On this page
Ask "which companies actually use AI agents?" and you get two kinds of answer. Surveys say most large companies are "exploring" agents, without naming any. Vendor pages name customers, but only their own, and only the good numbers.
I wanted one list that names names across vendors, so I built it. This page logs every enterprise AI agent deployment I could verify from onward: the company, what its agent does, which team it works for, whether it was bought or built in-house, the scale and result the company or its vendor claims, and the source. As of it holds 83 deployments.
One rule runs through the whole page: every figure is a claim. When a row says an agent resolves 70% of chats, that is what the company or its vendor says, not something I measured. For survey-level numbers, see the AI agent adoption statistics; for a ranked view of use cases, see the top AI agent use cases of H1 2026. This page is the named, dated evidence underneath them.
The numbers at a glance
- 83 named enterprise AI agent deployments logged at 82 companies in 16 industries and 16 countries, from 6 February 2025 to 4 October 2026. Each row has a dated source; 3 are labelled pilots and the rest are described as in production.
- Customer support is the most common job for an agent: 29 of 83 deployments (35%), followed by software engineering with 17 (20%). The single most common use case is customer service resolution (23 deployments).
- 72% come with a quantified result claim (60 of 83): a share of conversations resolved, hours saved, money saved or time cut. None of these figures is independently audited; each is what the company or its vendor says.
- 64% run on an outside vendor's agent product (47 of the 73 deployments where the builder is public); 36% were built in-house (26), and 10 do not say. Salesforce and Sierra tie as the most common vendors, with 11 deployments each. 49% of rows rest mainly on a vendor's own customer story or press release.
- Financial services account for 28% of deployments (23 of 83 across banking, fintech and insurance). Banking, fintech and retail tie as the largest single industries with 10 each; companies headquartered in the United States account for 49 (59%).
Updated . A log of deployments that companies chose to make public, not a census of every agent in production.
Cite this page
Cite this page: "Enterprise AI agent deployments log", Gravity, updated , https://gravity.fast/data/companies-using-ai-agents/. The data is licensed CC BY 4.0: copy the rows, chart them, quote the numbers, and link back here. Every row as a spreadsheet: data.csv. When you quote a company's result, quote it as the company's claim, as this page does.
The deployments
Newest first. "What the agent does" is one plain sentence from the source. "Scale claimed" and "Result claimed" are copied as stated and attributed to whoever said them. Click a source to read the original.
| Date | Company | What the agent does | Function | Built with | Scale claimed | Result claimed | Sources |
|---|---|---|---|---|---|---|---|
pilot | Swiggy Food and restaurants, India | An agentic marketing system is given a goal, reactivating dormant users, and decides and executes campaigns, timing and messages on its own. | Other | In-house | n/a | 5 per cent higher reactivation of dormant users using 50 per cent fewer campaigns than business-as-usual (company claim, as reported by the press) | ETBrandEquity (The Economic Times), 2026-10-02 |
| Barclays Banking, United Kingdom | Claude models classify, enrich and route incoming Global Markets operations emails so requests reach the right processing path. | Operations | In-house (on Anthropic Claude) | approximately 120,000 emails each day | reduced manual handling activity (the vendor's claim) | Anthropic, 2026-10-01 | |
| JPMorgan Chase Banking, United States | In-house AI agents, built by the bank's engineering teams, run back-end operations and coding tasks such as writing code and building software. | Software engineering | In-house (built by JPMorgan software engineering teams) | most code now assisted or written with agentic coding capabilities; over 200 applications built on LLM Suite | coding portion of the software development lifecycle reduced by 10% to 20% (company claim, as reported by the press) | American Banker, 2026-10-01 | |
| Walmart Retail, United States | Sparky, Walmart's AI shopping agent, helps customers find products and build baskets in conversation instead of searching item by item. | Sales | In-house (uses external frontier models where appropriate) | n/a | average basket size up 38% when customers interact through Sparky versus browsing the website (company claim, as reported by the press) | CNBC-TV18, 2026-10-01 | |
| BNP Paribas Banking, France | In securities services, an agent triages incoming email and dispatches trade processing work from step to step. | Operations | Undisclosed | trade processing cut from ten steps to six | agent handling 80 to 85 per cent of the work within a couple of weeks (company claim, as reported by the press) | The Fintech Times, 2026-09-30 | |
| Booking Holdings Travel and hospitality, United States | Lola, an AI assistant, searches real-time availability and pricing and books hotels, restaurants, events and trips inside one conversation. | Sales | In-house (Booking Holdings) | n/a | No figure given | Lola / Booking Holdings (press release), 2026-09-29 | |
| Unilever Consumer goods, United Kingdom | Nova reads order emails, enters the data into SAP and validates orders in claims disputes; one of eight agents rolled out worldwide. | Operations | In-house (platform not named) | eight agents rolled out worldwide | No figure given | Consumer Goods Technology, 2026-09-28 | |
| Virgin Media O2 Telecom, United Kingdom | AI voice agent handles routine broadband fault calls, being introduced gradually, with complex issues left to human advisors. | Customer support | Sierra | initially routine broadband fault calls, a small proportion of total call volumes | No figure given | Sierra, 2026-09-24 | |
