Verified
Capital One
Scaling- Banking and payments
- ~76,300 employees
- Proprietary AI built in-house
- Runs principally on AWS
- Evident AI Index #2 (2026)
Capital One builds its own AI instead of renting it: a central enterprise AI group under Prem Natarajan builds proprietary multi-agent systems on customized open-weight models such as Meta's Llama, on a technology stack that runs principally on AWS after the bank closed its last data center in 2020. Capital One uses AI in fraud, credit, marketing, servicing and software security, but Richard Fairbank said on Oct 20, 2026 that the company does not intend to quantify AI's benefits.
capitalone.comTheir approach to AI
- The platform
- Capital One built a proprietary multi-agent framework for the Chat Concierge car-buying assistant and reused it for customer service calls about fraud (Mar 2026); a central engineering team builds the enterprise platforms its agents run on (Sep 2026).
- The models
- Capital One customizes open-weight models in-house, building Chat Concierge on Meta's Llama and distilling fine-tuned models into smaller ones (Oct 2026), and wrote its VulnHunter code-security agent for Anthropic's Claude Opus 4.8 in Claude Code (Jul 2026).
- The reach
- Capital One launched an AI learning hub to more than 60,000 associates in 2026 and drew over 10,000 engineers to a live agentic coding session (Aug 2026), but has not published how many employees use AI tools.
- The catch
- Richard Fairbank said on Oct 20, 2026 that Capital One does not intend to quantify AI's benefits; its one AI dollar figure, ProtectID's nearly $150 million of auto lending fraud stopped in 2025, is a gross amount, not a reported saving.
Snapshot
- Employees
- ~76,300
- Dec 31, 2025, before the Brex acquisition (2025 Form 10-K)
- Technology budget
- Not disclosed
- Richard Fairbank calls for significant and sustained investment in technology and AI (Mar 2026)
- AI/ML headcount
- Hundreds of AI hires
- Recruited from global technology companies (Mar 2026); no total published
- Who owns AI
- Prem Natarajan
- EVP, Chief Scientist and Chief Enterprise AI Officer; joined from Amazon in 2023
- Platform
- Proprietary multi-agent framework
- Presented by AI Foundations head Milind Naphade at NVIDIA GTC (Mar 2026)
- Models
- Customized open-weight models
- Meta Llama under Chat Concierge, distilled into smaller models; Anthropic Claude Opus 4.8 for VulnHunter
- Cloud
- Principally AWS
- Last data center closed in 2020; Discover's network still partly in data centers
- Reach
- 60,000+ offered AI training
- AI learning hub launched to over 60,000 associates in 2026; no count of employees using AI tools
Key findings
Richard Fairbank wrote in Capital One's 2025 annual report, filed Mar 25, 2026, that while many companies rent AI capabilities, Capital One builds its own, has recruited hundreds of AI professionals from leading technology companies, and applies AI in software development, analysis, marketing, servicing, credit, fraud and cyber.
Richard Fairbank told analysts on Oct 20, 2026 that Capital One does not intend to quantify AI's benefits, and Capital One has published no business return credited to AI; its one AI dollar figure, nearly $150 million of auto lending fraud stopped by ProtectID in 2025 (Oct 2026), is a gross amount stopped, not a loss Capital One says it avoided.
Evident's 2026 AI Index, published Oct 6, 2026, ranks Capital One second among 50 banks behind JPMorgan Chase and first on talent and innovation, but 40th on its leadership pillar, which measures public communications and strategy, down 20 places.
Capital One built a proprietary multi-agent framework for Chat Concierge, a car-buying assistant on Meta's Llama, and reused it for customer service calls about fraud (Mar 2026); Prem Natarajan said in Oct 2026 that distilling fine-tuned open-weight models into smaller ones cut Chat Concierge's latency fivefold.
Capital One runs principally on AWS after closing its last data center in 2020, while Discover's payment network still runs partly in data centers; 100% of new Discover card originations were on Capital One's technology stack by the end of Sep 2026, with the back book due by the first quarter of 2027.
Where Capital One sits
- Stage
- Scaling
Capital One sits at Scaling: a named executive, Prem Natarajan, has owned enterprise AI since 2023; AI runs in production in fraud, credit, marketing, servicing, auto lending and code security; a proprietary multi-agent framework built for car buying is being reused for fraud servicing calls; and Capital One launched an AI learning hub to more than 60,000 associates in 2026. Capital One meets the Scaling test only in part, because it has published no business return credited to AI, and falls short of Embedded because it has published no count of employees using AI and Richard Fairbank said on Oct 20, 2026 that it does not intend to quantify AI's benefits.
How the stages are defined - Ownership
- Prem Natarajan, Chief Enterprise AI Officer, within Frank LaPrade's Enterprise Services
- Platform
- Built: Proprietary multi-agent framework built in-house
- Models
- Customized and distilled open-weight models such as Meta Llama; Anthropic Claude for code security
- Hosting
- Principally AWS, with some Azure and Google Cloud resources; Discover's network partly in data centers; inference location not published
- Reach
- AI learning hub launched to 60,000+ associates; tool use not disclosed
How they deploy
Build AI instead of renting it
Richard Fairbank wrote in Mar 2026 that while many companies are renting AI capabilities, Capital One is building its own, drawing on its heritage as a data and technology company. Fairbank told analysts in Oct 2025 that many companies will bring in third-party AI applications, but that changing the business model of banking requires AI embedded in a company's technology, operations, processes, risk management and customer experiences, and that this requires significant investment in AI and AI talent.
