What proprietary AI solutions has Capital One built?
Capital One builds its AI on open-weight models it customizes itself rather than on a licensed closed frontier model. Its generative models are open-source LLMs customized with synthetic and public data the bank began collecting in 2024: more than 20 trillion tokens and over 495,000 task examples. The models run inside MACAW, a multi-agent framework first built for Chat Concierge, the Llama-based car-buying assistant launched for dealers in March 2025, and then reused for a fraud-servicing workflow of millions of calls a year, where it writes post-call summaries for several hundred specialist agents. In September 2026 Prem Natarajan said customer-service agents now edit only a fifth of AI-drafted case summaries. Older in-house systems include Eno, whose natural-language processing the bank built itself (2017), and a generative agent servicing tool for thousands of customer-service agents (presented in March 2025). Tools first built for its own data estate are sold through Capital One Software: Slingshot, which manages tens of petabytes of the bank's own Snowflake data and announced AI query optimization in June 2026, and Databolt, built because 'we couldn't find a viable solution on the market'. The work sits in Natarajan's enterprise AI organization and Milind Naphade's AI Foundations team. Capital One trains engineers on Google and Anthropic frontier models but has not named a closed-model provider behind its production assistants.[1][2][3][4][5][6][7][8][9][10][11][12]
Data as of Oct 22, 2026 · Capital One profile updated Oct 19, 2026 · 34 sources cited on this page · 140 banks profiled on this site · Markdown
Which AI models, platforms and assistants has Capital One built in-house?
Systems Capital One describes as its own, with the base technology it names. 'Since' is the earliest public date; MACAW and the infrastructure optimizer were described at VB Transform 2026 in a session Capital One sponsored.
| Platform or model | What it does | Built on | Since | Status |
|---|---|---|---|---|
| Customized generative models | Generative AI models trained on more than 20 trillion tokens: multi-turn conversations, instruction tuning for math reasoning, code, web content and over 7,100 synthetic conversations (more than 495,000 task examples); curated to handle complex business rules by contrasting preferred and incorrect responses | Open-source LLMs customized with synthetic and public data owned by Capital One or licensed as open source | Data collection from 2024, ongoing | In use (disclosure retrieved September 2026)[1] |
| MACAW (multi-agentic workflow) | An understanding agent reads intent, a reasoning agent drafts a summary, a validation agent fact-checks it and an explaining agent formats it; used on a fraud-servicing workflow of millions of calls a year lasting four to sixty minutes, for several hundred specialist agents | Fine-tuned open-weight models and Capital One's own orchestration harness | Presented at NVIDIA GTC, March 2026 | In production[2][3] |
| Chat Concierge | Car-buying assistant on dealer websites: one agent converses, one plans from business rules and permitted tools, one evaluates the plan against bank policy, one explains and validates it with the buyer; Naphade cites engagement up to 55% higher in some cases | Meta's Llama customized with Capital One data; in-house technology, open-source tool chains and NVIDIA's inference stack (Triton, TensorRT-LLM) | March 2025 | In production[4][2][13] |
| Agent servicing tool | Generative AI and semantic search give thousands of customer-service agents knowledge-grounded answers, for example whether a declined transaction counts against a daily card limit | In-house build using the NVIDIA enterprise AI ecosystem | Presented March 2025 | In use[7][14] |
| Eno | Customer assistant in the app, on the web, by text and email | Natural-language processing built in-house in under three months, trained on customer-service chat logs | March 2017 | In use; no generative version announced[15][6][16] |
| Agentic infrastructure optimizer | Runs a researcher-designed search space of LLM hosting configurations, sets up and runs each experiment and summarizes results, to find the combination with the lowest latency | In-house agentic system | Described 2026 | Internal[2] |
| GPU HPC platform | Multi-tenant high-performance computing cluster with components built to recover from GPU XID errors; GPU-accelerated Spark on Kubernetes for data engineering | Kubernetes on AWS; NVIDIA RAPIDS | Presented March 2025 | Internal[7] |
| Slingshot | Manages, governs and optimizes Capital One's own Snowflake environments across tens of petabytes; announced in June 2026 AI query and table optimization and AI duplicate-pipeline detection | Built for Snowflake; Databricks support added 2025 | June 2022 (sold externally) | Commercial product[17][8][9][18] |
| Databolt | Replaces sensitive data with tokens that keep the original format; internally no sensitive data goes into a large language model unless it has been tokenized | In-house; integrated with Snowflake and Databricks | Commercial rollout completed April 2025 | Commercial product[10] |
Which Capital One teams build its proprietary AI?