| INRY Professional services, United States | ServiceNow EmployeeWorks lets employees open tickets, submit PTO and expense requests by chat, completing tasks across connected systems. | IT | ServiceNow EmployeeWorks | n/a | median submission time cut 60% for self-service allocation requests; request completion rates of 97%; live in six weeks (the vendor's claim) | ServiceNow, 2026-09-23 | |
| University of Texas Medical Branch (UTMB) Healthcare and pharma, United States | AI agents on Bunkerhill's Carebricks platform triage specialist referrals, support lung nodule detection and handle pharmacy prior authorization. | Operations | Bunkerhill Health Carebricks | 22 agents in place | No figure given | Healthcare Innovation, 2026-09-23 | |
| Novo Nordisk Healthcare and pharma, Denmark | NovoScribe, built with Claude, automatically generates clinical study reports and other regulatory documents for human review and approval. | Other | In-house NovoScribe (on Anthropic Claude and Claude Code, with MongoDB Atlas) | clinical study report generation automated across the organization; expanded to device protocols and patient materials | clinical study reports in minutes rather than months; documentation time cut from 10+ weeks to 10 minutes; CSR writing times cut by 90% (company claim, as reported by the press) | pharmaphorum, 2026-09-17; Anthropic | |
| Live Nation Entertainment Media and publishing, United States | Venue Agent answers fan questions 24/7 on Live Nation venue websites, covering parking, what to bring, accessible seating and door times. | Customer support | Salesforce Agentforce | Live Nation venues across the U.S. | 95% of questions handled by Venue Agent answered without handoff to a team member (the vendor's claim) | Salesforce, 2026-09-16 | |
| Transportation Security Administration (TSA) Public sector, United States | Agentforce agent Ace answers routine traveller questions, such as packing liquids or medical devices at checkpoints, without human escalation. | Customer support | Salesforce Agentforce | approximately 100,000 routine traveler conversations per month | resolved 96% of routine inquiries without human escalation (the vendor's claim) | Salesforce, 2026-09-14 | |
| Microsoft Technology, United States | Falcon Agent runs automated network troubleshooting and remediation on incident tickets for Microsoft's internal network operations. | IT | In-house (Falcon Agent) | about 60% of total incident volume handled | around 300 hours saved in the first few months, 10-15 minutes saved per ticket (the company's claim) | Microsoft Inside Track, 2026-09-10 | |
| Lowe's Retail, United States | Mylow, an agentic shopping assistant, turns customer and associate questions into specific product lists; a companion tool helps store associates on the floor. | Sales | In-house (model vendor not stated in source) | more than 25 million questions from customers and store associates since launch | online shoppers who use Mylow convert at three times the rate of those who do not (company claim, as reported by the press) | PYMNTS (citing Lowe's Q2 earnings call), 2026-08-27 | |
| Octopus Energy Energy and utilities, United Kingdom | Arlo, an AI customer assistant, responds to a share of incoming customer emails so human advisors handle harder questions. | Customer support | Undisclosed (Octopus Energy's Arlo tool) | around 4% of incoming emails | frees human advisors to tackle the tough questions (qualitative) (company claim, as reported by the press) | Marketing Week, 2026-08-26 | |
| Rohlik Group Retail, Czech Republic | Devin agents plan, write, review code and handle incident response across the online grocer's software development lifecycle. | Software engineering | Cognition Devin | about 90% of code is agent-generated | 2x engineering throughput from November through August (the vendor's claim) | Cognition, 2026-08-24 | |
| EBANX Fintech, Brazil | Devin authors pull requests on the core payments platform, which engineers review and merge. | Software engineering | Cognition Devin | 92% of merged PRs in the core payments platform created with Devin | a workstream estimated at six weeks completed in 12 hours; additional output of 150 engineers' worth (the vendor's claim) | Cognition, 2026-08-20 | |
| UPS Logistics, United States | Agentic AI handles customs classification, documentation and error checking so most dutiable packages clear without manual review. | Operations | In-house | 112,000 packages a day requiring clearance (September 2025) | share cleared without manual intervention rose from about 21% in March 2025 to 90% by September 2025 (company claim, as reported by the press) | PYMNTS (citing UPS Q3 2025 earnings call), 2026-08-17 | |
| GE Aerospace Manufacturing and automotive, United States | Devin completes user stories for a previously capacity-constrained software team building airline systems. | Software engineering | Cognition Devin | n/a | almost 2x output in capacity for user stories over three months (the vendor's claim) | Cognition, 2026-08-16 | |
| Abu Dhabi Government (TAMM platform) Public sector, United Arab Emirates | The TAMM AI Assistant delivers government services and processes transactions and service payments for residents. | Customer support | Undisclosed | 115 services; 180,000 transactions plus 37,000 service payment transactions processed | 93.96% satisfaction rate; AI-powered interactions 97.11%, 2.89% transferred to human agents (company claim, as reported by the press) | Gulf News, 2026-08-15 | |
| Jaloc (ジャロック) Manufacturing and automotive, Japan | Sales agents evaluate deals, flag gaps and suggest next steps, answer staff questions from CRM knowledge and handle website inquiries for a logistics equipment maker. | Sales | Salesforce Agentforce | company of about 50 employees | order win rate up 15 points over two years, report writing time cut 80%, deal conversion rate up 5% (company claim, as reported by the press) | TechTarget Japan (reporting a Salesforce Japan announcement), 2026-08-04 | |
