Reversal rate: the AI finance metric CFOs ignore

September 25, 2026 · 5 min · Free, with full transcript

Bill Gates has asked governments to act on AI risks he believes could cause mass casualties. Does that headline tell a finance chief anything about their own exposure? The position here is no. The biosecurity scenario is real but separate from the failure that will actually embarrass a CFO in 2026: a model quietly auto-approving payments at a scale the old fraud controls were never calibrated for, including a European logistics firm that paid 400,000 euros to an insolvent supplier. You come away able to take apart the AI finance automation budget line into its three parts, defend document ingestion, challenge the autonomous approval band, and replace automation rate with reversal rate plus a circuit breaker on the auto-approved queue.

Listen to the podcast

4 min

Chapters

Key takeaways

  • Keep document ingestion in the budget: one mid-market manufacturer cut invoice processing from 11 minutes to under two.
  • Treat the autonomous approval line, such as AI approving payments up to 50,000 without review, as the most dangerous item in the plan.
  • Judge models on correlated error, not accuracy, since a model repeats the same wrong decision a thousand times before anyone notices.
  • Add reversal rate to the dashboard and aim for under half a percent on the auto-approved band before approving more spend.
  • Set a hard circuit breaker so any single reversal above a threshold pauses the automated queue until a human signs off.
Read the full transcript

Host:You're listening to Leaders Insights. Today's subject, Gates warns of a billion deaths, but the real AI risk in finance is subtler than that. The mistake I see everywhere right now is CFOs reading the Gates headline about a billion deaths and then doing absolutely nothing to their own controls. Because the number is so big it feels like someone else's problem.

Expert:And that's exactly the trap. Gates is talking about biosecurity, AI helping someone design a pathogen. Real worth a government response. But it has nothing to do with the thing that's going to embarrass a finance chief in 2026, which is far more boring and far more likely.

Host:So what does embarrass you then?

Expert:A model quietly approving invoices it shouldn't. Let me take apart an actual artifact, ease the AI finance automation line most companies put in their 2026 budget. I've reviewed a dozen of these. They all have the same three components and two of them are wrong. Start with the component that works. The document ingestion piece, the software that reads a supplier invoice and pulls out the numbers. That part genuinely earns its keep. At a mid-market manufacturer I advised, it cut invoice processing from 11 minutes to under two. Vendor benchmarks from the likes of Apsen and Ramp claim similar. And I'd normally discount vendor numbers heavily. But this one I watched happen. It's mature, it's cheap, it's fine.

Host:And the piece that isn't fine?

Expert:The autonomous approval line. This is where the budget says AI approves payments up to 50,000 without human review. Everyone loves it because it looks like headcount savings. It's the single most dangerous line in the plan.

Host:Why if the model's accurate?

Expert:Because accuracy isn't the risk. Consistency is the risk. A human approver has bad days, but broadly random errors. A model has correlated errors. It gets the same wrong idea a thousand times before anyone notices. In 2025, a European logistics firm led a model auto-approve recurring vendor payments, and it kept paying a supplier that had quietly gone insolvent because the invoice pattern looked normal. 400,000 euros out the door before a human looked.

Host:That's not the model being stupid though. That's the model being obedient.

Expert:That's the whole point. And it's the thing the gates framing hides. The danger isn't a rogue superintelligence. It's a competent, cheerful system doing precisely what you asked, at a scale where your old controls assumed a tired human would catch it. Your fraud checks were calibrated for human throughput. The machine outruns them. The monitoring dashboard. And here's the con. It usually reports the wrong metric. Every one of these dashboards leads with automation rate. The percentage of transactions the AI handled without a human. Finance teams brag about hitting 94%. The reversal rate. The share of AI decisions a human later had to undo. That's the number that tells you if the thing is actually trustworthy. If your automation rate is 94% and nobody's tracking reversals. You don't have a dashboard. You have a mirror telling you you're pretty.

Host:That's a little cocky. Give me the number a good team hits.

Expert:Under half a percent reversal on the auto-approved band. And, uh, this is the part people skip. A hard rule that any single reversal above a threshold pauses the whole automated queue until a human signs off. Circuit breaker. Not a suggestion. It trims them and good. The pitch of fire the account's payable team was always fantasy. What you actually get is the same team supervising 10 times the volume, spending their hours on the 10% of transactions that are genuinely weird. That's a better job and a safer books. The savings are real, just smaller and more honest than the deck claimed.

Host:So the gates warning useful or noise for a finance leader. Useful as a reminder that

Expert:regulators are about to get twitchy. Noise is a guide to your own risk. Your billion death scenario is a rounding error you didn't catch because the dashboard flattered you.

Host:Give me the one thing to do Monday. Open your automation dashboard. If it shows automation

Expert:rate but not reversal rate, add reversal rate before you approve another dollar of that

Host:budget line. That's the whole job this week. This episode draws on accounting today CFO dive, Financial Times. That's it from us. The reading continues at MBA-training.com, new CFO analysis every day.

Read and go deeper