Teams and leaders as Capital One and the trade press name them. Headcount for the AI organization is not published.
| Team | What it builds | Named leaders | Source date |
|---|---|---|---|
| Enterprise AI | Firmwide 'proprietary, in-house AI development', including multi-agent systems with validation frameworks | Prem Natarajan, EVP, Chief Scientist and Head of Enterprise AI | April 2025[7] |
| AI Foundations and Applied Research | Foundation models across language, graphs and time series; LLM training, customization and alignment; multi-agent reasoning; the Chat Concierge agents | Milind Naphade, SVP, Head of AI Foundations | November 2025[11][13] |
| Enterprise AI platforms engineering | Puts generative and agentic applications into the retail bank, risk, legal and compliance with codified policy checks, runtime controls and evals | Rashmi Shetty, VP of Enterprise AI | September 2026[19] |
| Machine learning engineering | Model customization, the MACAW harness and the latency optimizer | Kel Vanee, MVP of machine learning engineering | 2026[2] |
| AI/ML Platforms | GPU HPC cluster and GPU-accelerated data pipelines | Ali Rodell, Senior Director; Chandra Shekar Mudarpu and Herik Webb, Directors | April 2025[7] |
| Data pipelines for foundation models | Distributed pipelines that curate datasets for 'laser-focused foundation models for financial applications' | Brian Nguyen, Sr. Engineering Manager; Nick Resnick, Lead AI Engineer | April 2026[3] |
| Capital One Software | Slingshot and Databolt for outside customers | Jeff Chou, VP Slingshot Product Management; Leon Bian, VP and head of Databolt product | 2025-2026[9][10] |
| Engineering base | More than 10,000 engineers in a month-long agentic coding program; an AI learning hub for over 60,000 associates; hands-on frontier-model training with Google and Anthropic | Capital One | August 2026[12] |
What does Capital One say about building versus buying AI?
'At Capital One, we're not just using AI, we're building AI,' Kel Vanee said at VB Transform 2026, in a session Capital One sponsored. He described a centralized enterprise AI platform with built-in governance, open-weight models fine-tuned on proprietary data and an orchestration harness the bank built itself, and said customizing a model for one use case lifts results across the portfolio because it learns 'Capital One use cases, policy, and nomenclature'.[2]
Naphade said in July 2025 that the team used open-weight rather than closed models because they allow significant customization, and that competitive advantage rests on proprietary data. In September 2026 Natarajan said the bank had moved toward customized open-weight models and smaller, specialized models with more targeted context to cut latency and token costs.[13][5]
The rule extends past models. Leon Bian said of Databolt: 'We couldn't find a viable solution on the market, so we decided to build our own.' Slingshot began as the tool Capital One built for its own Snowflake estate before it became a commercial platform.[10][8]
Reuse is the operating model: the multi-agent framework 'was originally built to support a car-buying solution and later applied to supporting customers with fraud resolution', and associates are encouraged to pull from the centralized stack. Outside models enter mainly through training, where Capital One partners with Google and Anthropic to give technologists hands-on work with frontier models.[3][12]
Which Capital One patents relate to its in-house AI systems?
Databolt's method, tokens that keep the original data format, matches two filings by Capital One Services, LLC: US12579308B1 'Tokenization with format preservation', granted March 17, 2026 (filed June 9, 2025), and US20250258958A1 'Vaultless tokenization', published August 14, 2025 from a September 1, 2022 priority. Neither filing names Databolt.[20][21][10]
For the agent systems, 2026 publications include applications on knowledge gap detection for agentic conversation (EP4793794A1) and on safeguarding user-facing AI models (EP4776166A1). They do not name MACAW, Chat Concierge or Eno, so the link is by subject only; the wider portfolio is on the research and patents page.[22][23]
What has Capital One not disclosed about its proprietary AI?
Base models beyond Meta's Llama for Chat Concierge, model sizes, GPU counts and training compute are not published; the training-data disclosure gives the token count and data types but not which open-source LLMs were customized.[1][2]
No usage figures are published for Eno or the agent servicing tool, and Capital One has not disclosed revenue for Capital One Software's Slingshot and Databolt; CIO Rob Alexander said in 2022 that winning enterprise software businesses take five to seven years to reach $100 million in revenue.[16][7][24]
What does the Capital One profile record on this topic?