| Bajaj Finance Banking, India | AI voice agents handle loan conversations and facilitate disbursements, alongside AI text bots live across most of its products. | Sales | Undisclosed (Bajaj Finance FINAI programme) | 10 voice agents live; 24 of 32 AI text bots live | voice agents facilitated loan disbursements of Rs 2,100 crore in Q1 FY27 (company claim, as reported by the press) | Business Standard, 2026-07-30; Moneycontrol, 2025-12-08 | |
| WellSpan Health Healthcare and pharma, United States | Ana, a generative AI voice agent, handles inbound and outbound patient calls and answers FAQs, doctor and location requests online. | Customer support | Hippocratic AI | more than 160,000 patient calls per month; more than 7,000 hours of conversation | No figure given | WellSpan Health (GlobeNewswire), 2026-07-30 | |
| OCBC Banking, Singapore | Helios, an agentic AI platform, front-loads customer due diligence for private banking onboarding and passes screened prospects to relationship managers. | Operations | Undisclosed (OCBC Helios platform) | in use by relationship managers at Bank of Singapore | No figure given | Computer Weekly, 2026-07-29 | |
| DBS Bank Banking, Singapore | DBS Joy, an agentic assistant, retrieves and analyses account and transaction data and fulfils requests for corporate and SME clients in conversation. | Customer support | Undisclosed (DBS Joy) | rolled out to 350,000 corporate clients in Singapore; chats up 22 per cent and active users up 61 per cent in H1 2026 | 7 per cent reduction in calls or e-mail to customer service (company claim, as reported by the press) | The Business Times, 2026-07-28 | |
| City of Raleigh Public sector, United States | ServiceNow AI agents, including a triage agent named Ral-E, route and resolve the city's IT support requests. | IT | ServiceNow AI Agents | nearly half of IT support requests resolved by an AI agent | tickets routed correctly 98% to 99% of the time (company claim, as reported by the press) | Smart Cities Dive, 2026-07-24 | |
| SoftBank Corp. Telecom, Japan | AI agent on SoftBank's business website answers small business customers' questions on products, services, pricing and contracts. | Customer support | Sierra | n/a | self-resolution rate rose from around 50% at launch to 70% (the vendor's claim) | Sierra, 2026-07-14 | |
| Bunnings Retail, Australia | Buddy, a shopping agent, works out products from a plain-language project description or a photo of a handwritten list and fills the cart. | Sales | Google Cloud | live in Australia since April 2026 | more than doubled online conversion rates and lifted basket sizes (company claim, as reported by the press) | iTWire, 2026-06-30 | |
| CVS Health Healthcare and pharma, United States | Agentforce agents handle prior authorization and order status work and give contact center staff personalized caller data. | Operations | Salesforce Agentforce Health | Salesforce's largest Agentforce customer in a regulated industry | No figure given | TechTarget, 2026-06-25 | |
| Salesforce Technology, United States | Agentforce service agent on help.salesforce.com resolves customer support inquiries end to end, with handoff to human engineers for the rest. | Customer support | Salesforce Agentforce (its own product) | 4.3 million inquiries handled (1.5+ million support requests after the first year, Sept 2025) | resolved 70 percent of inquiries; CEO Marc Benioff said support headcount went from 9,000 to about 5,000 (the company's claim) | Salesforce, 2026-06-25; Salesforce, 2025-09-18; CNBC, 2025-09-02 | |
| C.H. Robinson Logistics, United States | An AI agent answers every customer freight quote request, a task employees previously covered only partly and slowly. | Sales | In-house | 100% of customer quotes | quotes answered in 31 seconds versus up to 20 minutes, up from about 60% of quotes answered (company claim, as reported by the press) | Fortune (via Yahoo Finance), 2026-06-24 | |
| Coinbase Fintech, United States | Cloud and local coding agents in Cursor create most pull requests under an agent-first engineering model. | Software engineering | Cursor | 75% of all PRs created by agents | some teams cut time from idea to production from 20 days to less than 2 days; 55% more PRs merged per engineer; 7 hours saved per engineer per week (the vendor's claim) | Cursor, 2026-06-23 | |
| The Adecco Group Professional services, Switzerland | AI agents run pre-screening, interviews, talent pool management, onboarding and redeployment across seven stages of recruitment in 10 countries. | HR | Undisclosed | 1.2 million candidate-agent interactions, including 250,000 completed interviews across 50,000 jobs | time-to-deliver cut by 50% and fill rates above 80% in lead markets; customer satisfaction 4.3 out of 5 (the company's claim) | The Adecco Group (PR Newswire release), 2026-06-18 | |
| EY (Global Delivery Services) Professional services, United Kingdom | EY's Global Delivery Services unit runs AI agents across its operations, with a target to double the count by 2028. | Other | In-house (EY.ai Agentic Platform) | around 50,000 AI agents deployed; target 100,000 by 2028 | No figure given | CIOL, 2026-06-15 | |
| Dubai Electricity and Water Authority (DEWA) Energy and utilities, United Arab Emirates | Government utility uses agentic AI in customer workflows, including processing security deposit refunds without human intervention. | Finance | Undisclosed | n/a | security deposit refund time cut from 4 days to 8 minutes without any human intervention (the company's claim) | DEWA (press release via Zawya), 2026-06-12 | |