A technology rebuild aimed at AI
Capital One declared a technology transformation in 2013, hired thousands of engineers, product managers, designers and data scientists, rebuilt its data ecosystem and moved fully into the cloud. Richard Fairbank said on Oct 20, 2026 that AI was the objective of that rebuild; Capital One did not predict generative or agentic AI, but Fairbank said it would have built the same stack had it known.
Reinvent work, not only improve it
Richard Fairbank said on Oct 20, 2026 that every company will use AI to improve work as it is done today, but that reinventing how work is done needs a modern technology stack and an organization able to run changes that cut across business units. Fairbank named the data ecosystem, software development, credit, marketing, risk management, fraud, operations, servicing and cybersecurity as areas inside Capital One that AI will change.
Build once, reuse across the bank
Capital One's multi-agent framework was built for the Chat Concierge car-buying assistant and later applied to customer service calls about fraud. Capital One describes the approach as build once, reuse extensively, and encourages associates anywhere in the bank to find manual processes and build solutions from its centralized technology stack (Apr 2026).
Start agents where the risk is low
Prem Natarajan told Fortune in Dec 2025 that Capital One wanted to start agentic AI at the low end of the risk spectrum while still choosing use cases with real impact; its first proprietary agent, Chat Concierge, serves car buyers and dealers. Natarajan said Capital One had cut Chat Concierge's latency fivefold since launch and credited the decision to build its own stack, which lets it tune the agents for latency.
Put the controls in the platform first
Rashmi Shetty, Capital One's VP of enterprise AI, told CIO Dive in Sep 2026 that agentic AI must be treated as an end-to-end system, and that codified policies, runtime controls and security guardrails belong in the platform before developers build agents, because retrofitting governance onto ad hoc agent applications is far harder. Shetty credited Capital One's operational agents to years of investment in governed data pipelines and data lineage.
Buy AI-native software when it fits
Capital One agreed to buy Brex, which it describes as an AI-native corporate card and spend management platform, on Jan 22, 2026 and closed the deal on Apr 7, 2026. Richard Fairbank said Brex runs in-house AI agents for expense management and audit, with procurement, payments and accounting agents to follow, and that Capital One would add investment capacity in engineering and AI.
- Deployment model
- Capital One runs AI as a central enterprise capability that serves its lines of business. Enterprise AI, led by Prem Natarajan, is one of the functions in Enterprise Services, which Frank LaPrade manages alongside Technology, Enterprise Data, Capital One Software and Corporate Security (2026 proxy statement). Milind Naphade's AI Foundations organization, created in 2023, builds foundational AI capabilities for the businesses and houses an applied research job family; Rashmi Shetty, VP of enterprise AI, leads the engineering organization for enterprise platforms that puts generative and agentic AI applications into the retail bank, risk, legal and compliance (Sep 2026); and data science teams sit inside businesses such as the retail bank. Capital One builds most of its AI in-house on its own cloud stack and reuses components, such as the multi-agent framework, across businesses.
- Model providers
- Capital One built Chat Concierge on Meta's open-source Llama model, customized with proprietary data to meet its performance, risk and governance thresholds (Mar 2025), and Prem Natarajan said in Oct 2026 that Capital One has distilled its fine-tuned open-weight models into smaller models that support Chat Concierge. Capital One's NVIDIA GTC sessions in Mar 2026 described customized foundation models combined with multi-agent orchestration, and its knowledge tool for customer service uses a custom retrieval pipeline with models fine-tuned on Capital One data. VulnHunter, the code-security agent Capital One open-sourced in Jul 2026, was written for Anthropic's Claude Opus 4.8 inside Claude Code, though Capital One says the framework could run on other models and coding harnesses. Capital One says it works with Google and Anthropic on hands-on training with frontier models (Aug 2026), but has not published a full list of its model providers.
- Data and infrastructure
- Capital One describes itself as the only major U.S. bank to migrate entirely to the public cloud, and its 2025 Form 10-K names AWS as its cloud infrastructure provider. Richard Fairbank's 2025 letter dates the closing of Capital One's last data center to 2020, while Capital One's technology site gives 2021. Discover's Global Payment Network still runs through a combination of data centers and third-party vendors during the integration, and 100% of new Discover card originations were on Capital One's stack by the end of Sep 2026. A Capital One engineering post from Jun 2026 describes about 350 cloud resource types across AWS, Microsoft Azure and Google Cloud, and Capital One says it uses NVIDIA accelerated computing (Apr 2026) and that reliable access to GPUs matters for AI at scale (Mar 2025), without saying where those GPUs run.
