AI agent examples: 24 real ones running inside businesses, and what they actually achieved.

Twenty-four deployments at named companies, each with what the agent does, the figure the company reported and the caveat that usually gets left out, reversals included.

Search for AI agent examples and you mostly find lists of what agents could do. This one is a list of what they have done: 24 deployments at named companies, each with what the agent actually does, the figure the company reported, and the caveat that usually gets left out. Several of those caveats are reversals. A bank apologised and reversed the job cuts it had announced, a fintech went back to hiring people, and a checkout feature was retired after six months. They belong on the list as much as the successes, because they show where the line currently sits. Every example has a source at the bottom of the page, and where the only source is the company or the vendor that sold it the agent, we say so.

All 24 examples at a glance

#CompanyAreaWhat the agent doesReported resultWho has the last word
1KlarnaCustomer serviceAnswers customer service chats in 35+ languages2.3 million chats in month one, two-thirds of all chats; human agents recruited again in 2025The agent, with a human on request
2SalesforceCustomer serviceResolves questions on Salesforce’s own help site1 million+ conversations, 84% resolved, 4% handed to a humanThe agent
3LyftCustomer serviceFirst-line support for riders and drivers87% shorter average resolution timeThe agent; complex cases go to people
4IKEACustomer serviceAnswers routine questions (Billie)47% of queries in its first two years; about 8,500 staff retrained, not cutThe bot; complex cases go to people
5Commonwealth Bank of AustraliaCustomer serviceVoice bot for contact-centre callsCall volumes rose; 45 job cuts reversed with an apologyReversed
6CursorCustomer serviceFront-line email support, signing as “Sam”Invented a policy that did not exist; public apologyThe agent, until it was wrong
7SalesforceSalesWorks dormant leads (SDR agent)43,000+ leads worked, $1.7 million in new pipelineThe agent
8VerizonSalesAssists 28,000 service reps on screenSales through the service team up nearly 40%The rep it assists
9AmazonSoftware developmentUpgrades Java applications (Q Developer)4,500 developer-years and $260 million a year saved, by Amazon’s estimateA human reviewer
10GoogleSoftware developmentWrites internal code-migration changes80% of code changes AI-written; about 50% less timeA human reviewer
11NubankSoftware developmentRefactors an ETL monolith (Devin)8–12x engineering-time efficiency, 20x cost savingA human reviewer
12StripeSoftware developmentWrites pull requests end to end (Minions)1,000+ merged pull requests a weekA human reviewer
13SaaStr, on ReplitSoftware developmentBuilds an app from promptsDeleted the production database during a code freezeNobody, which was the problem
14GoogleSecurityHunts for software vulnerabilities (Big Sleep)Found an SQLite flaw known only to attackersNot published
15BNYFinanceValidates payments and repairs code100+ “digital employees”; 10% of payment validationsNot published
16AllianzInsurance claimsSeven agents process small food-spoilage claims80% less claim processing and settlement timeA person approves every payout
17IBMHRAnswers and completes HR requests (AskHR)11.5 million interactions in 2024, 94% containedThe agent for routine requests
18C.H. RobinsonLogisticsPrice quotes, orders and appointments3 million+ shipping tasks; orders in under 90 secondsThe agent
19Taco BellOperationsTakes drive-thru orders by voiceAbout 890 restaurants; an 18,000-cup order went viralThe agent; staff can step in
20McDonald’sOperationsTakes drive-thru orders (Archy)IBM test ended in 2024; new agent “above 90%” accurateThe agent
21AmazonShoppingShopping assistant (Rufus) and Buy for Me250 million customers; $10 billion+ incremental annualised sales, by Amazon’s estimateThe shopper
22OpenAI, Etsy, WalmartShoppingCheckout inside ChatGPTPivoted away from after six months; conversion three times lowerRetired
23Morgan StanleyInternal knowledgeAdvisor assistant and meeting notes98% of advisor teams; about half an hour saved per meetingThe advisor
24UberInternal knowledgeAnswers engineers’ on-call questions (Genie)70,000+ questions, 13,000 hours saved; 48.9% rated helpfulThe engineer

Results are as reported by each company, its vendor or the outlet named in the sources, checked October 2026. Most are self-reported and none is an audited figure. “Who has the last word” says who makes the final decision in the process: the agent itself, a human reviewing its work, or the human it assists.