| Hitachi Manufacturing and automotive, Japan | A quality-assurance AI agent searches past trouble cases, drafts response reports and analyses issues at Hitachi's Omika works before external sale. | Operations | In-house (HMAX Industry) | n/a | search time for trouble-response cases cut by about 90%; report writing and analysis time each cut by more than 80% (the company's claim) | Hitachi (press release via Nikkei), 2026-06-04 | |
pilot | R1 RCM Healthcare and pharma, United States | A Sierra voice agent replaces IVR menus, understanding patient requests in plain language and resolving or routing calls. | Customer support | Sierra | n/a | resolving up to 40% of incoming calls in early testing (the vendor's claim) | Sierra, 2026-05-18 |
| Paychex Professional services, United States | Voice AI agent processes routine payroll requests by phone, such as paying an employee for hours worked, for Paychex clients. | Customer support | Sierra | n/a | early results: 95% positive client satisfaction scores and 93% positive customer effort scores (the vendor's claim) | Sierra, 2026-05-14 | |
| PayPal Fintech, United States | Engineering teams use Cursor agents for large code upgrades and feature work. | Software engineering | Cursor | Java upgrade across 3,000 applications | 3,000-app Java upgrade completed 6x faster; high-adoption teams deploy daily, up from monthly releases (the vendor's claim) | Cursor, 2026-05-11 | |
| Microsoft Technology, United States | Employee Self-Service Agent is the first point of contact for IT support, answering and resolving issues before escalation to human help desk. | IT | In-house (Microsoft Employee Self-Service Agent) | more than 300,000 employees and vendors in 103 countries and regions | better first-touch answers and faster issue resolution, reducing escalations and ticket volume (qualitative) (the company's claim) | Microsoft Inside Track, 2026-05-07 | |
| ecoATM Retail, United States | Voice AI agent Echo answers inbound customer calls about device value, kiosk locations and data handling before they reach a human. | Customer support | Sierra | n/a | handling more than 30% of inbound calls within the first month, now in the mid-30% range (the vendor's claim) | Sierra, 2026-05-04 | |
| AES Energy and utilities, United States | A multi-agent system on Claude processes audit documents, breaks down tasks and generates reports for internal health and safety audits. | Operations | In-house (on Anthropic Claude via Google Cloud) | about 1,550 internal safety audits annually | 99% reduction in audit costs; audit results from 14 days to one hour; audit accuracy up by up to 20% (the vendor's claim) | Google Cloud, 2026-04-22; Google Cloud; Anthropic | |
| Danfoss Manufacturing and automotive, Denmark | AI agents automate email-based customer order processing, making most transactional order decisions without staff. | Operations | Go Autonomous on Google Cloud | operations in over 100 countries | 80% of transactional decisions automated; average customer response time cut from 42 hours to near real-time (the vendor's claim) | Google Cloud, 2026-04-22 | |
| Reliance (Reliance Retail) Retail, India | A shopping agent takes a goal such as planning a birthday party, guides the customer and assembles the cart. | Sales | Google Cloud | millions of customers | greater revenue per session (the vendor's claim) | Google Cloud, 2026-04-22 | |
| Telecom Italia (TIM) Telecom, Italy | A voice agent handles a large share of inbound customer calls. | Customer support | Google Cloud | n/a | efficiency up 20% (the vendor's claim) | Google Cloud, 2026-04-22 | |
| Engine Travel and hospitality, United States | Agentforce agent Eva handles hotel, flight and car rental support requests end to end, including cancellations and group booking changes. | Customer support | Salesforce Agentforce | over 800,000 annual customer support cases | fully handles 50% of chat cases with no human intervention; 15% decrease in average handle time; first version built in 12 days (the vendor's claim) | Salesforce, 2026-04-15 | |
| BNY Banking, United States | More than 140 'digital employees' on BNY's Eliza platform execute operations work such as payment repair and report to human managers. | Operations | In-house (Eliza platform) | more than 140 agents, each with roughly two dozen skills; about 100 managers oversee digital employees | CEO says a digital employee can do three quarters of a task in 10 minutes instead of two weeks (company claim, as reported by the press) | Business Insider, 2026-03-23; The Fintech Times, 2026-09-30 | |
| Starling Bank Banking, United Kingdom | Starling Assistant carries out banking tasks for customers, such as setting savings goals with automatic transfers and organising bill payments. | Customer support | In-house (on Google Gemini, Google Cloud) | rolled out to personal current account holders from 20 March 2026 | No figure given | The Paypers, 2026-03-23 | |
| Vivid Seats Retail, United States | AI agent recognizes logged-in fans, pulls up their orders and resolves common ticket questions such as locating tickets and delivery timelines. | Customer support | Sierra | n/a | containment increased by 40% and customer satisfaction rose by 35% since launch (the vendor's claim) | Sierra, 2026-03-17 | |
| Singtel Telecom, Singapore | AI assistant Shirley resolves mobile and home troubleshooting cases and completes roaming sign-ups across chat and voice. | Customer support | Sierra | over 70,000 customer cases in the first six weeks since launch | 73% of troubleshooting cases resolved without a Customer Care officer; 76% of roaming sign-up requests completed without one (the vendor's claim) | Sierra, 2026-03-04 | |
| Formula 1 Media and publishing, United Kingdom | Agentforce resolves routine fan support issues such as login or streaming problems; F1 also launched a fan companion agent. | Customer support | Salesforce Agentforce | n/a | 80% of queries handled within four hours; chat handling times reduced by 30% (the vendor's claim) | Salesforce, 2026-03-03 | |