Platforms and internal tools
Multi-agent AI framework2025
Capital One's proprietary framework orchestrates specialized agents that understand a request, plan actions, check the plan against internal APIs, policies and business rules, and explain the result. Capital One built the framework for Chat Concierge and reused it for customer service calls about fraud, which it presented at NVIDIA GTC in Mar 2026.
Enterprise AI platforms
Capital One's enterprise AI engineering organization, led by VP Rashmi Shetty, builds the enterprise platforms used to put generative and agentic AI applications into production in the retail bank, risk, legal and compliance. Shetty said in Sep 2026 that codified policies, runtime controls and security guardrails belong in the platform before developers build agents, with observability and evaluations tracking agent trajectories, tool accuracy and latency.
Chat ConciergeBy Mar 2025
Chat Concierge, Capital One's first proprietary multi-agent assistant, runs on participating dealers' websites, in Capital One's Navigator platform and in dealer CRMs, where it compares vehicles, answers car buyers' questions and schedules test drives.
- Models
- Meta Llama, customized with Capital One data and distilled into smaller models
ProtectIDJul 2025
ProtectID is a fraud screening tool inside Dealer Navigator that runs Capital One's proprietary fraud models on every auto application, alerts dealers in real time and sends flagged applicants an identity check before funding.
VulnHunterJul 16, 2026
VulnHunter is an agentic code-security tool, built internally and released under the Apache 2.0 license, that starts from attacker entry points, tries to disprove its own findings and proposes targeted code fixes. Capital One ran VulnHunter across thousands of its own repositories before releasing it.
- Models
- Anthropic Claude Opus 4.8 in Claude Code
Knowledge retrieval tool for customer serviceBy Mar 2025
Capital One's generative AI tool helps customer service agents answer customer questions, using a custom retrieval-augmented generation pipeline and models fine-tuned on Capital One data to ground answers in its knowledge base.
Transformer-based recommender
Capital One built a transformer-based recommender engine that replaced tree-based models and drives real-time personalized experiences across its digital channels.
DataAgentsJun 2026
DataAgents is a Capital One engineering pattern that pairs an AI agent with an authoritative data product and a human validation loop to classify large catalogs of entities, first used to write dormancy detection logic for cloud resources.
Vendors and partners
Amazon Web ServicesMid-2010s
Principal cloud infrastructure provider · Capital One's 2025 Form 10-K says that, other than the Discover integration, it has moved substantially all of its core infrastructure and customer-facing applications to third-party cloud platforms, principally AWS.
MetaBy Mar 2025
Llama open-source model, the base for Chat Concierge
AnthropicBy Jul 2026
Claude Opus 4.8 and Claude Code, the model and harness behind VulnHunter; hands-on training partner on frontier models · Capital One names Anthropic and Google as AI research labs it partners with for hands-on training with frontier models (Aug 2026).
GoogleBy Aug 2026
Hands-on training partner on frontier models
NVIDIA
Accelerated computing for AI · Capital One gave two sessions at NVIDIA GTC in Mar 2026 on multi-agent systems and on data pipelines for foundation models, and NVIDIA listed VulnHunter among Open Secure AI Alliance members' tools in Aug 2026.
BrexApr 2026
AI-native corporate card and spend management platform, acquired · Brex founder Pedro Franceschi continues as CEO of Brex and reports to Frank LaPrade.
Open Secure AI AllianceBy Aug 2026
Industry group sharing open AI security tools · NVIDIA listed Capital One's VulnHunter among the tools alliance members contribute on Aug 4, 2026, as an alliance working group drafted the Shared AI Findings Exchange guidelines for AI security incidents and the Linux Foundation requested comments.
University of Southern California, University of Illinois and Columbia University
Academic centers of excellence in responsible and generative AI · Capital One funds the USC Center for Responsible AI and Decision Making in Finance, the Illinois Center for Generative AI Safety, Knowledge Systems, and Cybersecurity, and Columbia's Center for AI and Responsible Financial Innovation.
Leadership
AI, data and platform leaders
Executives
AI org map
Central Enterprise AI organization in Enterprise Services; data science teams inside each business
Capital One does not publish an AI org chart. The 2026 proxy statement lists Enterprise AI, Technology and Enterprise Data among the functions managed by Frank LaPrade, Chief Enterprise Services Officer and Chief of Staff to CEO Richard Fairbank, to whom Brex CEO Pedro Franceschi also reports. Prem Natarajan has led enterprise AI since 2023, Milind Naphade built the AI Foundations organization from May 2023, Rashmi Shetty leads enterprise AI platform engineering, and CIO Rob Alexander oversees all technology activities.
Capital One has not published a count of its AI and machine learning staff. Richard Fairbank wrote in Mar 2026 that Capital One has recruited hundreds of AI professionals from leading global technology companies, on top of the thousands of engineers, product managers, designers and data scientists it hired after 2013. Capital One's AI work sits in the central Enterprise AI organization under Prem Natarajan, including Milind Naphade's AI Foundations and its applied research job family and Rashmi Shetty's engineering organization for enterprise AI platforms, in Technology under CIO Rob Alexander, in Enterprise Data, and in data science teams inside the businesses. Evident ranked Capital One first among 50 banks on AI talent in Oct 2026, and Capital One researchers and academic partners had 16 papers accepted at EMNLP 2026.