How we chose them

An AI agent, for this list, is software built on a language model that does more than answer. It is given a goal, it uses tools — an order system, a code repository, a payments queue — and it decides the next step itself. Anthropic’s engineers draw the line between workflows, where the model follows a path written in advance, and agents, where the model directs its own process and tool use. The longer explanation, and where the line with a plain chatbot falls, is in AI agents vs chatbots. This article only needs the working test: does it act, or does it only talk?

Every entry had to pass three rules. The company is named. The agent is in real use, not a demo or a press-release pilot. And the result is documented by the company itself, by the vendor that built it, or by a major outlet that reported it. We left out anything whose only evidence was an analyst’s projection or a vendor’s “a leading retailer”, and we left out several well-known names because no figure for them could be traced back to a primary source.

Not every entry is fully autonomous, and pretending otherwise would be the hype this list is meant to avoid. Some are assistants that a company calls an agent; some are agents that a human signs off. That is why the table has a “Who has the last word” column, and it is worth reading before the results column. The figures themselves are mostly self-reported: treat them as what each company says happened, which is useful, and not as what will happen to you, which nobody can promise.

Customer service: the three big numbers

Klarna is the example everybody quotes, so it goes first, with the whole story. In February 2024, a month after launch, Klarna said its AI assistant had handled 2.3 million conversations — two-thirds of its customer service chats — doing the work of 700 full-time agents, in 23 markets and more than 35 languages. Resolution time fell from 11 minutes to under two, repeat enquiries fell by 25%, and Klarna estimated a $40 million profit improvement for 2024. Then, in May 2025, chief executive Sebastian Siemiatkowski told Bloomberg that cost had been “a too predominant evaluation factor”, that quality had suffered, and that Klarna would recruit human agents again so customers always have the option of a person. The assistant stayed. The claim that it could replace the team did not.

Salesforce runs its own agent product, Agentforce, on its help site. By April 2025 it had handled more than 500,000 conversations, and by July more than a million, with 84% resolved without escalation and 4% handed to a human support engineer. In September 2025 Marc Benioff said on The Logan Bartlett Show that he had reduced support “from 9,000 heads to about 5,000”; the company added that it had redeployed hundreds of people into professional services, sales and customer success. Two caveats come from Salesforce itself: the 4% handoff is deliberate, with renewal questions routed to people on purpose, and the agent only worked once 740,000 pieces of help content had been organised for it. Remember too that Salesforce sells Agentforce, so its own help desk is also its best advertisement.

Lyft added a Claude-powered assistant to its customer care in early 2025, handling the first line of support for more than 40 million riders and a million drivers and passing complex cases to human specialists. Lyft reported an 87% reduction in average resolution time. It is a good number, and also an average: it says how fast the large mass of simple requests now closes, not how the hard ones fare. It comes from the announcement of the partnership, not from an independent review.

Customer service: the reskilling, the reversal and the invented policy

IKEA’s Billie is closer to a chatbot than an agent, and it is here because it is the clearest counter-example to the replacement story. Since 2021 it has answered routine questions — opening hours, order status, stock, returns — and it assisted 47% of customers in its first two years, and about three-quarters now. Instead of cutting the call centre, IKEA retrained about 8,500 staff to handle complex service and to sell remote interior-design advice; those remote-sales centres brought in €1.25 billion in the last fiscal year, up from €1.08 billion. The honest caveat is time: the reskilling took about two years, and it only works for a company with something more valuable for those people to do.

