Capital One AI Strategy: A Bank Rebuilt to Run on AI
How Capital One approaches AI: a technology rebuild for AI, in-house models, teams of agents in fraud servicing and car buying, and no published results. Updated November 2026.
US bank and card issuer: credit cards, consumer and commercial banking, and the Discover payment network
Capital One’s AI page sets out its approach: customer-driven, proprietary AI at scale. It sets no targets.
Read the company sourceWhere Capital One Is Going with AI
Capital One rebuilt its technology from the ground up so that it could run on AI, and now builds its own AI rather than buying it. It fine-tunes open models on its own data, runs teams of AI agents in fraud servicing and car buying, and checks every agent with an evaluation system built for the purpose. CEO Richard Fairbank says AI will transform the customer experience, but Capital One does not intend to put a number on the benefits.
The rebuild was for AI. Founder, Chairman and CEO Richard Fairbank said in October 2026: “the objective function of the rebuild was AI”, describing a technology overhaul that began in 2013. He added: “We did not predict generative AI, and we did not predict agentic AI”, but had the bank known, “we would have built exactly what we built” (third-quarter 2026 call, via AIM). He said: “AI will be transformative for the customer experience.”
The focus has four parts.
- Fraud servicing: a team of specialized agents that understands, reasons about, checks and explains complex fraud calls, supporting several hundred service agents (VentureBeat, August 2026)
- Car buying: Chat Concierge, a customer-facing assistant on dealer websites that compares cars, estimates trade-ins and books test drives (VentureBeat)
- Employees: a generative AI knowledge tool built on the bank’s own documents, and AI assistants for about 20,000 customer service agents (Capital One, Euromoney, February 2025)
- Business payments: Brex, bought for about $5 billion in April 2026, which brings a suite of agentic AI products (American Banker)
What It Takes: A Chief Scientist, In-House Models and Agents That Are Always Evaluated
A chief scientist at the top. Prem Natarajan, executive vice president, Chief Scientist and Chief Enterprise AI Officer, leads enterprise AI (Capital One). He favors a careful rollout: “The first step is getting a baseline solution to everyone”, with pilots that grow from about ten people to a few hundred, then a few thousand (Euromoney).
Build, do not buy. Capital One runs a central AI platform with governance built in, customizes open-weight models with its own data and has built its own system for running teams of agents. Kel Vanee, a managing vice president of machine learning engineering, said: “At Capital One, we’re not just using AI, we’re building AI” (VentureBeat).
Every agent is evaluated. Maulin Patel, managing vice president of product for AI and ML platforms, said: “At that scale we cannot govern autonomous AI agents just by raising chats.” The bank scores agents continuously, uses AI judges calibrated by people, and has subject matter experts audit complex financial logic (Constellation Research, September 2026). He said: “Human oversight is super important, especially in regulatory industries.” It is a strong example of why agents need managers.
Headcount grew through deals. Capital One had about 76,300 employees at the end of 2025, up from 52,600 a year earlier, after it bought Discover in May 2025, and about 77,500 in September 2026, after it added Brex. It has not linked headcount to AI.
AI fluency, coded against the Lead with AI Fluency Matrix (September 2026 edition):
- Fluency floor: Not published. AI assistants reach about 20,000 service agents, and an AI knowledge tool serves employees, but Capital One has published no training requirement or usage figure.
- Highest level evidenced: Level 3 (Integrating), company-claimed. Teams of AI agents are built into fraud servicing and car buying, with evaluation and human oversight. Capital One has published no results for them, so Level 4 is not evidenced.
- Portfolio design: Platform-led and research-heavy. One central platform with built-in governance, in-house research, and models trained on the bank’s own data.
How Capital One Measures Results
Capital One reports no AI results. Fairbank said in October 2026: “We certainly do not intend to quantify the benefits” (AIM). The bank describes its agents in detail but has published no figures for accuracy, time saved or cost. It uses AI to monitor credit in some areas, where Fairbank said: “Instead of manual, it is automated,” while calling it early days.
The gap is by choice. Capital One treats its AI as a competitive advantage and keeps the numbers private. That choice limits outside judgment of the step from AI adoption to AI value.