Reporting lines
Founder, Chairman and CEO
Richard Fairbank
Enterprise Services
Frank LaPrade, Chief Enterprise Services Officer
Technology, Enterprise Product Management, Design, Enterprise AI, Enterprise Data, Capital One Software, Corporate Security and other functions (2026 proxy)
Technology
Rob Alexander, CIO
Listed within Enterprise Services in the 2026 proxy; Alexander's direct reporting line is not published
Cybersecurity
Chris Nims, CISO (unconfirmed)
The 2025 Form 10-K puts the CISO and CIO in the first line of defense for technology risk; reporting line inferred
Enterprise AI
Prem Natarajan, Chief Enterprise AI Officer
Listed within Enterprise Services in the 2026 proxy; Natarajan's direct reporting line is not published
AI Foundations
Milind Naphade, SVP
Builds foundational AI capabilities and houses applied AI research; line to Enterprise AI inferred, since Naphade's title is SVP, Technology
Enterprise AI platform engineering
Rashmi Shetty, VP, Enterprise AI
Builds the enterprise platforms for generative and agentic AI applications (CIO Dive, Sep 2026); reporting line inferred from her title
Enterprise Data
Amy Lenander, Chief Data Officer (role not confirmed for 2026)
Listed within Enterprise Services in the 2026 proxy; builds the data platforms and products behind Capital One's AI
Brex
Pedro Franceschi, Founder and CEO
Franceschi reports to Frank LaPrade (Brex, Jan 2026); results reported in Domestic Card since the second quarter of 2026
Financial Services (Auto Finance)
Sanjiv Yajnik, President
Owns Chat Concierge and ProtectID; reporting line assumed from executive officer status
Prem Natarajan is not among the executive officers listed in Capital One's 2025 annual report, and Capital One has not said to whom he reports. Natarajan's title changed from Chief Scientist and Head of Enterprise AI (2023 to Mar 2025) to Chief Scientist and Chief Enterprise AI Officer (2026). Capital One's 2025 Form 10-K describes an unnamed CISO with nearly five years as CISO at a major global technology company before joining, which is consistent with Chris Nims's RSA Conference profile (CISO of Verizon Media and, before that, AOL); Capital One itself has not named Nims as CISO in a filing.
Where the AI workforce sits
Central Enterprise AI and AI Foundations, including applied research
Capital One has not published the size of these teams.
Not disclosedTechnology organization engineers building AI into products
Capital One hired thousands of engineers after 2013 but gives no current count.
Not disclosedData science teams in the lines of business
A retail bank data science leader co-presented the fraud servicing agents at NVIDIA GTC (Mar 2026).
Not disclosed
Named people
Use cases
Capital One applies AI in software development, analysis, marketing, servicing, credit, fraud and cyber, Richard Fairbank wrote in Mar 2026, and has described named systems such as Chat Concierge, ProtectID, VulnHunter, a knowledge tool for customer service and a transformer-based recommender. Capital One has published no business return credited to AI for any of them: its one AI dollar figure, nearly $150 million of auto lending fraud stopped by ProtectID in 2025 (Oct 2026), is a gross amount stopped rather than losses Capital One says it avoided. Capital One reports no firm-level AI return, and Fairbank said on Oct 20, 2026 that it does not intend to quantify AI's benefits.
| Use case | Area | Status | ROI |
|---|---|---|---|
| ProtectID fraud screening ProtectID runs Capital One's proprietary fraud models on every application submitted through Dealer Navigator, alerts dealers in real time and sends flagged applicants an identity check before funding. Capital One, which calls ProtectID AI-powered, said it stopped nearly $150 million in auto lending fraud in 2025 (Oct 2026), after giving more than $120 million of potentially fraudulent applications for the 12 months to Mar 2026. | Auto Finance | In production | |
| Chat Concierge for car dealers Chat Concierge is a multi-agent assistant on participating dealers' websites, the Navigator platform and dealer CRMs that compares vehicles, answers car buyers' questions and schedules test drives, checking each action plan against policies and business rules before it acts. Capital One distilled fine-tuned open-weight models into smaller ones to cut Chat Concierge's latency fivefold (Oct 2026). | Auto Finance | In production | |
| Multi-agent support for fraud servicing calls Capital One applied its multi-agent framework and customized foundation models to customer service calls about fraud, cutting summarization latency and supporting quality reviews, work it presented at NVIDIA GTC in Mar 2026 as showing promise. | Fraud and servicing | Pilot | |
| Fraud and scam detection Capital One's fraud teams combine data-driven models built with data science teams, real-time monitoring and human review to pause high-risk payments, such as wire transfers tied to scams, before money moves; the 2025 Form 10-K lists identifying potentially fraudulent transactions among the processes that use models and AI. | Fraud | In production | |
| Knowledge tool for customer service agents Capital One's generative AI tool helps agents find answers to customer questions quickly, such as whether a declined transaction counts against a daily card limit, using a custom retrieval pipeline and models fine-tuned on Capital One data. | Customer service | In production | |
| Real-time personalization A transformer-based recommender engine, which replaced tree-based models, drives real-time personalized experiences across Capital One's digital channels. | Digital channels | In production | |