The rows below are the entries of the Capital One AI strategy profile that bear on this page, as of Oct 19, 2026; the profile holds the full record of 46 dated moves, 13 use cases and 75 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Chat Concierge | Multi-agent assistant on dealer websites; compares vehicles, schedules appointments, routes to sales. | In production | [4] |
| Eno | Customer assistant in the mobile app, by text and email, in notifications (including on smart watches) and online from a computer: answers account questions and watches for charges that might surprise customers, free-trial endings and recurring charges. Natural-language processing built in-house. | In production | [16][6] |
| Databolt (Capital One Software) | Tokenisation and data-security tooling sold to enterprises for secure AI adoption. | In production | [25] |
| MACAW fraud-call summaries | Understanding, reasoning, validation and explanation agents on fine-tuned open-weight models summarise long fraud calls for several hundred specialist agents; the same architecture runs Chat Concierge. | In production | [2] |
| Proprietary LLMs customised from open-source models | Generative AI models built by customising open-source LLMs with synthetic and licensed open-source data; the disclosed training set exceeds 20 trillion tokens. | In production | [1][2] |
| ProtectID dealer fraud screen | Proprietary fraud models and AI score every application sent through Dealer Navigator and flag synthetic and third-party fraud. | In production | [26] |
| VulnHunter | Agentic code-security tool with attacker-first analysis and a falsification engine; open-sourced in July 2026. | In production | [27] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 29, 2026 | Natarajan: cutting latency and token costs with smaller, open-weight models | Chief enterprise AI officer Prem Natarajan says the bank has put more emphasis on agentic AI infrastructure to reduce latency and lower token costs, building customised open-weight and specialised models with less context targeted at the desired output; customer-service agents now edit only one-fifth of AI-drafted copy. | [5] |
| Aug 13, 2026 | MACAW and open-weight models, described | A machine-learning engineering leader explains the multi-agent fraud-call workflow and fine-tuning open-weight models on proprietary data. | [2] |
| Jul 16, 2026 | VulnHunter open-sourced | An internally built agentic AI code-security tool released on GitHub under Apache 2.0; the CISO cites the brief window before AI attack tools spread. | [27] |
| Jun 16, 2026 | Databolt Connect on Databricks Marketplace | Secure multi-party data sharing in Databricks Clean Rooms. | [28] |
| Mar 23, 2026 | Databolt expands for secure AI adoption | Capital One Software's data-security product adds capabilities for sensitive unstructured data. | [25] |
| Mar 12, 2026 | Chat Concierge and ProtectID win BIG Innovation Awards | Chat Concierge 55% more successful at converting leads than a leading competitor on Q1 2025 results; ProtectID prevented more than $120 million of potentially fraudulent applications in 12 months. | [29] |
| Jan 22, 2026 | Q4 2025 call: a stack built 'working backwards from the AI revolution' | Fairbank says the bank is 'now building AI solutions across our businesses' and that Brex has deployed in-house AI agents for expense management and audit. | [30] |
| Jul 23, 2025 | ProtectID launches for dealers | Real-time fraud screening inside Dealer Navigator using proprietary fraud models and a soft credit pull. | [31][32] |
| Jun 10, 2025 | Databricks partnership deepens; Sync Computing team joins | Databolt integrates with Unity Catalog; 4,000-plus Databricks users and 5 million-plus jobs at the bank. | [18] |
| Jun 4, 2025 | ASKS Center for generative-AI safety at Illinois | Capital One and the University of Illinois launch the Center for Generative AI Safety, Knowledge Systems and Cybersecurity; AI Foundations SVP Milind Naphade among its contributors. | [33] |
| Mar 4, 2025 | Chat Concierge launches for auto dealers | A multi-agent assistant that compares vehicles, books appointments and hands off to sales, on the Auto Navigator platform. | [4][34] |
| Jun 1, 2022 | Capital One Software launches with Slingshot | Enterprise B2B software business; first product manages Snowflake cost and governance after 27% in-house savings. | [17][24] |
Does Capital One build its own AI models?
Yes, by customizing open models. Its generative AI models are open-source LLMs that Capital One customizes with synthetic and public data, more than 20 trillion tokens collected from 2024, and its AI Foundations team researches foundation models for language, graphs and time series. Engineers presented at NVIDIA GTC 2026 on pipelines that curate data for 'laser-focused foundation models for financial applications'. The bank has not said it pretrains a large language model from scratch.
What is Capital One's MACAW?