Next markers: any published result from the fraud agents or Chat Concierge, the use of Brex’s agentic products across business payments, the completion of the Discover migration, and a training or usage figure for employees.
Change log
- 2013: Capital One begins rebuilding its technology stack.
- February 2025: Chat Concierge is live as the bank’s first customer-facing agentic AI.
- May 2025: Capital One completes its purchase of Discover.
- April 2026: Capital One completes its purchase of Brex.
- August 2026: The bank describes its multi-agent fraud servicing system.
- October 2026: Fairbank says the rebuild was for AI and declines to quantify the benefits.
The Bottom Line
- Disclosure: 3 of 4. Capital One answers where it is going, where it will focus and what it will take. It has chosen not to report results.
- The foundation came first, with a technology rebuild that began in 2013.
- It builds its own AI, on open models trained on its own data.
- Agents are evaluated continuously, with people calibrating the automated judges.
- The say-do gap is disclosure. Capital One calls AI transformative but says it will not quantify the benefits.
- Watch Brex. The test is whether agentic products spread across business payments.
Revenue, Headcount and Revenue per Employee
What the numbers show: Before Discover, from 2022 to 2024, revenue rose 14% while headcount fell 6%, so revenue per employee rose 21%. The Discover deal in May 2025 lifted year-end headcount to about 76,300 but added only about seven months of revenue, so the 2025 figure of about $700k understates the combined bank. These figures do not show that AI caused any change.
Revenue, headcount and revenue per employee, FY2022 to FY2025
- FY2022: $34.25bn revenue, 55,943 employees, $612k per employee
- FY2023: $36.79bn revenue, 51,987 employees, $708k per employee
- FY2024: $39.11bn revenue, 52,600 employees, $744k per employee
- FY2025: $53.43bn revenue, 76,300 employees, $700k per employee
Change, FY2022 to FY2025: revenue up 56%, headcount up 36%, revenue per employee up 14%.
Notes: Revenue is total net revenue, as reported. Headcount is approximate, at year-end. Capital One bought Discover on May 18, 2025, so 2025 revenue includes about seven months of Discover while year-end headcount includes all of its employees, which lowers 2025 revenue per employee. Capital One also bought Brex in April 2026.
These figures give context on scale and productivity. They do not show that AI caused any change. Sources: Capital One 10-K filings.
How Capital One Compares with Financial Services Peers
We looked at how Capital One compares with other financial services companies in the index. Here is how they stack up.
| Company | Disclosure | AI owner | Fluency floor | Governance | AI value reported | Revenue per employee |
|---|---|---|---|---|---|---|
| Allianz | 3 of 4 | No named AI owner; board COO covers IT | Level 1, by participation | Eight principles and a register of every use case | Results by use case, no total | $1.35M |
| American Express | 3 of 4 | EVP, Enterprise Data and AI | Not published; tools for nearly all colleagues | Agents only through approved API or MCP channels | No: coding time only | $940k |
| Bank of America | 4 of 4 | CTIO; no Chief AI Officer | Not disclosed | Controls, no published principles | Yes: $250M to $300M in 2026, with a 2027 figure | $531k |
| BBVA | 4 of 4 | AI Transformation area at the top of the bank | Level 1, training required before access | Eight published principles | Results by process and agent | $328k |
| Berkshire Hathaway | 1 of 4 | None at group level; each business decides | Not published | Not published | None published | $958k |
| BlackRock | 2 of 4 | No firmwide AI leader named; Aladdin engineering builds the tools | Not published | Guardrails built into the agent platform | None published | $973k |
| BNY | 4 of 4 | Chief Data and AI Officer | Level 1, near-universal training | Every digital employee has a human manager; reviews before release | Speed results by use case, no total | $417k |
| Capital One | 3 of 4 | Chief Scientist and Chief Enterprise AI Officer | Not published | Every agent evaluated continuously, with human oversight | No: declines to quantify benefits | $700k |
| Charles Schwab | 3 of 4 | CIO for technology, operations and data; no AI head named | Not published | Guardrails and human handoffs; outside agents cannot move money | No: counts of use cases only | $725k |