| Credit, pricing, marketing and anti-money laundering models Capital One's 2025 Form 10-K gives pricing products, grading loans, extending credit and predicting deposits and loan losses as examples of its use of models and AI, and Capital One named anti-money laundering, multichannel marketing and product valuations among areas where AI improves its operations (Mar 2025). | Card and Consumer Banking | In production | |
| AI-based monitoring Capital One uses AI-based monitoring in some areas that checks full files in real time instead of samples in batches and flags issues with a root-cause diagnosis and suggested fixes; Richard Fairbank called it early days in Oct 2026. | Risk management | In production | |
| VulnHunter code security VulnHunter traces attacker paths through source code, tries to disprove its own findings and proposes code fixes; Capital One ran it across thousands of its own repositories before open-sourcing it in Jul 2026. | Technology | In production | |
| AI-assisted software development Capital One gives its software developers AI coding tools, ran a month-long agentic coding program whose live session drew over 10,000 engineers (Aug 2026), and Richard Fairbank listed software development first among the areas where Capital One applies AI (Mar 2026). | Technology | In production | |
| DataAgents for cloud resource analysis Capital One engineers paired an AI agent with a cloud asset data product to classify about 350 cloud resource types across AWS, Azure and Google Cloud for dormancy, with humans checking every result before the detection queries went to production (Jun 2026). | Technology | In production | |
| Capital One Shopping Capital One Shopping uses real-time machine learning to save customers money when they shop online and to help merchants target new customers. | Shopping | In production | |
| Brex AI agents Brex, acquired in Apr 2026, runs in-house AI agents for expense management and audit, with procurement, payments and accounting agents in development, Richard Fairbank said in Jan 2026. | Business payments | In production |
Adoption
Capital One has not published how many employees use generative AI tools or how often. Capital One said in Aug 2026 that it had launched an enterprise-wide AI learning hub to more than 60,000 associates that year and run a month-long agentic coding program whose live session drew over 10,000 engineers, counts of access and attendance rather than of regular use. Capital One's AI spread in layers: statistical modeling since its founding, which Richard Fairbank said moved from regression models to neural network machine learning models over time (Apr 2026), a technology rebuild from 2013 that moved the company onto AWS by 2020, a central Enterprise AI organization and AI Foundations team from 2023, and proprietary agents from 2025, starting with Chat Concierge for car dealers and extending to fraud servicing calls and code security in 2026. Capital One encourages associates anywhere in the bank to build on its centralized technology stack (Apr 2026), and Richard Fairbank wrote in Mar 2026 that Capital One had broken into the top 10 U.S. companies for generative and agentic AI patent filings.
By the numbers
Rollout, first pilot to today
2013
Capital One declared a comprehensive technology transformation, from the bottom of the technology stack up, to prepare for real-time AI.
2020
Capital One closed its last data center and moved fully into the cloud, according to Richard Fairbank's 2025 letter; Capital One's technology site gives 2021.
May 2023
Prem Natarajan joined Capital One from Amazon's Alexa AI organization as Chief Scientist and Head of Enterprise AI.
Jun 2024
Milind Naphade said a year after joining that his new AI Foundations organization had launched an applied research job family and a PhD research internship, and that Evident had named Capital One the top bank for AI talent.
Mar 2025
Capital One presented Chat Concierge, built on Meta's Llama for car buyers and dealers, as its new and first proprietary multi-agent AI assistant.
May 2025
Capital One completed its acquisition of Discover on May 18, 2025, adding the Discover, PULSE and Diners Club networks.
Jul 2025
Capital One Auto launched ProtectID, a fraud screening tool built on proprietary fraud models, inside Dealer Navigator.
Mar 2026
Capital One presented at NVIDIA GTC how its multi-agent framework, first built for car buying, was reused for customer service calls about fraud.
Apr 2026
Capital One closed its acquisition of Brex, an AI-native corporate card and spend management platform.
Jul 2026
Capital One open-sourced VulnHunter after running the agentic code-security tool across thousands of its own repositories.
Aug 2026
Capital One said it had launched an enterprise-wide AI learning hub to more than 60,000 associates in 2026 and run a month-long agentic coding program for engineers.
Sep 2026
All new Discover card originations ran on Capital One's technology stack, with the back book due to move by the first quarter of 2027.
Economics & controls
Spend and ROI
Technology and AI spending
Capital One does not disclose a technology budget or a figure for AI spending. Richard Fairbank wrote in Mar 2026 that Capital One's opportunities require significant and sustained investment in marketing, infrastructure, technology, talent and AI, said on Apr 21, 2026 that Capital One continues to invest in building AI infrastructure and specific AI experiences, and said on Oct 20, 2026 that it continues to invest in foundational technology and AI. Capital One's domestic card non-interest expense rose 9% year over year in the third quarter of 2026, driven by technology investments, higher marketing and the addition of Brex.
