Getting finance teams to trust an AI with their numbers
Drivepoint's GL Mapping Agent connects a consumer brand's accounting system to the financial model it uses for planning and reporting.
Before this, Customer Success mapped hundreds of accounts by hand for every new customer. I designed the agent end to end as the only designer, including UX, UI, interaction, agent behavior, and copy.
Mapping was both a growth tax and a financial risk
For Drivepoint, mapping was a bottleneck to growth. Customer Success set up every brand by hand, hundreds of accounts at a time, and the work never really ended. New accounts appeared every month, creating recurring manual work for every customer.
For customers, the stakes were higher. Account mapping is the foundation underneath Drivepoint's reporting and profitability model. If an account lands in the wrong channel, the error propagates into the P&L, making one part of the business look more or less profitable than it actually is.
The product needed to automate the work without making financial teams less confident in the numbers.
AI automapping
The hardest part wasn't the interface
It was understanding enough about financial reporting to know when the product itself was getting the domain wrong.
I came in with experience designing complex financial systems at Gitcoin, but consumer-brand FP&A had its own logic and vocabulary. I built that fluency through interviews with finance leaders and by working closely with analysts, learning to read a P&L the way they do.
It's why I could see the mapping model itself was wrong.
We were mapping the wrong thing
The original approach was to map accounting classes directly to Drivepoint channels.
It looked logical on paper, both were ways of grouping money. But customers revealed a problem.
A cash account like Chase Checking might be tagged with Amazon, DTC, and Finance classes. The classes describe how the account is used, but the account itself is still cash. It doesn't belong to any sales channel.
Mapping by class would therefore assign channels to accounts that don't have one, corrupting the channel-level reporting the product exists to provide.
So I changed the agent's underlying logic.
Instead of mapping classes to channels, the agent reasons about each account and how it's actually used.
The AI proposes mappings where it has enough evidence. People make the judgment calls where it doesn't.
New workflow for mapping accounts
Three decisions that earned their trust
Audit, don't approve
The system was designed so the AI handles obvious mappings and only escalates uncertain cases. ~90% of accounts are resolved automatically, leaving a small set for human judgment.
Explain impact, not errors
Finance teams don’t act on “inconsistencies”, they act on business impact.
So instead of flagging issues, the agent translates them into financial consequences (e.g. “This would have inflated CAC by ~$34K/year”).
Add friction where it protects trust
Because mapping affects live financial models, changes move through a deliberate flow: Approve → Staged → Live.
That extra step ensures teams can review high-impact changes before they affect reporting.
Manual bulk approval
What it looks like when it works
Shortly after launch, the product was used in a walkthrough with the VP of Finance at a women's health brand.
During the session, the agent surfaced a mapping issue involving $377K of Amazon fulfillment costs. The costs were categorized correctly in the accounting system, but had been mapped to "shipping" instead of "fulfillment" by the time they reached the model.
She caught the issue in real time, remapped the accounts, and reimported the data. The channel-level discrepancy disappeared before the meeting ended.
That's the experience I was designing for. Not a quarterly cleanup when something has gone wrong, but a system a finance team can use the moment a number looks off.
What shipped
Self Serve Mapping
Customers can map their own accounts directly from the Account Mapping experience, without CS setup or scheduling.
Channels that reflect how brands actually sell
Customers can create custom channels for businesses like Costco, TikTok Shop, and regional distributors instead of being forced into a fixed taxonomy.
Financial dimensions preserved
Revenue and expenses by department, location, or class flow through from QuickBooks and NetSuite without being flattened into a generic P&L.
An agent that improves with use
Corrections become signals the system can learn from, allowing mapping to improve as a brand grows.
What I'd do differently
Tune confidence thresholds from real usage
We initially set the agent's confidence thresholds partly through domain judgment. With live customer data, I would have used the actual distribution of successful and corrected mappings to continuously tune where the agent should act versus ask.
Design the ongoing state earlier
Most of the early product effort went into first-run mapping. But mapping doesn't end at onboarding, new accounts appear continuously.
The next iteration would treat ongoing maintenance as a first-class experience, with the same attention given to initial setup.
© 2025 melissa neira