not a replacement for thought. The best finance hire should understand the business, spot changes fast, and help drive action.
Finance memos and personas 13:00
The finance team works in a document-heavy culture, so it writes investment and M&A memos for leadership and product leaders. The team has built unique personas for each leader, then runs drafts against those views to sharpen the questions and cut the time spent in meetings. The speaker sees three uses for AI: it augments work, replaces some workflow steps, and helps people learn. That has made him more optimistic about hiring and growing a finance team that wants to keep getting better.
Build, buy, and slop 15:01
Some tools are built directly in code, some with ChatGPT agents and workflow tools, and some inside Figma Make, which can turn prompts into apps or websites. He wants the company to give product feedback too. He also worries about AI slop: too many half-baked docs, weak thinking, and people losing sight of what good work looks like. He says you should not outsource your thinking, even if AI makes first drafts faster and easier to review.
Shared finance builders 25:30
People on the finance side are being treated like core product builders. Some come from accounting backgrounds, and that helps because they know the hard parts have to be right: checks and balances, refunds, and uncollectible items. That shared understanding lets them move faster and push other people’s thinking. The team also talks about who owns key definitions, especially early on when founders may set reporting metrics for investors and business decisions. Even as AI makes it easier to reach the Stripe logic or code, the goal stays the same: own the outputs and keep the historical numbers consistent.
Automation with judgment 28:30
The most automated people are spread across the team, from forecast owners to BISOPS staff doing competitive benchmarking and pricing studies. Speed comes from knowing the data sets, how they are stitched together, and where a model is deterministic versus inferential. Show-and-tell sessions, internal hackathons, and outside finance conversations help spread that knowledge. Core systems of record are often bought, but workflow automation is where building is more fun. The work shifts from reconciliation toward experiments, strategic questions, audit trails, and the people skills of communication, trust, and followership.
Automation and finance judgment 38:31
The discussion turns to what finance will look like in the next 12 to 18 months. The hope is to add people only when business complexity grows faster than the team can support it. Even so, new products and acquisitions keep pushing complexity up. The speakers expect more work to become fully automated, but only where the steps are deterministic and the output can be trusted. Audit-ready results matter most for business-critical systems, and that will have to satisfy auditors too. Excel is not seen as under threat. It remains a way to build trust, test assumptions, and understand models by hand. AI can help read a 20-tab model and spot the biggest drivers or sensitive areas, but it does not remove the need for judgment, fundamentals, and clear storytelling.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