Reported ROI
Capital One reports no firm-level return on AI and no business return credited to AI for any use case. Asked on Oct 20, 2026 whether Capital One would quantify revenue or market share gains from AI, Richard Fairbank said it does not intend to, and compared AI to Capital One's information-based strategy, which shows up across the business model instead of in one app. Capital One's one AI dollar figure is for ProtectID, which it calls AI-powered: nearly $150 million of auto lending fraud stopped in 2025 (Oct 2026), after more than $120 million of potentially fraudulent applications in the 12 months to Mar 2026. Both are gross amounts stopped, not losses Capital One says it avoided, and the two do not reconcile for a tool launched in Jul 2025. Chat Concierge's figures, leads converted 55% better than a leading competitor (Q1 2025) and dealer customer engagement up 50% (Oct 2026), measure results for dealers.
Efficiency
Richard Fairbank said in Jan 2026 that Capital One's investments would put some upward pressure on its efficiency ratio in the near term, and said on Oct 20, 2026 that Capital One continues to seek meaningful efficiencies outside its investment areas, powered by its technology transformation, while giving no efficiency ratio guidance. Capital One has not attributed any efficiency figure to AI.
Headcount
Capital One had about 76,300 employees at Dec 31, 2025. Capital One's 2025 Form 10-K says advances in automation and AI may lead to workforce evolution that requires more training and competition for candidates with advanced technology skills. Asked about AI and jobs on Apr 21, 2026, Richard Fairbank said he is on the optimistic side, that demand for some jobs such as software development could rise, and that Capital One is not changing credit policy in anticipation of AI-driven unemployment.
The Brex acquisition
Capital One announced the Brex deal at $5.15 billion on Jan 22, 2026; total consideration at closing on Apr 7, 2026 was about $4.5 billion, including $2.6 billion in cash, and Capital One booked $510 million of Brex's developed technology as an intangible asset. Brex integration expenses were $96 million in the second quarter and $47 million in the third quarter of 2026.
Governance and risk
Board oversight
Capital One's Board oversees material and emerging risks, naming generative AI as an example, and the Board and its Risk Committee take part in periodic education sessions on generative AI (2026 proxy statement).
Model and AI risk
Capital One's 2025 Form 10-K gives pricing, fraud detection, loan grading, credit, deposit and loss forecasting, capital assessment, software development and personalization as examples of where it uses models and AI, says that generative AI can produce false or hallucinatory output, and that many of its data management, modeling and AI processes are manual. The filing warns that staff use of third-party AI tools could expose confidential information to outside training sets. Capital One's 2026 proxy statement names Michael Zamsky, Chief Credit and Financial Risk Officer, as the leader of Enterprise Model Risk Management.
Guardrails inside the agents
Chat Concierge's agents plan an action, check the plan for hallucinations and errors, simulate it against Capital One's policies and business rules, and explain it to the customer before acting (Mar 2025). Prem Natarajan said in Dec 2025 that Capital One wanted to start agentic AI at the low end of the risk spectrum. Rashmi Shetty, VP of enterprise AI, described in Sep 2026 validating agents with rule-based checks, sandbox simulations and evaluations, tracking agent trajectories and tool accuracy, and deciding where humans must approve high-risk actions.
Lines of defense for technology risk
Capital One's CISO and CIO form the first line of defense for cybersecurity and technology risk, a Chief Technology Risk Officer provides independent oversight as the second line, and Internal Audit is the third, according to the 2025 Form 10-K.
Agentic commerce and fraud
Capital One's 2025 Form 10-K warns that agentic commerce, in which AI agents initiate and execute transactions for users, may raise fraud losses and change when merchants or issuers are liable, under legal frameworks Capital One calls nascent. The filing also flags synthetic video, images and identity documents as new sources of identity fraud.
Open security tools
Capital One released VulnHunter under the Apache 2.0 license in Jul 2026 so others can inspect and improve it, and NVIDIA listed VulnHunter among the tools Open Secure AI Alliance members contribute on Aug 4, 2026, as an alliance working group drafted shared guidelines for reporting AI security incidents.
Responsible AI research and regulation
Capital One funds academic centers on responsible and generative AI at the University of Southern California, the University of Illinois and Columbia University. Capital One's 2025 Form 10-K notes state AI laws in Colorado and California and says it may need to adjust offerings where legal frameworks differ.
Recent signals
Oct 22, 2026
Prem Natarajan told an Evident AI event in New York that Capital One had distilled its fine-tuned open-weight models into smaller models for Chat Concierge, cutting latency fivefold and raising dealer customer engagement by 50%, CIO Dive reported.
Oct 22, 2026
Capital One said its researchers and academic partners would present 16 accepted papers at EMNLP 2026 in Budapest, on red-teaming, retrieval-augmented generation and LLM evaluation.
Oct 20, 2026
Richard Fairbank told analysts that Capital One does not intend to quantify AI's benefits, said AI was the objective of its 2013 technology rebuild, and described AI-based monitoring that checks full files in real time.