| Box Technology, United States | Cursor rolled out organization-wide as Box's core IDE and agent platform for roadmap work and migrations. | Software engineering | Cursor | rolled out throughout the organization | 30-50% improvement in roadmap delivery and 80-90% faster migrations (the vendor's claim) | Cursor, 2026-02-13 | |
| Stripe Fintech, United States | Homegrown one-shot coding agents called Minions take tasks end to end and open pull requests that engineers merge. | Software engineering | In-house (Minions) | more than a thousand pull requests merged each week | No figure given | Stripe, 2026-02-09 | |
| NVIDIA Technology, United States | Cursor's coding agent finds and fixes bugs and writes code for NVIDIA developers. | Software engineering | Cursor | 30,000 developers | 3x more code committed when using Cursor (the vendor's claim) | Cursor, 2026-02-06 | |
| SoFi Fintech, United States | AI agent handles member support conversations across SoFi's banking, investing and lending products. | Customer support | Sierra | more than 50,000 conversations weekly | 61% containment; chat-contained NPS up 33 points (the vendor's claim) | Sierra, 2026-02-04 | |
| WEX Fintech, United States | Developers use GitHub Copilot agent mode and the Copilot coding agent to ship code more autonomously. | Software engineering | GitHub Copilot coding agent | 1,700+ seats | ~30% more code shipped; ~30% higher developer productivity; ~60% ROI (the vendor's claim) | GitHub, 2026-01-24 | |
| McKinsey & Company Professional services, United States | AI agents work alongside consultants across the firm; CEO Bob Sternfels counts them as part of a 60,000-strong workforce. | Other | In-house (QuantumBlack) | about 25,000 AI agents alongside 40,000 humans | No figure given | Business Insider, 2026-01-12 | |
| The Home Depot Retail, United States | Conversational AI across SMS, chat and phone understands customer intent and resolves service issues in real time, replacing menu-driven automation. | Customer support | Google Cloud (Gemini Enterprise for CX) | live across SMS, chat and phone | materially better engagement and resolution outcomes (the company's claim) | The Home Depot and Google Cloud (press release), 2026-01-11 | |
| Allianz Insurance, Germany | Project Nemo coordinates seven AI agents for coverage, weather, fraud and payout checks on small food spoilage claims in Australia. | Operations | Undisclosed (Allianz Project Nemo) | food spoilage claims under AUD 500 after catastrophe events in Australia, live since July 2025 | 80% reduction in claims processing and settlement time; claims processed in under a day, often within hours (company claim, as reported by the press) | Insurzine, 2025-12-31 | |
| Itau Unibanco Banking, Brazil | Devin runs code migrations, security fixes, test writing and documentation across the bank's software development lifecycle. | Software engineering | Cognition Devin | 300,000+ repos documented; 59 services migrated; 800 database objects migrated | .NET to Java migration 6x faster at 5x lower cost; about 70% of security vulnerabilities remediated automatically; test coverage up from ~50% to ~90+% (the vendor's claim) | Cognition, 2025-12-02 | |
| DHL Supply Chain (DHL Group) Logistics, Germany | AI agents autonomously handle phone and email work such as appointment scheduling, transport status calls and high-priority warehouse coordination. | Operations | HappyRobot | hundreds of thousands of emails and millions of voice minutes annually | No figure given | HappyRobot / DHL (GlobeNewswire release), 2025-11-25 | |
| Klarna Fintech, Sweden | Klarna's AI assistant resolves customer service chats end to end, such as refunds, returns and payment questions, inside the Klarna app. | Customer support | In-house | 28m annual conversations | solves 81% of customer service chats; does the equivalent work of 853 full time agents; $58m of annual cost savings; customer satisfaction on par with human agents (the company's claim) | Klarna (Q3 2025 investor presentation, SEC Form 6-K Exhibit 99.2), 2025-11-18 | |
| Spotify Media and publishing, Sweden | Background coding agent (Honk) in the Fleet Management platform makes code migrations and opens pull requests across repositories. | Software engineering | In-house (Honk background coding agent) | more than 1,500 merged AI-generated pull requests; hundreds of developers interact with the agent | total time saving of 60-90% on these migrations compared to writing code by hand (the company's claim) | Spotify Engineering, 2025-11-06 | |
| Rocket Mortgage Fintech, United States | Rocket's Digital Assistant guides homebuyers from first questions to preapproval across chat and voice, including outbound calls. | Sales | Sierra | more than 400,000 successful chat conversations and over one million outbound dials each month | conversion rates four times higher when clients use both AI chat and a banker (the vendor's claim) | Sierra, 2025-10-27 | |
| Chime Fintech, United States | AI agent resolves member questions such as direct deposits, account access and password resets instantly. | Customer support | Sierra | n/a | resolution rates increased from 50% to over 70% (the vendor's claim) | Sierra, 2025-10-23 | |
| Tubi Media and publishing, United States | AI agent handles viewer support, from streaming troubleshooting to password resets, and recommends content. | Customer support | Sierra | n/a | containment reached 80%; CSAT up seven percentage points; resolution times shifted from hours or days to minutes (the vendor's claim) | Sierra, 2025-10-15 | |
| Williams-Sonoma, Inc. Retail, United States | AI agents across its brands answer and resolve customer service chats; a sous chef agent named Olive offers cooking and hosting help. | Customer support | Salesforce Agentforce | brand portfolio including Williams Sonoma, West Elm and Pottery Barn | No figure given | Salesforce, 2025-10-14 | |