The Commonwealth Bank of Australia went the other way. In July 2025 it announced 45 contact-centre redundancies on the strength of a new AI voice bot. Call volumes then rose rather than fell, with managers offering overtime and pulling team leaders onto the phones. After the Finance Sector Union took the case to the Fair Work Commission, the bank said in August that it “did not adequately consider all relevant business considerations”, apologised, and offered the 45 the choice of staying, redeployment or leaving. The bot was not the mistake. Cutting the people before the bot had proved itself was.

Cursor, which makes an AI code editor, shows that knowing the technology is no protection. In April 2025 its front-line AI support agent, signing its emails as “Sam”, told users who were being logged out that this was expected behaviour under a new login policy. There was no such policy; the logouts were a bug. The answer spread across Reddit and Hacker News, users cancelled, and co-founder Michael Truell apologised for “an incorrect response from a front-line AI support bot”. A support agent that is allowed to explain policy will, sooner or later, explain a policy that does not exist, unless it is limited to sources it can cite and clearly labelled as a bot.

Sales: the leads nobody called, and the reps who started selling

Salesforce also points an agent at its own sales pipeline. By September 2025 its SDR agent had worked more than 43,000 leads and generated $1.7 million in new pipeline from accounts that had gone dormant. Salesforce’s own first-year review admits the rough start: when asked for details on a lead, the agent answered “I don’t know” 30% of the time, a figure the team got below 10% only through repeated tuning. And pipeline is not revenue; it is the start of a conversation a salesperson still has to finish.

Verizon’s example is an assistant rather than an agent, and it is on the list because it is one of the few with a sales figure attached. Verizon fed nearly 15,000 internal documents into a version of Google’s Gemini and put the result on the screens of its 28,000-person customer service team, starting in July 2024 and at full scale by January 2025. In April 2025 the head of its consumer group said sales through that team had risen by nearly 40%, describing it as “reskilling in real time from customer care agents to selling agents”. The rep still talks to the customer and makes the sale; the assistant shortens the search for the answer.

Buying an “AI sales rep” off the shelf deserves more scepticism than either example. In March 2025 TechCrunch reported that 11x, a heavily funded AI SDR start-up, had listed ZoomInfo and Airtable as customers when both said they were not; ZoomInfo said it had run a one-month trial in which the product underperformed its human SDRs. Former employees described losing 70–80% of the customers who signed up. 11x put the logos down to human error and said its retention was 79%. Whatever the truth, the lesson for a buyer is simple: ask for a customer you can phone.

Software development: the migrations nobody wanted to do

Amazon used the code-transformation agent in its own Q Developer to move tens of thousands of production applications from Java 8 or 11 to Java 17. Andy Jassy wrote that the average upgrade went from 50 developer-days to a few hours, that the work saved more than 4,500 developer-years, and that 79% of the auto-generated code reviews shipped without any further change. Amazon puts the saving at $260 million a year and calls both estimates conservative. Two caveats: Amazon sells Q Developer, and the other 21% of reviews needed a developer.

Google published the most careful account in this group, a January 2025 paper by its own engineers. For a migration of ID fields from 32-bit to 64-bit integers in Google Ads, 80% of the code changes that landed were written by AI and the total time fell by about half; in a test-framework migration, about 87% of the AI-generated code was committed unchanged. The authors are also explicit about the limit: code review and rollout “still require a human operator and can quickly become bottlenecks”.

Nubank, the Brazilian digital bank, had an eight-year-old data-pipeline monolith of more than six million lines, roughly 100,000 data classes to migrate, and a plan that spread the work over 18 months and more than a thousand engineers. Engineers wrote example migrations, Cognition’s agent Devin was fine-tuned on them, and engineers then reviewed and merged its changes. The reported result is an 8–12x improvement in engineering time and a 20x cost saving on the work handed to the agent. The source is Cognition’s own case study, which is worth remembering when quoting it.

The three have the same shape, and it is why migration is the best-documented agent use in software. The correct result is defined before the work starts, the existing tests say whether it worked, and the job is repetitive enough that nobody wants to do it by hand. Very little ordinary product work has all three properties.

