5 lessons from putting AI to work with enterprise finance teams

We’ve been hard at work since we launched Ramp’s AI Solutions team back in June. We help our enterprise partners build and implement AI workflows across systems they already use, drawing on data they can actually trust. Our clients’ asks may vary, but common themes have emerged.

After a few months of engagements, the team got together at an offsite to ask: what have we learned so far? Our AI Strategists and FDEs compared examples from both client work and internal implementations with the goal of teasing out important parallels and articulating what moves the needle when our team gets involved.

These conversations produced five lessons. I’m sharing them to help other teams decide where to start and what to look for as they build. We used examples wherever possible to illustrate a frequent theme, but these learnings are meant to be directional and extrapolated beyond one specific problem or client. Happy building!

1. Define the question you want AI to answer

A specific business question helps focus the workflow on the evidence your team needs to make a decision.

• We worked with a distributor’s procurement team that wanted to know whether it was paying suppliers the prices it had negotiated. The company already matched invoices to purchase orders, but an outdated rate could appear on both and pass the check.
• Procurement was still sampling purchases against its agreements. We focused the workflow on that task: connect each invoice line to the applicable contract terms and give buyers the evidence to investigate a possible pricing discrepancy.
• The review queue included a reviewed calculation and the supporting contract term for each possible discrepancy. Buyers checked the flags against the agreements and identified questions to take back to suppliers. They had the evidence for those discussions in the queue; whether a flagged charge warranted recovery still needed to be resolved.

2. Give AI the business definitions behind the numbers

Finance workflows allow for a 0% margin of error. For AI to produce a useful comparison, it needs the team’s agreed definitions of what goes into each number and how shared costs are treated.

• A software company’s finance team wanted to compare customer acquisition cost across products. Different workbooks allocated shared costs differently. Payroll, marketing, and travel costs also sat in separate systems, and sales and finance used different customer identifiers.
• We worked with finance to define the calculation as we connected those records. The team had to agree which costs belonged in the comparison and how to divide shared costs. We made the approved allocation rules available to the agent. Repeatable calculations produced the amounts, which the agent could explain and trace to their inputs.
• During review, finance followed an unexpected difference back to a shared-cost assumption, corrected its treatment, and reran the comparison. The explanation gave them a way to investigate the number. With the assumption corrected, they could discuss products on a consistent basis without reconstructing the analysis in another workbook.

3. Check the evidence before approving an AI recommendation

Before acting on an AI recommendation, reviewers need to understand which constraints it satisfies and why alternatives were ruled out. The workflow should also specify who can approve the recommendation and who has authority to carry it out. A prompt saying “be careful” is not an approval process.

• During a treasury review, the team wanted to understand why a lower-cost funding option had been excluded. They needed to check whether the cheaper option was available under the agreements.
• Analysts had been combining asset spreadsheets with agreements, eligibility rules, and facility limits to assemble each scenario. An asset could fail the eligibility requirements, or a proposed allocation could exceed a facility’s capacity, regardless of the rate.
• We brought those inputs into a workflow that prepared scenarios with reviewed calculations, remaining capacity, and explanations of why assets qualified or failed. Treasury could trace the exclusion and check feasible alternatives against the agreements, then take its recommendation through approval. Execution remained with authorized treasury staff.

4. Measure the work AI leaves for your team

The value of a workflow depends on how much checking, correction, and follow-up it leaves behind, not just how much output it produces.

• Even when the data and rules are in place, the workflow can still leave the team doing too much work. We saw that in bank reconciliation: too many proposed matches sent accountants back into the records to investigate, even on cases they had already completed.
• For this multi-entity services business, we had grouped related ledger entries and attached evidence to proposed matches. Matching amounts alone missed timing differences and payments covering multiple entries. A false match could conceal an accounting difference. Sending everything to an exception queue would leave accountants to investigate it all anyway.
• We tightened the rules and separated routine timing differences from unexplained discrepancies. On the next review, accountants worked through a smaller, better-supported exception queue and could trace accepted matches to the underlying records.
• We still needed to measure the time spent checking, correcting, and following up, along with the cost of running the workflow. Testing with the accountants showed us which parts of the reconciliation still required investigation and where to focus the next changes.

5. Test rule changes on the cases they should and shouldn’t affect

When a business rule changes, testing should confirm both that the right cases receive the new treatment and that unrelated decisions stay intact. For your first workflow, stay close to the person who will use the output. Ask them to show you where it helps, where they still have to redo the work, and what changes from one case to the next. That gives you a practical way to decide what to improve and when the workflow is ready to take on more.

• A reviewer caught an expense rejection based on an outdated allowance. The allowance applied to a particular client and effective date, so the correction needed to preserve how other requests were handled.
• The services company processed expenses and supplier invoices under client-specific policies, amendments, and exceptions. We had converted approved guidance into versioned rules and linked recommendations to policy evidence. We could trace the rejection to its source and identify which version needed changing.
• We updated the approved allowance for the affected client and effective date, then reran previously reviewed cases. The next run applied the amendment to eligible requests while leaving other clients’ decisions governed by their own policies.

What are you working on?

What is taking too much time in your finance team? What have you tried, and where are you getting stuck? We’d be glad to hear about the work itself.