| OpenTable Travel and hospitality, United States | Agentforce agent autonomously resolves support inquiries from diners and restaurants, routing complex cases to human reps. | Customer support | Salesforce Agentforce | n/a | resolved 70% of diner and restaurant inquiries autonomously (the vendor's claim) | Salesforce, 2025-10-13 | |
| Reddit Technology, United States | Agentforce agent handles advertiser support cases, deflecting routine cases and speeding responses to advertisers. | Customer support | Salesforce Agentforce | n/a | deflected 46% of support cases; cut resolution times by 84%; average response time down from 8.9 minutes to 1.4 minutes (the vendor's claim) | Salesforce, 2025-10-13 | |
| Litera Technology, United States | Devin-based testing agents write tests and run regression work for legal software products. | Software engineering | Cognition Devin | 250 engineers; 5 major products in 3 weeks | 40% increased test coverage and 93% faster regression cycles (the vendor's claim) | Cognition, 2025-08-25 | |
| IBM Technology, United States | AskHR agent on watsonx Orchestrate automates more than 80 HR tasks and answers employee HR questions. | HR | In-house (IBM watsonx Orchestrate) | over 2.1 million employee conversations; more than 80 HR tasks automated | 94% containment rate of common questions; 40% reduction in HR operational costs over four years; 75% fewer support tickets since 2016 (the company's claim) | IBM, 2025-08-21 | |
| Accenture Professional services, Ireland | An internal agent platform (AATA) runs IT operations tasks such as VPN configuration and auto-remediates employee IT issues to close tickets faster. | IT | In-house (Accenture Advanced Technology Agent) | more than 100 active agents; supports roughly 800,000 employees | VPN configuration cut from 30 minutes to two minutes, more than a 93% reduction (company claim, as reported by the press) | CIO, 2025-07-18 | |
| AIG Insurance, United States | Claude is built into AIG's underwriting process to review submitted business faster and improve data accuracy for underwriters. | Finance | Anthropic Claude | early rollouts in underwriting | timeline to review business compressed by more than 5x; data accuracy improved from 75% to over 90% (the vendor's claim) | Anthropic, 2025-07-15 | |
pilot | Goldman Sachs Banking, United States | Autonomous AI software engineer Devin tested to join the bank's 12,000 human developers on coding work. | Software engineering | Cognition Devin | hundreds of Devins initially, possibly thousands depending on use cases | No figure given | CNBC, 2025-07-11 |
| 1-800Accountant Professional services, United States | Agentforce agent answers and resolves client support chats, such as tax return status questions, during tax season surges. | Customer support | Salesforce Agentforce | over 1,000 client engagements resolved in the first 24 hours | autonomously resolved 70% of administrative chat engagements during 2025 tax weeks; 90% case deflection rate during tax week (Salesforce, Oct 2025) (the vendor's claim) | Salesforce, 2025-06-23; Salesforce, 2025-03-04; Salesforce, 2025-10-13 | |
| Chipotle Mexican Grill Food and restaurants, United States | Ava Cado, an AI hiring agent, chats with candidates, collects information, schedules interviews for managers and sends offers to selected candidates. | HR | Paradox | n/a | hiring time reduced by 75% (company claim, as reported by the press) | CNBC (via NBC4 Washington), 2025-06-13 | |
| Heathrow Airport Travel and hospitality, United Kingdom | Agentforce agent Hallie answers traveller questions on WhatsApp, such as gate directions, amenities, wayfinding and flight status, without human transfer. | Customer support | Salesforce Agentforce | n/a | 90% chat resolution rate without human agent transfer; 40% improvement in digital contact efficiency; 50+% fewer chat exchanges to get an answer (the vendor's claim) | Salesforce, 2025-06-11 | |
| Rakuten Retail, Japan | Claude Code runs long autonomous coding jobs, including a complex open-source refactor, and speeds feature delivery for engineering teams. | Software engineering | Anthropic Claude Code | 7 hours of sustained autonomous coding on one refactoring task | average time to market for new features cut from 24 working days to 5 days, a 79% reduction (the vendor's claim) | Anthropic, 2025-05-22; Anthropic | |
| eHealth Insurance, United States | AI voice agents run initial screening calls with prospective insurance customers when staff cannot keep up with volume and after hours. | Sales | Undisclosed | n/a | No figure given | Mint (The Wall Street Journal syndication), 2025-05-21 | |
| Ramp Fintech, United States | Devin handles repetitive maintenance and technical debt tasks across the engineering organization and opens pull requests. | Software engineering | Cognition Devin | up to 80 merged PRs each week | 10,000+ hours saved each month on rote tasks (the vendor's claim) | Cognition, 2025-03-25 | |
| Airbnb Travel and hospitality, United States | LLM-driven automation pipeline with validation and retry loops migrated test files from Enzyme to React Testing Library. | Software engineering | In-house (on frontier LLMs) | 75% of target files migrated automatically in 4 hours on the first bulk run | migration estimated at 1.5 years of engineering time finished in 6 weeks (the company's claim) | Airbnb Engineering, 2025-03-13 | |
| Lyft Travel and hospitality, United States | Customer care AI assistant resolves rider and driver support inquiries and hands complex cases to human specialists. | Customer support | In-house (on Anthropic Claude via Amazon Bedrock) | thousands of daily customer inquiries | reduced customer service resolution time by 87% (the vendor's claim) | Anthropic, 2025-02-06 |
"Scale claimed" and "Result claimed" are copied as the source states them and attributed to whoever made the claim: the company, its vendor, or a press report quoting company executives. I have not audited any of them. Download every row as data.csv.
