Om Agarwal is the Founder and CTO at Minerva Intelligence Inc.
When I ran month-end accounting for my own business, it looked the same every time: a dozen bank tabs, invoices that wouldn’t parse and a spreadsheet that was always one column short.
After automating one too many reconciliations by hand, I wrote down the loop I kept repeating—observe, decide, act, verify—and had an agent follow it. It wasn’t perfect on day one, but it was reliable enough that reviews shifted from “hunt and fix” to “spot-check and approve.”
Since then, I’ve treated finance agents as quiet infrastructure. They don’t replace judgment—they make it easier to exercise. The broader trend backs this up: Finance leaders report AI usage in the function jumping from 37% to 58% year over year. At the same time, maturity is uneven—only about 1% say more than three-quarters of their finance processes are automated—so the opportunity for practical wins is still wide open.
What Changed For Me (And Why Business Owners Should Care)
The first change was cash clarity. When reconciliations ran daily instead of weekly, I wasn’t flying blind on spend. Next was payables: Invoices arrived pre-coded, tax detected and PO tolerances checked, so my role shifted from firefighter to reviewer. The last change was the calm that comes with evidence. Entries carried the source file, the rationale and a clean reversal path, which turned audit from a panic into a download.
I also felt the macro push. In a tougher labor and cost environment, “do more with the same team” stopped being a slogan and became a requirement. Recent CFO surveys reflect that shift: Nearly two-thirds say automating employee tasks is now a strategic priority, and most of those planning to automate expect to use AI in the next year.
That’s encouraging, but it also heightens the cost of bad data. I learned early to normalize vendors, memos and bank descriptors because the rework tax compounds at scale. Forrester’s 2024 analysis found that many organizations estimate losses above $5 million annually due to poor data quality.
The Owner-Friendly, Five-Step Pilot Plan I Wish I’d Had
1. Measure a baseline. Write down where the hours go (recs, coding, aging, flux), your cycle time and exception volume. Pick two workflows with tight acceptance criteria.
2. Codify policy. Turn mappings, approvals, tax/FX rules, thresholds and memo formats into versioned rules with tests. Ambiguity in policy becomes ambiguity in the books.
3. Run in draft mode. Let the agent propose entries and matches; keep human review until boredom sets in. Optimize for precision first; trust compounds when reviewers have nothing to fix.
4. Make auditability default. Attach source evidence, show rationale and confidence, keep immutable change logs and make reversals explicit.
5. Report results in business terms. Track percent auto-approved, exception reasons, hours saved and cycle-time deltas—those are the metrics leaders actually act on.
What To Start With, What To Avoid And How This Scales
I’ve had the best results starting where the evidence is clean and “done” is unambiguous. Bank recs are the classic early win; cash application tends to follow. Accounts payable intake is next—vendor normalization, tax detection and PO tolerance checks remove a surprising amount of rework.
As those pieces settle, the benefits spill into planning: Earlier, cleaner actuals shrink the gap between operations and finance. Benchmarks keep pointing finance leaders in the same direction. Fewer days spent closing means more days advising the business (registration required).
Here are some common traps to avoid:
1. Speed Without Design: Agents juggle files, portals and models; without batching, caching and parallelization, you’ll feel “AI lag” and blame the idea instead of the implementation.
2. Edge Cases Without Context: When confidence dips—new vendor, odd memo—escalations must include the exact evidence, top candidate codings and the “why,” so review is a decision, not a hunt.
3. Controls Bolted On Later: Treat policy as a controlled artifact, keep roles and approvals clean and assume an auditor will read your logs.
4. Ignoring The Browser: APIs don’t cover everything. A hardened, allow-listed browser approach—with full action traces—bridges legacy portals without compromising governance.
5. Skipping Data Hygiene: If you don’t normalize vendors and descriptors early, the rework cost grows with every new entity or integration.
The Takeaway
Agents create the conditions for better judgment. Start with one loop you run every month, write your rules as if an auditor will read them (they will) and measure what changes. The market is moving in that direction—adoption up, maturity uneven—which is why small, credible wins beat grand promises right now.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?