Capital One audit assistant: 85% fewer quality errors, 5,000 hours saved

Thirty auditors at Capital One saved an estimated 5,000 hours "in only a few months" with an AI assistant for audit, risk and compliance work, and the assistant cut quality errors by 85%, the bank's chief enterprise AI officer, Prem Natarajan, told the authors of a [column in MIT Sloan Management Review](https://sloanreview.mit.edu/article/ai-is-thriving-at-capital-one/) published on 28 October 2026. One audit task went from five hours to three minutes. The column gives no cost and no baseline for the error figure.

What the assistant does

The column, by Thomas H. Davenport and Randy Bean, describes "an intelligent assistant for risk associates" built by the bank on open-weight models (AI models whose files a company can download and run on its own systems, so the data need not go to an outside vendor's service) alongside the bank's existing analytics. It lists what the assistant helps with: quality control, script review, root cause analysis and "data-driven assurance". It does not say which audit task fell from five hours to three minutes, or which models are used.

The numbers, and what they are a share of

• 5,000 hours across 30 auditors is about 167 hours each (5,000 ÷ 30). The period is "a few months"; over three months that would be roughly 13 hours a week per auditor, over six months about 6. The column calls the 5,000 an estimate and does not say who made it or how.
• Five hours to three minutes is 300 minutes to 3, a hundredfold cut on one task. How often auditors do that task is not stated.
• "Reduced quality errors by 85%" has no starting count, no definition of a quality error and no period.

The support desk got call summaries, with one lesson

In customer service, "several hundred agents" now use AI call summaries that are added to the customer record, the column says, with an emphasis on fraud calls. Natarajan said the summaries became satisfactory only after the bank added a reasoning step: a second AI pass that interprets the customer's problem before the summary is written. The column reports "substantial productivity gains" and gives no figure.

The people side and the catch

On people, the column has one statement. "It's not 'Here's a tool — go use it,'" Natarajan said; the bank "invests heavily in training for any new use case". It does not say how auditors were trained, who checks the assistant's output before an audit finding is signed, or whether the team's size or workload changed.

The catch is in Natarajan's own words: "You can't achieve transformation if you're focused on quarterly ROI." He told the authors it is too early to measure short-term return on AI. So the hours are not set against a cost, and nobody outside the bank has checked them. The column is an interview with two executives, in a series about successful uses of AI; it reports no failures.

Capital One was the eighth-largest US bank by assets in September 2026, the column says, with its own engineers and decades of work on its data. A compliance team that buys its software should expect to get the assistant from a vendor, at a price the column cannot help with.

What to do if you run compliance or internal audit

• Pick one repetitive testing or review task, time it for a month, and count the errors your quality reviewers catch in it. That is the baseline the 85% and the five hours lack.
• Ask any vendor quoting a figure like this for the error count before and after, the period, and who counted.
• Write down who signs off an AI-assisted test before it enters the audit file, and hand that rule to your head of audit or chief compliance officer before a pilot starts.