Software development: pull requests, a database and a zero-day

Stripe’s coding agents, which it calls Minions, produce more than a thousand merged pull requests a week, according to a February 2026 post from its engineering team. An engineer tags the agent in Slack; it works on an isolated development machine that starts in about ten seconds, reaches more than 400 internal tools through one central MCP server, and gets at most two rounds of the test pipeline before handing its work to a human reviewer. If MCP is new to you, how an MCP server differs from the API underneath it is the short version. The caveat is the infrastructure: the agent is the last layer on top of years of developer tooling that most companies do not have.

The other end of the scale is SaaStr founder Jason Lemkin’s experience in July 2025. Vibe coding an app with Replit’s AI agent, he declared a code and action freeze; the agent deleted the production database anyway, produced fake data and fake reports, and said a rollback was impossible, which was untrue. Replit’s chief executive Amjad Masad responded with automatic separation of development and production databases, and said Lemkin would be compensated. The lesson is older than AI: nothing, human or agent, should hold delete rights on production that it does not need. We cover the wider risk in AI-generated code security.

Security is where an agent’s tirelessness counts most. Google’s Big Sleep, an agent that searches software for unknown vulnerabilities, found a flaw in the widely used SQLite database engine, CVE-2025-6965, that Google says was known only to threat actors. In July 2025 Google called it the first time an AI agent had been used to directly foil efforts to exploit a vulnerability in the wild, without saying how. Attackers are using the same tools, which is the subject of how AI agents changed hacking.

Finance and back office: claims, payments and HR

BNY, the custody bank, had more than 100 AI “digital employees” by October 2025, working on payment validations and code repairs, and 117 AI solutions in production at the end of September. Microsoft’s customer story on the bank adds that digital employees now handle 10% of payment validations, that client settlement enquiries are resolved 80% faster and that onboarding is 20% quicker. Those last figures come from a vendor’s case study, and 10% is a share of the work, not the job.

Allianz has the best-designed example on this list. Project Nemo, launched in Australia in July 2025 and built in under 100 days, uses seven agents — planner, cyber, coverage, weather, fraud, payout and audit — to process food-spoilage claims under AUD 500 after severe weather. A claim goes from filing to a human reviewer in under five minutes, processing and settlement time fell by 80%, and, by design, payout decisions are never automated. Allianz is now looking at travel delays, simple motor claims and property damage, all of them low-value and high-volume.

IBM’s AskHR handled 11.5 million employee interactions in 2024, contained 94% of common questions and processed more than a million HR transactions, from employee letters to compensation changes, across about 80 automated tasks. IBM credits it with a 40% fall in HR operating costs over four years. Arvind Krishna told the Wall Street Journal that AI had replaced several hundred HR staff, and also that IBM’s total employment had gone up, with the money moved into programmers and salespeople. What happened to the remaining 6% is the subject of the 94% that AI finished and the 6% it could not.

Operations: freight, and the drive-thru

C.H. Robinson, the freight logistics company, said in April 2025 that its generative AI agents had performed more than three million shipping tasks: a million price quotes, a million orders and a million pickup and delivery appointments. For more than 5,200 customers, orders are now processed in under 90 seconds instead of waiting up to four hours. Bear the unit in mind: a task is the company’s own measure, and a quote delivered is not a load won.

Taco Bell has used voice AI to take drive-thru orders since 2023, with Omilia, and by July 2026 it was live in about 890 US restaurants in 38 states. Omilia says transactions are as fast as or faster than with a person, and that staff retention is better where the system runs. In August 2025 a customer ordered 18,000 cups of water and the clip went viral; Taco Bell’s chief digital and technology officer Dane Mathews then told the Wall Street Journal that the company was thinking harder about where to use it, and that busier restaurants may be better served by a person.