Breakdown
The same rows, counted four ways. Each deployment has one function, one industry and one main use case, so each table adds up to the total.
| Function | Deployments | Share | With a quantified result | |
|---|---|---|---|---|
| Customer support | 29 | 35% | 23 of 29 | |
| Software engineering | 17 | 20% | 15 of 17 | |
| Operations | 13 | 16% | 7 of 13 | |
| Sales | 10 | 12% | 4 of 10 | |
| IT | 5 | 6% | 4 of 5 | |
| Other | 4 | 5% | 2 of 4 | |
| HR | 3 | 4% | 3 of 3 | |
| Finance | 2 | 2% | 2 of 2 |
| Industry | Deployments | Share | |
|---|---|---|---|
| Banking | 10 | 12% | |
| Fintech | 10 | 12% | |
| Retail | 10 | 12% | |
| Technology | 8 | 10% | |
| Professional services | 7 | 8% | |
| Travel and hospitality | 6 | 7% | |
| Healthcare and pharma | 5 | 6% | |
| Manufacturing and automotive | 4 | 5% | |
| Media and publishing | 4 | 5% | |
| Telecom | 4 | 5% | |
| Energy and utilities | 3 | 4% | |
| Insurance | 3 | 4% | |
| Logistics | 3 | 4% | |
| Public sector | 3 | 4% | |
| Food and restaurants | 2 | 2% | |
| Consumer goods | 1 | 1% |
| Use case | Deployments | Share | |
|---|---|---|---|
| Customer service resolution | 23 | 28% | |
| Code generation and review | 13 | 16% | |
| Contact-centre voice agent | 6 | 7% | |
| Claims, refunds and authorisations | 5 | 6% | |
| Document, audit and compliance work | 5 | 6% | |
| Email and order processing | 5 | 6% | |
| Sales conversations and quotes | 5 | 6% | |
| Shopping assistant | 5 | 6% | |
| Code migration and modernisation | 4 | 5% | |
| Agent workforce across functions | 3 | 4% | |
| IT service desk | 3 | 4% | |
| IT operations automation | 2 | 2% | |
| Recruiting and hiring | 2 | 2% | |
| HR service desk | 1 | 1% | |
| Marketing campaigns | 1 | 1% |
| Vendor | Deployments | Share | |
|---|---|---|---|
| Salesforce | 11 | 23% | |
| Sierra | 11 | 23% | |
| Cognition | 7 | 15% | |
| Cursor | 4 | 9% | |
| 4 | 9% | ||
| Anthropic | 2 | 4% | |
| ServiceNow | 2 | 4% | |
| Bunkerhill Health | 1 | 2% | |
| GitHub (Microsoft) | 1 | 2% | |
| Go Autonomous | 1 | 2% | |
| HappyRobot | 1 | 2% | |
| Hippocratic AI | 1 | 2% | |
| Paradox | 1 | 2% |
Vendor shares are out of the 47 deployments that use an outside agent product; 26 built in-house and 10 with an undisclosed builder are left out. Model providers count as the vendor only when the company bought their agent product (for example Claude Code or ChatGPT Enterprise agents); a company that built its own agent on a model API counts as in-house.
Trend by quarter
| Quarter | Logged | |
|---|---|---|
| 2025 Q1 | 3 | |
| 2025 Q2 | 5 | |
| 2025 Q3 | 5 | |
| 2025 Q4 | 11 | |
| 2026 Q1 | 12 | |
| 2026 Q2 | 19 | |
| 2026 Q3 | 24 | |
| 2026 Q4 (to date) | 4 |
The date is when the deployment was announced or reported, not when it went live. Read the trend as companies talking about agents in public, not as the true rate of adoption.
What the log shows
Customer support is where agents are counted in public, and the numbers are not like for like. Klarna told investors in November 2025 that its assistant solves 81% of customer service chats, does the work of 853 full-time agents and saves $58 million a year. Salesforce says the Agentforce agent on its own help site has handled 4.3 million inquiries and resolved 70% of them, and Marc Benioff said in September 2025 that support headcount went from 9,000 to about 5,000. The TSA says its agent, Ace, resolves 96% of routine questions. Read those side by side with care: "resolved", "contained", "deflected" and "answered without handoff" are four different measures, and a vendor that reports containment is not reporting resolution.
Coding agents carry the boldest claims. Coinbase says agents create 75% of its pull requests (Cursor's customer story, June 2026). EBANX says 92% of merged pull requests on its core payments platform are created with Devin. Stripe's own engineering blog says its in-house agents, called Minions, account for more than a thousand merged pull requests a week, and Spotify's says its background agent has produced more than 1,500 merged pull requests. Most of these numbers come from the coding-tool vendors' own pages; Stripe and Spotify are the exceptions that published their own.
Support teams buy; back offices build. Of the support deployments where the builder is public, 22 of 26 run on a vendor's product, mostly Salesforce Agentforce or Sierra. Operations work goes the other way: BNY runs more than 140 "digital employees" on its own Eliza platform, Barclays routes about 120,000 operations emails a day through an in-house system on Claude, and UPS says on its earnings call that its customs agents now clear 90% of dutiable packages without manual review, up from about 21% six months earlier. Back-office work touches a company's own systems and data, so it gets built close to them.
Agent counts are not deployments. McKinsey says it runs about 25,000 agents alongside 40,000 people; EY's Global Delivery Services unit says about 50,000 agents are deployed across it. Those are real statements, logged as such, but a count of agents says little about what each one does. The rows that tell you most name one job, one volume and one before-and-after figure, like C.H. Robinson's quote agent answering every freight quote request in 31 seconds instead of about 60% of them in up to 20 minutes.
The evidence is mostly marketing, so weigh it. Nearly half the rows rest on a vendor's customer story, and those are the rows most likely to carry a number: 35 of the 41 vendor-sourced rows give a quantified result, against 16 of 28 rows sourced from press reports. Only 3 rows are labelled pilots, which tells you more about how companies announce things than about how many agents are still in pilot. If you are choosing a platform, the enterprise AI agent platforms guide and the AI agent pricing index compare what the vendors on this page charge, and the ROI case studies roundup covers how to read their claims.
