Capital One says an audit assistant cut errors 85% and saved 5,000 hours

Capital One says an AI assistant for its audit, risk and compliance teams cut quality errors by 85% and saved an estimated 5,000 hours of audit time across 30 auditors in a few months. The figures come from Prem Natarajan, the bank's chief enterprise AI officer, in a column by Thomas H. Davenport and Randy Bean published by MIT Sloan Management Review on 28 October 2026. It is a first-party account relayed by the column's authors, not an independent evaluation.

What Capital One says it built

• Audit, risk and compliance: an intelligent assistant for risk associates, built with open-weight models, advanced analytics and analytical AI. Natarajan said it supports faster risk assessments, script review, root cause analysis and quality control.
• Reported results: quality errors down 85% and about 5,000 hours saved across 30 auditors in a few months. One audit task fell from five hours to three minutes, according to the column.
• Call summarisation: agentic AI splits a summary into subtasks, one per agent, to interpret complex customer issues, particularly on fraud-related calls. Summaries are added to the customer record.
• Customer service: generative AI surfaces the right answers to human agents faster.

The detail worth copying

Natarajan said the summarisation process only reached satisfactory performance once the bank added a reasoning agent to it. The column reports that several hundred agents now use the tool and have seen substantial productivity gains, and that early adopters saw a "massive improvement in customer experience", in Natarajan's words. No time saved, accuracy or satisfaction figure is given for this use case.

How the bank frames adoption

Natarajan told the authors that "you can't achieve transformation if you're focused on quarterly ROI", and said the bank views short-term ROI measurement as premature for a technology changing this fast. He also said most initiatives target enterprise and line-of-business capabilities rather than individual productivity, and that the bank invests heavily in training for each new use case: "It's not 'Here's a tool — go use it.'"

Chief data officer Amy Lenander said outputs improve when data is organised. According to the column, the bank is curating unstructured data such as call transcripts, manages it as "data products" owned by the teams closest to it, and uses knowledge graphs to map relationships between entities. A central team she oversees checks that the data products fit together.

What the column leaves out

The column does not say which open-weight models are used, what the 85% is measured against, how many audits or errors were sampled, what the system costs to run, or whether auditors review the assistant's output before it is used. It describes the 5,000 hours as an estimate. The authors say they spoke with Natarajan and Lenander.

The column also notes that Capital One now sells Databolt, a tool built by its software team that tokenises sensitive data so organisations can use it with large language models without exposing personally identifiable information. That is a product the bank markets externally, which is context for how it describes its AI work.

What to do

• If you automate audit or QA review, define an error before you start, count errors on a fixed sample before and after, and publish the denominator with the percentage.
• If a summarisation or extraction pipeline underperforms, test splitting it into subtasks and adding a reasoning step, as Capital One says it did, and compare cost per summary as well as quality.
• Ask each vendor or internal team quoting hours saved whether the figure is measured or estimated; this column calls its figure an estimate.