Short answer. The AI agent examples with published results share one shape: one trigger, a short list of tools, a clear line on what the agent may not do alone, and one number that shows whether it works. Below are ten real agents broken down that way: eight from companies that published how theirs work (Klarna, Salesforce, IBM, Ramp, Uber, Shopify, Anthropic, Carbonorm), one we run ourselves, and Air Canada as the warning. Then six patterns a smaller business can run, and how to pick your first one.
What counts as an AI agent example
An AI agent is software that uses a language model to decide its next step and calls tools to carry it out: it looks up an order, queries a database, files a ticket. A bot that only answers from an FAQ is a chatbot, and a fixed sequence of steps is a workflow. The difference is explained with a real run in agentic AI vs AI agents vs generative AI.
Many "AI agent examples" lists mix in thermostats, Mars rovers and product brochures. This one only includes LLM agents whose owners published how they work and what came out of them, plus one of our own that we can show from the inside.
10 real AI agent examples at a glance
- #: 1Agent: KlarnaJob: Customer service chatsLimit on autonomy: Since 2025 a person is always availablePublished result: 2.3M chats in month one; resolution under 2 min vs 11
- #: 2Agent: SalesforceJob: Support on its own help siteLimit on autonomy: Escalates to peoplePublished result: Escalations about 10,000 → 5,000 a week
- #: 3Agent: IBM AskHRJob: Employee HR requestsLimit on autonomy: Sensitive cases go to an HR personPublished result: 94% of common questions contained
- #: 4Agent: RampJob: Fix wrong merchant matchesLimit on autonomy: Fixed action list; high-impact changes restrictedPublished result: Under 10 s instead of hours
- #: 5Agent: Uber FinchJob: Finance data questions in SlackLimit on autonomy: Expert validation on requestPublished result: Sub-agents tested against "golden" queries
- #: 6Agent: Shopify SidekickJob: Run a store from the adminLimit on autonomy: Changes staged for the merchant to confirmPublished result: Weekly active shops 4× year on year
- #: 7Agent: Anthropic ResearchJob: Research across many sourcesLimit on autonomy: Lead agent plans and checksPublished result: +90.2% vs one agent; ~15× chat tokens
- #: 8Agent: CarbonormAIJob: EU packaging complianceLimit on autonomy: The grade comes from a rule engine, not the modelPublished result: Same packaging, same grade; every quote checked
- #: 9Agent: Semroi CMS agentJob: Edit this website's contentLimit on autonomy: Drafts only; a person publishesPublished result: This article, written as a draft
- #: 10Agent: Air CanadaJob: Policy answers on the websiteLimit on autonomy: None, and that was the problemPublished result: Airline held liable; CA$812.02
| # | Agent | Job | Limit on autonomy | Published result |
|---|---|---|---|---|
| 1 | Klarna | Customer service chats | Since 2025 a person is always available | 2.3M chats in month one; resolution under 2 min vs 11 |
| 2 | Salesforce | Support on its own help site | Escalates to people | Escalations about 10,000 → 5,000 a week |
| 3 | IBM AskHR | Employee HR requests | Sensitive cases go to an HR person | 94% of common questions contained |
| 4 | Ramp | Fix wrong merchant matches | Fixed action list; high-impact changes restricted | Under 10 s instead of hours |
| 5 | Uber Finch | Finance data questions in Slack | Expert validation on request | Sub-agents tested against "golden" queries |
| 6 | Shopify Sidekick | Run a store from the admin | Changes staged for the merchant to confirm | Weekly active shops 4× year on year |
| 7 | Anthropic Research | Research across many sources | Lead agent plans and checks | +90.2% vs one agent; ~15× chat tokens |
| 8 | CarbonormAI | EU packaging compliance | The grade comes from a rule engine, not the model | Same packaging, same grade; every quote checked |
| 9 | Semroi CMS agent | Edit this website's content | Drafts only; a person publishes | This article, written as a draft |
| 10 | Air Canada | Policy answers on the website | None, and that was the problem | Airline held liable; CA$812.02 |
Customer service and HR
1. Klarna: customer service at scale, then a step back
On 27 February 2024 Klarna reported that its OpenAI-powered assistant had handled 2.3 million conversations in its first month, two-thirds of its service chats: the work of 700 full-time agents, customer satisfaction on par with people, 25% fewer repeat inquiries, and errands resolved in under 2 minutes instead of 11. It handles refunds, returns, payment issues, cancellations and disputes in 35+ languages.