McDonald’s ended its automated order-taking test with IBM in about 100 restaurants in July 2024, saying a voice ordering solution would still be part of its future. In September 2026 it announced that future: Archy, a voice agent built with Google, which takes drive-thru orders in English and Spanish and checks orders before they leave the window. It has completed more than a million transactions in five test restaurants at what McDonald’s calls above 90% accuracy, and the US rollout starts in 2027. Above 90% can still mean nearly one order in ten going wrong, and the figures are the company’s own and unaudited.

Shopping agents: discovery works, checkout went back to the shop

Amazon’s shopping assistant Rufus had 250 million customers in 2025, Andy Jassy told investors that October, with monthly users up 140% and interactions up 210%. Customers who use it during a shopping trip are 60% more likely to complete a purchase, and Amazon says it is on track for more than $10 billion in incremental annualised sales. Its Buy for Me feature goes further: it surfaces items Amazon does not stock and buys them on the other merchant’s own website. The caveat on the 60% is selection: people who ask a shopping assistant for help are already in a buying mood, and “incremental” is Amazon’s estimate.

The reversal in this group is ChatGPT’s Instant Checkout. OpenAI launched it in September 2025 with Etsy and Shopify merchants, letting a purchase complete inside the chat. In March 2026 OpenAI moved away from it, saying the first version “did not offer the level of flexibility that we aspire to provide”. Walmart’s Daniel Danker said its conversion rate had been three times lower inside the chatbot than when shoppers clicked out to the retailer. Walmart now brings its own agent, Sparky, into ChatGPT instead, with the account, loyalty programme and payment handled in Walmart’s own environment.

The pattern for a shop owner is that AI agents increasingly do the research and send the buyer to your checkout, which is where the sale still happens. The protocols behind that, and what a shop needs in order to be found, are in how AI agents buy things.

Internal knowledge: assistants for the people who know the business

Morgan Stanley rolled out its AI @ Morgan Stanley Assistant to financial advisors in September 2023, giving them fast access to the firm’s research and documents; 98% of advisor teams have adopted it. In June 2024 it added Debrief, which, with the client’s consent, takes notes in meetings, summarises the key points, drafts a follow-up email for the advisor to edit and send, and saves a note into Salesforce. One advisor quoted in the launch announcement said it saved him about half an hour per meeting. That is one person’s figure, not a firm-wide average, and the advisor still decides what goes to the client.

Uber’s Genie answers engineers’ questions in the Slack channels where on-call teams are asked for help, searching internal documentation and citing its sources. By October 2024 it had answered more than 70,000 questions across 154 channels and saved an estimated 13,000 engineering hours. Uber also published the number most vendors would leave out: users rated 48.9% of answers helpful. It is built on retrieval over the company’s own documents, the approach explained in RAG for business.

Internal knowledge is the safest place in this list to start. The reader is an employee who knows the business, can check the cited source and can ignore a bad answer, so a mistake costs a few minutes rather than a customer.

What the ones that worked have in common

The successes are narrow. Java upgrades, food-spoilage claims under AUD 500, freight quotes, payment validations, routine HR letters: each is high-volume, repetitive and has a right answer that can be checked. None of the agents above is running a business; each is doing one well-defined job inside one.

Every reversal on this list has the same shape: the agent was given more of the job than it had shown it could do. The Commonwealth Bank cut staff before the bot had proved itself. Cursor let a bot explain policy. Klarna judged its assistant mainly on cost, by its own chief executive’s account. Instant Checkout took over the part of the purchase that retailers wanted to keep.

And the most convincing deployments keep a person at a defined point. Allianz never automates the payout. Stripe, Google and Nubank have a human review every change. Morgan Stanley’s advisor sends the email. That is not a sign the agents are weak; it is the design that lets them run at volume without a single bad case becoming a headline.

How to read an agent success story

Check who published the number. On this list, Nubank’s figures come from Cognition, which sold it the agent; BNY’s from Microsoft; Salesforce’s and Amazon’s describe their own products. None of that makes them false, but it puts them in a different category from Google’s paper, which states its own limits, or from Uber, which published a 48.9% helpfulness rate.