MACAW is Capital One's multi-agentic workflow: an understanding agent, a reasoning agent, a validation agent and an explaining agent, running on fine-tuned open-weight models with governance and guardrails built in. It writes post-call summaries in a fraud-servicing workflow that handles millions of calls a year, for several hundred specialist agents, and its division of labor also underpins Chat Concierge for car buyers.
What AI does Capital One Software sell?
Two products first built for the bank's own use. Slingshot, sold since June 2022, manages and optimizes Snowflake and Databricks environments and announced AI query optimization, table optimization and duplicate-pipeline detection in June 2026. Databolt, rolled out commercially in April 2025, replaces sensitive data with format-preserving tokens so it can be shared or fed to AI; Capital One does not feed untokenized sensitive data into its own LLMs.
Does Capital One use OpenAI, Anthropic or Google models?
For training staff, yes: Capital One partners with Google and Anthropic to give technologists hands-on training with frontier models, tools and workflows. For its production assistants the disclosed models are open-weight ones it customizes, with Meta's Llama behind Chat Concierge. It has not announced an OpenAI, Anthropic or Google model behind Eno, Chat Concierge or MACAW.
Who builds Capital One's proprietary AI?
Prem Natarajan, EVP, Chief Scientist and Head of Enterprise AI, leads the enterprise AI organization; Milind Naphade, SVP, runs AI Foundations, including the Applied Research team; Rashmi Shetty, VP of Enterprise AI, leads platform engineering for generative and agentic applications; Kel Vanee leads machine learning engineering on MACAW. Capital One Software's product teams build Slingshot and Databolt. The bank has not published the AI organization's headcount.
Which patents cover Capital One's in-house AI?
The clearest match is Databolt: Capital One Services holds US12579308B1 'Tokenization with format preservation' (granted March 17, 2026) and published US20250258958A1 'Vaultless tokenization' (August 14, 2025), though neither names the product. Its 2026 applications on agentic conversation and safeguarding user-facing AI models cover the same ground as MACAW and Chat Concierge but do not name them.
- GAI Training Data Transparency Disclosure
- Why Capital One built its multi-agent AI platform around open-weight models (presented by Capital One)
- Capital One at NVIDIA GTC 2026
- AI for Car Dealerships: Smarter Car Shopping
- How bottom-line-focused financial firms are finding ROI in AI
- How and Why Capital One Built Eno's NLP In-House
- Highlights from NVIDIA GTC AI Conference 2025 (Capital One sessions and speakers)
- Smarter Snowflake optimization with Slingshot's contextual AI
- Capital One Software announces new observability and AI-powered optimization capabilities for Slingshot
- Why Capital One's software division created its own data security platform (Databolt)
- Capital One's Applied AI research
- Advancing AI fluency at Capital One
- How Capital One built production multi-agent AI workflows to power enterprise use cases
- How AI is transforming financial services & banking
- Capital One launches Eno, gender-neutral virtual assistant
- Eno, your Capital One assistant
- Capital One Launches Capital One Software (Slingshot for Snowflake)
- Capital One Software strengthens partnership with Databricks
- Capital One's agentic AI strategy hinges on data, platform-first mindset
- US12579308B1: Tokenization with format preservation (granted)
- US20250258958A1: Vaultless tokenization (application)
- EP4793794A1: Knowledge gap detection for agentic conversation (application)
- EP4776166A1: Systems and methods for safeguarding user-facing artificial intelligence models (application)
- Exclusive: Banking Giant Capital One Enters B2B Software Industry With Launch Of New Business
- Capital One Software Expands Databolt Capabilities
- Capital One Auto's ProtectID Wins 2026 'Fraud Prevention Innovation Award' for its Fraud Protection Tool
- Capital One releases VulnHunter, an open-source AI tool that finds software flaws before hackers do
- Capital One Software Introduces Databolt Connect for Secure Data Collaboration on Databricks Marketplace
- Capital One Auto earns two 2026 BIG Innovation Awards (Chat Concierge and ProtectID)
- Capital One (COF) Q4 2025 Earnings Call Transcript
- Capital One Auto Unveils Real-Time Fraud Detection Tool (ProtectID)
- Auto Buying Fraud is Exploding. Capital One Is Using AI to Fight Back
- Capital One and University of Illinois launch ASKS Center
- Capital One's 'Chat Concierge' Puts Agentic AI on Dealer Websites
When Capital One moves on AI, you'll read it here first.
when one of the banks we cover moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning