How Figma’s Finance Team Uses AI to Build and Automate

Seema Amble and Ivan Makarov of a16z welcome Figma CFO Praveer Melwani. They argue that finance leaders increasingly need to design systems and automation, and ask how his team is changing its work.

Melwani says he has always wanted finance to act as an operational leader. When he joined Figma, the automation was not in place, so he used SQL to build data tables and bring financial information into conversations across the company.

He joined product and usage metrics with Stripe financial data, writing long SQL queries to create shared tables. Later, data scientists helped put those tables into recurring ETL jobs and feed go-to-market systems. Consistent definitions let teams spend less time arguing about what the numbers meant.

Makarov recalls that a spike in product usage once took weeks to investigate. The group contrasts that pace with modern tools that can help teams examine revenue or usage changes much faster, although the underlying data still has to be reliable.

Melwani says he checks daily or hourly dashboards, looks for anomalies and asks questions alongside his team. His aim is to make the same data available to colleagues so finance can act as a business partner rather than a gatekeeper.

In Figma’s early days, he reviewed individual paid sign-ups to see who customers were and what they might need. Today he wants more people to feel accountable for metrics and investigate changes promptly instead of waiting for the next weekly meeting.

Asked about hiring, Melwani emphasizes intellectual curiosity, data fluency, good questions and clear communication. AI can reduce repetitive work, he says, but a strong finance hire still needs to understand the business and pursue the right outcome.

For a finance leadership case study, he expects candidates to use available tools and public information. He then probes whether they understand the model they produced, have tested its sensitivities and can explain the business implications.

Anyone can produce a model with a language model, Melwani says; the important question is whether the tool improves the person’s thinking. He wants candidates to use AI without outsourcing their judgment.

Asked for actual workflows, he describes an employee in credit and collections who connected invoice tracking, collections and reconciliation with NetSuite in a homegrown system. The goal is to spend less time on manual work and more time reducing outstanding receivables.

The same work expanded into credit checks and combining data sets. Melwani sees these tools as a way for a specialist to improve a process they know intimately, then make the result available to colleagues.

Figma’s corporate development team also tries to prototype a target company’s product during diligence. Melwani says the exercise helps the team form an intuition about the technology and its difficulty; product teams still participate in the assessment.

The team writes investment and acquisition memos and has created simulated leadership perspectives to challenge drafts. Running a memo through those perspectives can surface recurring questions or areas needing more diligence before a real leadership discussion.

On build versus buy, Melwani says the examples use a mixture of coding agents, workflow automation, ChatGPT and Figma Make. Finance staff can also give product feedback when they build with their own company’s tools.

The hosts raise the problem of AI slop: a flood of documents that are easy to generate but costly for others to read. They ask how teams maintain standards and develop the experience needed to recognize high-quality work.

Melwani says Figma has internal policies for identifying AI-written material. Whoever sends a document should still read, review, simplify and understand it, because consuming it costs someone else time and attention.

He contrasts poor output with a travel-and-expense analysis that combined information from several systems. The team had to reconcile ambiguous records, such as whether a person was actually traveling for an off-site. Getting the data into a useful form was much of the work.

For that planning discussion, Melwani wanted directionally useful information rather than a perfect answer. He says the team visualized suggested changes and incorporated feedback quickly. What counts as good enough depends on the decision; it is not the same as an audit-ready result.

Amble describes a company whose staff built dozens of agents, many of which were later unused. Her point is that experimentation can still develop an automation-first habit; the figures are an anecdote, not a measured success rate.

Melwani sees repeated patterns across internal audit, governance and accounts payable: teams ask for information, follow up and reconcile records. Instead of everyone building an isolated version, a systems team could turn common patterns into a shared platform.

The hosts note that individually built dashboards can also show different numbers. Melwani proposes that domain specialists prototype what they need, while a central systems organization handles dependencies, hosting, reuse and continuity when the original builder leaves.

Figma is still in what he calls the messy middle: some tools are centralized, while other teams built faster than supporting infrastructure could catch up. He wants a CIO organization that identifies successful experiments and creates foundations others can build on.

He compares the emerging finance-engineer role with data scientists embedded in business teams. Some people supporting Figma finance have accounting backgrounds, which helps them understand checks, refunds and uncollectible amounts as well as the software they build.

Makarov asks who owns metric definitions as AI makes code and payment logic easier to inspect. Melwani says his team owns the outputs and values consistency in disclosures and historical numbers, even while people explore new ways to understand the business.

Automation is spread across forecast owners and business operations, Melwani says. He credits knowledge of the underlying data architecture and the ability to distinguish deterministic steps from model inference; not every task tolerates the same uncertainty.

Show-and-tell sessions, internal hackathons and conversations with outside finance leaders help the team learn from one another. The objective is not merely to produce more agents, but to give people examples of useful changes to their own work.

On buying software, Melwani says Figma purchases core systems of record and solutions for tasks it cannot afford to get wrong. The team often builds around those systems to automate the repetitive parts of its own workflow.

He says the freedom to explore has made work more enjoyable for some employees and may help recruiting and retention. Access to tools alone is not the whole appeal: candidates also see colleagues learning and building with them.

Amble recalls a finance pyramid in which reconciliation occupied most of the team’s time and strategy occupied the smallest part. She asks whether automation has actually created more room for strategic work.

Melwani says he wants teams supporting the self-service business to spend less time reconciling prior periods and more time examining experiments, trends and spending decisions. Automatically preparing consistent materials can leave more time for annotation and discovery.

The hosts discuss finance products that provide both an answer and an audit trail showing how a reconciliation was reached. Melwani adds that tools can produce a first pass on changes in general-ledger accounts, giving accountants leads to investigate rather than removing review.

Melwani reflects that he once underestimated communication, relationship building and influence. Finance is a team sport: knowing the answer matters, but a leader must also build trust and bring others along.

Looking ahead 12–18 months, he hopes regular processes become more efficient, while new products and acquisitions continue to add complexity. He says that if business complexity stayed flat, a year from now the team might not need the same number of people; that is a conditional assessment, not an announced staffing plan.

He expects deterministic tasks to be candidates for further automation, but says the bar for trustworthy, audit-ready output in a critical system is very high. Auditors will also have to be comfortable with the process.

Is Excel threatened? Melwani says he still values the spreadsheet as a hands-on way to build intuition and trust. Natural-language tools can produce a first draft, but direct manipulation can make it easier to explore alternatives.

He has also experimented with using a model to understand a complex, multi-tab financial model and identify sensitive assumptions. He treats that as a starting point for investigation, not a substitute for understanding the model.

The speakers agree that tools change quickly, while foundational finance skills remain. Melwani’s advice is to hire a leader who understands the business, can explain its model and has the operational ownership to help a CEO make change happen.

He adds that creativity matters: finance can frame the business differently, contribute ideas and influence internal strategy. Strong accounting and planning fundamentals remain necessary alongside that broader perspective.

Asked where companies should begin automating, Makarov starts with data integrity and access. Early-stage businesses may first need to establish tracking, correct definitions and reliable reporting before layering on metrics and a coherent narrative.

The conversation closes with the hosts observing a shift from generating documents toward reconciling data and automating core workflows. Melwani says Figma’s internal maker week now feels much more accessible to finance staff with ideas to try.