Nearly every company here has a platform team. If you do not, the pattern worth copying is the narrow one: one job, one queue, a figure you can check. That is the gap I am building Gravity for, with agents built and kept working by their builders rather than by you. It is in private alpha; pricing starts with one free agent, and the cheapest AI agent platforms list covers the alternatives.
Change log
One dated line for each batch of rows added, and for each row a later source changed.
- : Log opened with 83 verified deployments announced or reported between 6 February 2025 and 2 October 2026, each read against its source before entry.
Methodology
Here is exactly what goes in, so you can decide how far to trust the counts.
What counts. A named company publicly announced, or was reported by the press quoting the company or its executives, to run an AI agent in production: software that takes actions or finishes work end to end, such as resolving a support conversation, writing and shipping code, triaging and fixing IT tickets, processing claims or invoices, qualifying leads, or answering and acting on HR requests. Each row records what the agent does and the scale or result the source claims.
What does not. Survey answers and "we are exploring AI" statements. Partnerships with no deployed agent. Plans and roadmaps. Assistants that only draft text for a person to send. Consumer chatbots sold by AI labs themselves. Pilots, unless the source calls it a pilot, in which case the row says so.
Vendor claims. A vendor's customer story or press release counts only when it names the customer. Those results are marked "the vendor's claim" in the table, because a vendor picks its happiest customers and its best numbers. Results from the company itself, its filings or its own blog are marked "the company's claim"; results from press reports quoting the company or its executives are marked "company claim, as reported by the press".
Sources. Every row cites one to three sources, each opened and read before the row went in, dated 1 January 2025 or later. The date in the table is the date of the main source, not the go-live date, which companies rarely give.
Classification. Function is the team the agent works for: customer support, sales, IT, finance, HR, operations, software engineering or other. "Built with" names the agent product when the company bought one, or says "In-house" when it built its own agent on a model API or framework. A result counts as quantified when the claim includes a number: a share, a time, a sum of money or a volume.
Refresh. Weekly. A script pulls candidate news stories each week; each candidate is checked against these rules and its source before a row is added, and the numbers on this page are recomputed from the rows. Nothing in the table is edited by hand. A newer source for an existing deployment updates that row and the change log says so.
Known gaps. This is a log of what companies chose to make public, so it skews to deployments that went well and to vendors with busy press teams. Failures rarely get a press release; for those, see the AI agent incidents database I keep alongside this one. Coverage is mostly English-language and skews to the US and Europe. The CSV includes every field so you can apply your own filters.
FAQ
Which companies are using AI agents?
As of 4 October 2026 this log names 82 companies with 83 deployments, among them Klarna, Salesforce, BNY, Barclays, Coinbase, Stripe, UPS, Walmart, Lowe's, Heathrow Airport, the TSA, Singtel and Novo Nordisk. Each row in the table says what the agent does, who built it, the scale and result the company or its vendor claims, and links the source.
What are the most common enterprise AI agent use cases?
Resolving customer service conversations end to end is the most common, with 23 of 83 deployments as of 4 October 2026. Writing and reviewing code comes second with 13, then contact-centre voice agents with 6. By team, customer support accounts for 29 deployments and software engineering for 17.
Which AI agent vendors do enterprises use most?
In this log, Salesforce Agentforce and Sierra tie with 11 named customer deployments each, followed by Cognition's Devin with 7, and Cursor and Google with 4 each. Of the 73 deployments where the builder is public, 26 were built in-house, four of them on Anthropic's Claude models.
Do enterprise AI agents actually deliver results?
Companies say they do: 60 of the 83 deployments (72%) come with a quantified result, such as a share of chats resolved or hours saved. None of those figures is audited, and rows sourced from a vendor's own customer story carry a number far more often than rows from press reports, so treat them as claims.
Is this a complete list of companies using AI agents?
No. It is a log of deployments that companies or their vendors chose to make public, checked against a dated source. Most agents in production are never announced, failures are rarely announced at all, and coverage skews to English-language sources and US companies.
Can I use this data?
Yes. The table and the CSV are licensed CC BY 4.0: copy them, chart them, quote the numbers, and credit "Gravity enterprise AI agent deployments log" with a link to this page. Please quote each result as the company's or vendor's claim, the way the table does.
How do I add a deployment to the log?
Email support@gravity.fast with a link to a public, dated source: a company press release, an earnings call, a vendor customer story that names the customer, or press reporting that quotes the company. I check each one against its source and the rules in the methodology before it goes in.
Sources
- Per-deployment sources: linked in each table row and listed in full in data.csv (up to three per row, with publisher, date and whether the claim came from the company, the vendor or the press).
- Klarna, Q3 2025 investor presentation, SEC Form 6-K Exhibit 99.2, 18 November 2025: sec.gov
- Salesforce, "Agentforce Help Agent" announcement, 25 June 2026: salesforce.com; CNBC, Benioff on support headcount, 2 September 2025: cnbc.com
- Stripe, "Minions: Stripe's one-shot, end-to-end coding agents", 9 February 2026: stripe.dev
- Business Insider, BNY's digital employees, 23 March 2026: businessinsider.com
- Vendor customer-story indexes used to find named customers (each row links its own page): Sierra, Cognition (Devin), Salesforce newsroom, Cursor, all read 4 October 2026