| Citigroup | 4 of 4 | Group Head of AI, in the COO’s organization | Level 1, required prompt training | Controls built into the AI platform, human in the loop | Capacity and speed results; says benefits exceed costs, no figure | $377k |
| Goldman Sachs | 3 of 4 | CIO, sponsored by the top three executives | None declared | Firmwide AI risk committee | No: weekly use only | $1.23M |
| JPMorgan Chase | 4 of 4 | Chief Data and Analytics Officer, passing to the CTO | None declared (use is opt-in) | Little published | Earlier savings estimate, no longer updated | $573k |
| Lloyds Banking Group | 4 of 4 | Chief Data and AI Officer | Level 1, mandatory training | Responsible AI team and framework | Yes: about £50M in 2025, £100M+ target | $377k |
| Mastercard | 3 of 4 | Chief AI and Data Officer | Level 1, data training required before access | Every model explainable, governed and monitored | No: few internal results | $824k |
| Morgan Stanley | 3 of 4 | Head of Firmwide AI | Not published; 98% of adviser teams use the AI assistant | Evaluations before every launch; human oversight | Results by use case, no total | $851k |
| U.S. Bancorp | 4 of 4 | Dedicated AI teams; no single AI leader named | Not published; AI courses offered to all | Not published | Results by use case, no total | $417k |
| Visa | 4 of 4 | President of Technology | Not published; generative AI access for all since 2023 | Data-use principles built into model monitoring | Product development results, no total | $1.32M |
| Wells Fargo | 3 of 4 | Head of AI, reporting to the CEO | Not published; 90,000+ trained | Business heads accountable; rules for customers’ AI agents | No: coding efficiency only | $408k |
| Klarna AI Pacesetter | 3 of 4 | CEO | Not disclosed | None published | Yes: savings and work done by AI | $1.24M |
Disclosure counts how many of the four strategy questions (where it is going, where it will focus, what it will take and how it will know it is working) a company answers in public. Fluency floor is the lowest level of AI fluency a company expects of all its people, on the five levels of the Lead with AI Fluency Matrix: Avoiding (0), Chatting (1), Collaborating (2), Integrating (3) and Transforming (4). The other columns summarize each company’s profile. Revenue per employee is for each company’s latest fiscal year, in US dollars. Rows marked AI Pacesetter or Notable are companies outside the S&P 100 and STOXX Europe 50. They are shown for contrast and are not counted in any index figure. See all companies in the AI Strategy Index.
Frequently Asked Questions
What is Capital One’s AI strategy?
Capital One rebuilt its technology so that it could run on AI and builds its own AI on open models trained on its data. It uses teams of AI agents in fraud servicing and car buying.
Who leads AI at Capital One?
Prem Natarajan, executive vice president, Chief Scientist and Chief Enterprise AI Officer. Founder and CEO Richard Fairbank sets the direction.
How does Capital One govern AI agents?
It evaluates agents continuously, uses AI judges calibrated by people, and has subject matter experts audit complex financial logic.
Does Capital One report AI results?
No. Fairbank said in October 2026 that the bank does not intend to quantify the benefits of AI.
What is Chat Concierge?
It is Capital One’s customer-facing AI assistant on car dealer websites. It compares cars, estimates trade-in values and books test drives, using several agents that plan, check and explain.
How is this profile put together?
Lead with AI profiles use public sources only, such as annual reports, earnings materials, filings and on-record interviews, dated at the time of writing. AI fluency is coded against the Lead with AI Fluency Matrix. Company figures are marked as company-claimed unless they have been independently verified.
About the AI Strategy Index
The Lead with AI AI Strategy Index profiles how the largest companies in the US and Europe approach AI, starting with the S&P 100 and the STOXX Europe 50. Each profile answers four questions: where the company is going with AI, where it will focus, what it will take and how it will know it is working. AI Pacesetters and notable companies outside the two indexes are profiled for contrast and are not counted. Profiles use public sources only, and AI fluency is coded against the Lead with AI Fluency Matrix.
See all company profiles