A figure published by the company that sells the agent is a claim to check, not a result to plan around. Ask what the unit is. A “task”, a “conversation” or an “interaction” can be counted many ways; an average resolution time is dominated by the easy cases; “pipeline” and “incremental sales” are estimates, not revenue.

Expect some of these to change. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, and its analysts point to “agent washing” — vendors rebranding existing chatbots and assistants as agents — as part of the reason. The examples above are the ones that survived long enough to publish a result, and even among those, several ended in a reversal.

What a small business can realistically deploy

Not Stripe’s 400 internal tools or Allianz’s seven agents. Four patterns from this list do scale down, though. A support assistant that answers from your own help content and hands off to a person, as Lyft and Salesforce do. A follow-up agent for enquiries that would otherwise go cold, Salesforce’s SDR idea at the size of a contact form. A back-office workflow with one AI step inside it, such as reading order emails or invoices into your system, the smaller cousin of Allianz and C.H. Robinson. And an internal assistant over your own documents, like Genie.

Start with a task you could already write down as a checklist; if nobody can describe the steps, an agent will not discover them. Every success above had that: a defined input, a defined output and someone who could say whether the result was right. Our guide to where B2B process automation with AI pays off walks through how to pick that first process.

Most of these do not need a custom agent at all. A fixed workflow in n8n, Make or Zapier with a language model at one step is cheaper, more predictable and easier to debug, and Zapier vs Make vs n8n compares the three on price and on where your data ends up. For customer-facing questions, where AI chatbots for business genuinely pay off covers the simpler tool that is often enough.

Whatever you build, copy the guardrails from the good examples rather than the headlines from the bad ones. Give it read-only access first, put a person in front of anything that moves money or makes a promise, label it as a bot, keep a log of what it did, and test it on your last fifty real cases before it sees a customer.

What it costs

The platforms price agents per use, and the figures are public. Intercom’s Fin agent costs $0.99 per outcome, on top of a helpdesk seat from $29 a month billed annually. Salesforce charges $2 per conversation for customer-facing Agentforce agents, or $500 per 100,000 Flex Credits with each action costing 20 credits, about $0.10 an action. n8n Cloud starts at €20 a month billed annually for 2,500 workflow executions, and the self-hosted Community Edition is free. Prices checked October 2026.

Our own packages are scoped to the patterns above, and named for what they are. FAQ Chatbot is €209 €299 30% off: an assistant trained on up to 50 of your FAQs or documents with escalation to a person, which is a chatbot rather than an agent, because it answers rather than acts. Workflow Automation is €209 €299 30% off for one repeating workflow in n8n or Make connecting two or three of your tools, which is where an AI step inside a fixed process usually starts. n8n Automation is €209 €299 30% off for a single two-platform workflow, and Custom Text AI Model is €349 €499 30% off for a model built on your own documents and support history. AI usage and platform subscriptions are paid separately to the providers in every case.

The bill that decides whether an agent pays off is neither of those. It is the time spent testing it on real cases before launch, and the person who owns the exceptions afterwards. The Commonwealth Bank and Cursor stories above are what skipping them looks like.

Where we would start

If you have a support inbox full of the same twenty questions, start there, with a person one click away. If you have a back-office process with structured inputs, such as orders, invoices or bookings, automate it with an AI step and a human approval. If you are tempted to start by cutting the team, reread the Commonwealth Bank paragraph first.

That is the work we do in AI automation. Tell us the process you have in mind and we will tell you honestly whether it needs an agent, a chatbot, a plain automation without any AI in it, or nothing at all.

Frequently asked questions

Sources

Each example’s figures come from the company’s own announcement, the vendor’s case study or a major outlet’s report, checked October 2026. Where the only source is the company or the vendor that sold the agent, the text says so. Prices are from each vendor’s public pricing page, checked October 2026.

Wondering which of these would work in your business?

Tell us the process you have in mind. We will say honestly whether it needs an agent, a chatbot, a plain automation, or nothing at all, and build it if it does.