Getting finance teams to trust an AI with their numbers
Drivepoint's GL Mapping Agent keeps a consumer brand's financial reports accurate. It's the kind of work nobody notices until something breaks, and this is how I designed it.
I led design end to end as the only designer on the project: UX, UI, interaction design, the AI agent's behavior, and copy. The work ran from January to March 2026 at Drivepoint, an AI financial planning platform for consumer brands, and launched publicly that May. I built it alongside a PM and the engineering team.
Launched publicly in May 2026. In an early walkthrough, a VP Finance spotted $377K of costs mapped to the wrong place, and fixed it before the meeting ended.
One wrong account can quietly cost you six figures
Every consumer brand runs its finances on a translation between two systems that describe the same money in different ways. On one side is the accounting system, organized the way a bookkeeper files transactions. On the other is the financial model the business actually makes decisions with, organized by sales channel, margin, and product line. Mapping is the layer that connects the two.
Get one account wrong and the error does not stay contained. A channel looks more profitable than it is, so the brand pours money into it and starves the ones that are actually working. The model is off, the reports are off, the decisions are off.
Before this project, Drivepoint's customer success team set up the mapping for each brand by hand, working through hundreds of accounts inside a cramped Excel add-in panel. The screen was small and hard to read, so getting a brand fully mapped took a long time. And it was never really finished. New accounts show up every month as brands add products, channels, and vendors, and each one meant going back to CS.
AI automapping
I joined knowing nothing about FP&A
My finance background was crypto-native. At Gitcoin I spent years designing onchain funding mechanisms, so I was fluent in one corner of finance. But consumer-brand FP&A was a different world. I could not have told you what a chart of accounts was.
I built that fluency the only way that sticks: interviewing finance leaders, sitting with analysts while they worked, and studying the domain on my own until I could read a P&L the way they do. That mattered more than any UX pattern. You cannot design a tool finance teams trust if you cannot tell when the product itself is reasoning about their numbers incorrectly. Which is exactly what happened next.
The assumption that would have broken everything
The original plan was to map accounting "classes" straight to Drivepoint "channels." On paper it looked clean. Both are ways of grouping money, so line them up.
Talking to customers, I realized it was fundamentally wrong. A single cash account like "Chase Checking" might be tagged with Amazon, DTC, and Finance classes, but it is still just a cash account. It belongs to no channel. Mapping by class would have silently handed channels to accounts that do not have one, corrupting the exact channel-level reporting the product exists to deliver.
That insight reset the design. Instead of class-level rules, the agent maps based on the account itself and how it is actually used, with AI proposing patterns that a person reviews. A designer without domain fluency would not have caught this. I almost didn't.
New workflow for mapping accounts
Three decisions that earned their trust
Audit, don't approve. The agent is confident on about 90% of mappings and handles those automatically. It surfaces only the handful it is unsure about. Reviewing ten judgment calls is a task a finance team will actually do. Reviewing four hundred is one they will abandon.
Show the consequence before the change. The agent never says "I found an inconsistency." It says what that inconsistency does. For example: "this would have inflated your CAC by about $34k a year." Finance teams care about reporting impact, not data hygiene, so I wired the impact into how the agent speaks.
Make committing deliberate. Changes move through approve, then staged, then saved, then live. For production financial data, a little friction is a feature. The two-step save makes sure no one rewrites the numbers their company runs on by accident.
Manual bulk approval
What it looks like in practice
Shortly after launch, during a walkthrough with a women's health brand, the VP Finance noticed something mid-call. $377K of Amazon fulfillment costs were sitting under "shipping" instead of "fulfillment." Categorized correctly in the accounting system. Mapped to the wrong place when it reached the model. Quietly distorting that channel's profit for a full month.
She remapped it, saved, and reimported. The gap closed before the meeting was over. That is the whole point of the design. Not a quarterly cleanup project you schedule when things get bad enough, but something you reach for the second a number looks off.
What shipped
The GL Mapping Agent launched publicly in May 2026. It maps, validates, and maintains a brand's chart of accounts across their model, reports, and other agents, and it gets smarter with every correction.
The agent maps about 90% of accounts on its own and surfaces only the 10% that need human judgment, learning from every correction.
Fully self-serve. Customers map their own accounts from the Account Mapping tab, with no CS setup or scheduling.
Channels that match how brands actually sell. Custom channels for cases like Costco, TikTok Shop, and regional distributors, instead of forcing everything into DTC, Wholesale, Retail, and Amazon.
Dimensional detail preserved. Revenue and expense by department, location, or class flows in from QuickBooks or NetSuite with its structure intact, instead of collapsing into a flat P&L.
What I'd do differently
I would pull real customer data into the confidence thresholds earlier. We tuned what counts as high confidence partly on judgment, and seeing the live distribution sooner would have let me set those cutoffs with more precision and fewer assumptions.
I would also design the ongoing maintenance experience with the same care as first-run setup. Most of the early energy went into the initial mapping. But the product's real value is that mapping is now continuous, not a one-time task, and that state deserved as much attention from day one.
© 2025 melissa neira