Oct 20, 2026
Capital One said ProtectID, which it calls AI-powered, stopped nearly $150 million in auto lending fraud in 2025, as Capital One Auto won a 2026 Inc. Power Partner award.
Oct 6, 2026
Evident's 2026 AI Index ranked Capital One second among 50 banks, behind JPMorgan Chase, first on talent and innovation and 40th on leadership.
Sep 28, 2026
Rashmi Shetty, Capital One's VP of enterprise AI, told CIO Dive that Capital One runs agentic AI internally to improve customer service and externally in Chat Concierge, and builds controls into its platform before developers build agents.
Aug 4, 2026
NVIDIA listed Capital One's VulnHunter among the tools Open Secure AI Alliance members contribute, as an alliance working group drafted shared guidelines for AI security incidents.
Aug 3, 2026
Capital One said it had launched an enterprise-wide AI learning hub to more than 60,000 associates in 2026 and that a live session of its agentic coding program drew over 10,000 engineers.
Jul 16, 2026
Capital One open-sourced VulnHunter, an agentic code-security tool written for Anthropic's Claude Opus 4.8 and Claude Code, after running it across thousands of its own repositories.
Jun 9, 2026
Capital One engineers described DataAgents, which classified about 350 cloud resource types across AWS, Azure and Google Cloud in 10 days against an estimate of six to nine months of expert analysis.
Apr 21, 2026
Richard Fairbank said Capital One continues to invest in building AI infrastructure and specific AI experiences, and that it is not changing credit policy in anticipation of AI-driven job losses.
Apr 7, 2026
Capital One completed its acquisition of Brex, which it calls an AI-native spend platform, for about $2.56 billion in cash and 10.6 million Capital One shares.
Mar 25, 2026
Richard Fairbank's annual letter said Capital One builds its own AI instead of renting it, has recruited hundreds of AI professionals, and broke into the top 10 U.S. companies for generative and agentic AI patent filings.
Jan 22, 2026
Capital One agreed to buy Brex in a stock and cash deal valued at $5.15 billion, and Richard Fairbank said Capital One's technology stack was built working backwards from the AI revolution.
Dec 15, 2025
Prem Natarajan told Fortune that Capital One wanted to start agentic AI at the low end of the risk spectrum and had cut Chat Concierge's latency fivefold since launch.
Quick answers
- How does Capital One deploy AI?
- Capital One builds its own AI instead of renting it: a central enterprise AI group under Prem Natarajan builds proprietary multi-agent systems on customized open-weight models such as Meta's Llama, on a technology stack that runs principally on AWS after the bank closed its last data center in 2020. Capital One uses AI in fraud, credit, marketing, servicing and software security, but Richard Fairbank said on Oct 20, 2026 that the company does not intend to quantify AI's benefits.
- What does Capital One use AI for?
- Capital One applies AI in software development, analysis, marketing, servicing, credit, fraud and cyber, Richard Fairbank wrote in Mar 2026, and has described named systems such as Chat Concierge, ProtectID, VulnHunter, a knowledge tool for customer service and a transformer-based recommender. Capital One has published no business return credited to AI for any of them: its one AI dollar figure, nearly $150 million of auto lending fraud stopped by ProtectID in 2025 (Oct 2026), is a gross amount stopped rather than losses Capital One says it avoided. Capital One reports no firm-level AI return, and Fairbank said on Oct 20, 2026 that it does not intend to quantify AI's benefits.
- Who leads AI at Capital One?
- AI at Capital One is led by Prem Natarajan (EVP, Chief Scientist and Chief Enterprise AI Officer); Amy Lenander (Chief Data Officer); Milind Naphade (SVP, Technology, and Head of AI Foundations). Ownership of AI on the org chart: Prem Natarajan, Chief Enterprise AI Officer, within Frank LaPrade's Enterprise Services.
- What AI platform does Capital One use?
- Capital One built a proprietary multi-agent framework for the Chat Concierge car-buying assistant and reused it for customer service calls about fraud (Mar 2026); a central engineering team builds the enterprise platforms its agents run on (Sep 2026).
- Which AI models does Capital One use?
- Capital One customizes open-weight models in-house, building Chat Concierge on Meta's Llama and distilling fine-tuned models into smaller ones (Oct 2026), and wrote its VulnHunter code-security agent for Anthropic's Claude Opus 4.8 in Claude Code (Jul 2026).
- How many Capital One employees use AI?
- Capital One launched an AI learning hub to more than 60,000 associates in 2026 and drew over 10,000 engineers to a live agentic coding session (Aug 2026), but has not published how many employees use AI tools.
- How far has AI spread at Capital One?
- Capital One sits at Scaling: a named executive, Prem Natarajan, has owned enterprise AI since 2023; AI runs in production in fraud, credit, marketing, servicing, auto lending and code security; a proprietary multi-agent framework built for car buying is being reused for fraud servicing calls; and Capital One launched an AI learning hub to more than 60,000 associates in 2026. Capital One meets the Scaling test only in part, because it has published no business return credited to AI, and falls short of Embedded because it has published no count of employees using AI and Richard Fairbank said on Oct 20, 2026 that it does not intend to quantify AI's benefits.
Open questions
- How many of the more than 60,000 associates offered Capital One's AI learning hub use generative AI tools each week, and does Capital One offer a general-purpose AI assistant to all associates?
- How much does Capital One spend on technology and on AI each year, given that it publishes neither figure?
- How many AI and machine learning specialists does Capital One employ beyond the hundreds of AI hires Richard Fairbank cited in Mar 2026, and how large are Enterprise AI and AI Foundations?
- Which foundation models does Capital One run besides Meta's Llama and Anthropic's Claude, how many are open-weight, and where does inference run: on AWS, on NVIDIA hardware elsewhere, or both?
- What business return does Capital One get from deployed AI with no published figure, such as Chat Concierge's effect on its own auto loan originations, the knowledge tool for customer service, the recommender engine and VulnHunter?
- How much of the auto lending fraud ProtectID stopped would have become Capital One losses rather than dealers' losses, and how do nearly $150 million for 2025 (Oct 2026) and more than $120 million for the 12 months to Mar 2026 reconcile for a tool launched in Jul 2025?
- When did Chat Concierge enter production, given that CIO Dive said in Sep 2026 it had been in production for over two years while Capital One presented it as new in Mar 2025?
- To whom do Prem Natarajan, Milind Naphade and Rashmi Shetty report, and does Enterprise AI sit inside Technology or beside it in Enterprise Services?
- Will Discover's payment network systems move from data centers to AWS, and will Capital One's AI models run on them?
- Why did Evident rank Capital One 40th on its leadership pillar in 2026, down 20 places, while ranking it first on talent and innovation?
- How will Capital One combine Brex's AI agents with its own multi-agent framework and models?
Sources
Capital One
- AI at Capital One
- Driving the future of car buying with agentic AI, Chat Concierge (Mar 5, 2025)
- Announcing VulnHunter (Jul 16, 2026)
- Advancing AI fluency at Capital One (Aug 3, 2026)
- Capital One at NVIDIA GTC 2026 (Apr 8, 2026)
- DataAgents: How we turned 9 months of analysis into 10 days (Jun 9, 2026)
- AI in financial services, AWS re:Invent 2024 talk (Mar 27, 2025)
- Meet Milind Naphade, SVP of AI Foundations (Jun 5, 2024)
- Getting to know Amy Lenander, Chief Data Officer (May 3, 2024)
- Capital One Tech: Cloud
- Capital One Science: Introducing Prem Natarajan (May 24, 2023)
- Capital One Science: Academic partnerships
- Capital One Science: Capital One at EMNLP 2026 (Oct 22, 2026)
- Capital One Auto unveils ProtectID (Jul 23, 2025)
- Capital One Auto earns two 2026 BIG Innovation Awards (Mar 12, 2026)
- Capital One Auto wins 2026 Inc. Power Partner Award (Oct 20, 2026)
- Capital One completes acquisition of Brex (Apr 7, 2026)
- The last line of defense: stopping fraud in the moment (Jul 15, 2026)
Filings, letters and investor materials
- Capital One 2025 Form 10-K (filed Feb 19, 2026)
- 2025 Annual Report with Richard Fairbank's letter (filed Mar 25, 2026)
- 2026 proxy statement (Mar 25, 2026)
- Form 10-Q for the second quarter of 2026 (Jul 28, 2026)
- Form 8-K exhibit: Capital One to acquire Brex (Jan 22, 2026)
- Form 8-K: completion of the Brex acquisition (Apr 7, 2026)
- Third quarter 2026 earnings release (Oct 20, 2026)
- Q3 2026 earnings call transcript, The Motley Fool (Oct 20, 2026)
- Q1 2026 earnings call transcript, The Motley Fool (Apr 21, 2026)
- Q4 2025 earnings call transcript, The Motley Fool (Jan 22, 2026)
- Q3 2025 earnings call transcript, The Motley Fool (Oct 21, 2025)
Partners and benchmarks
- Evident AI Index for Banks, 2026 results (Oct 2026)
- NVIDIA: AI leaders propose SAFE guidelines, Open Secure AI Alliance (Aug 4, 2026)
- Brex: Brex and Capital One join forces, by Pedro Franceschi (Jan 2026)
- DataIQ 100: Amy Lenander, Chief Data Officer, Capital One (2025)
- RSA Conference speaker profile: Christopher Nims
News
- CIO Dive: Smaller AI models help Bank of America, Capital One manage costs (Oct 28, 2026)
- CIO Dive: Capital One's agentic AI strategy hinges on data, platform-first mindset (Sep 28, 2026)
- VentureBeat: Capital One releases VulnHunter (Jul 17, 2026)
- Fortune: AI agents are delivering results at some companies, Capital One's Chat Concierge (Dec 15, 2025)
Data confidence
- Confirmed
- Stated in a primary or reputable public source.
- Estimated
- Derived or approximated from public reporting.
- Unconfirmed
- Reported, but not yet corroborated.
Capital One profile · Enterprise AI diffusion report by Imagine · facts checked against their sources on Nov 1, 2026