In May 2025 the company changed course: its CEO said cost-driven automation had led to lower quality, and Klarna began recruiting people for customer service again so that customers "always" have a human if they want one. Lesson: handle time moved first, quality on hard cases mattered most. Measure the hard cases separately and make the route to a person easy to find.
2. Salesforce: an agent on its own help site
Salesforce runs its own agent platform, Agentforce, on help.salesforce.com. In December 2024 its CEO said the help site handled 32,000 conversations a week, and that escalations to people had fallen from about 10,000 to about 5,000 a week. In 2025 he said customer service headcount had gone from about 9,000 to about 5,000 as agents took over conversations.
Lesson: the number Salesforce reports is escalations, not conversations. Escalation rate per week is the simplest health metric for a support agent, and it also tells you how many people you still need.
3. IBM AskHR: 80 HR tasks behind one chat
AskHR is IBM's internal HR agent. It classifies an employee's request, routes it to a domain (benefits, payroll, career) and either answers or triggers one of around 80 automated tasks: employment letters, vacation requests, payroll access, and manager workflows such as compensation changes. IBM reports a 94% containment rate for common questions, 75% fewer support tickets since 2016, 11.5 million interactions in 2024, and a contribution to 40% lower HR operating costs over four years.
Sensitive topics such as low performance or misconduct are sent to an HR person, not answered. Lesson: write down the topics your agent never handles, and route them before it tries.
Finance and operations
4. Ramp: an agent that fixes records in seconds
Ramp, a corporate card company, built an agent for one job: a customer reports that a transaction was matched to the wrong merchant. The model receives the transaction and candidate merchants pulled with embeddings and fast database queries, and may take only one of a fixed set of actions. Low-impact actions are allowed on nearly every request; high-impact ones only in carefully chosen cases, and a post-processing step catches hallucinated merchants.
Ramp reports the agent handles nearly all of these reports correctly and unblocks customers in under 10 seconds, where they used to wait hours for a person. The team started by reviewing answers by hand at tens of requests a day. Lesson: a short action list is what makes an agent safe enough to act on live records.
5. Uber Finch: finance questions answered in Slack
Finch lets Uber's finance teams ask for data in plain language in Slack. A supervisor agent routes each question to a sub-agent, such as a SQL writer or a document reader, and returns the result. Each sub-agent is tested against a set of expected answers (comparing the query's results, not its wording), routing is tested separately, and simulated questions check the whole system before updates ship.
Uber is open that Finch is "not 100% reliable" and plans a "Request validation" button that sends an answer to an expert for checking. Lesson: test the routing as well as the answers, and let people ask for a human check when a number matters.
E-commerce and research
6. Shopify Sidekick: the agent proposes, the merchant confirms
Sidekick is the agent inside the Shopify admin. Since the Spring '26 Edition apps such as Klaviyo, Loop and Judge.me can plug into it: data extensions let Sidekick answer with an app's data, and action extensions let it "stage changes for confirmation", so the merchant approves before anything changes. Shopify says the number of weekly active shops using Sidekick was up 4× year on year in Q1 2026.
Lesson: when an agent acts inside someone else's business, propose-then-confirm is the default. For what this means for stores selling through AI assistants, see agentic commerce.
7. Anthropic Research: several agents, at a price
In Anthropic's research system, a lead agent plans the work and starts sub-agents that search in parallel, then writes the answer from what they return. On Anthropic's internal research evaluation, a Claude Opus 4 lead with Claude Sonnet 4 sub-agents outperformed a single Claude Opus 4 agent by 90.2%.
The cost is spelled out too: agents typically use about 4× more tokens than a chat, and multi-agent systems about 15×. Lesson: use several agents only when the value of each task pays for it. How to work out the cost per task is in how to build an AI agent.
Regulated work and our own
8. CarbonormAI: a compliance agent that does not decide the grade
Carbonorm helps manufacturers comply with the EU Packaging and Packaging Waste Regulation (PPWR). Its agent, CarbonormAI, grades packaging portfolios, checks compliance documents against the regulation and answers regulation questions. The design choice that matters: the grade comes from a deterministic engine that encodes the regulation's articles as rules (recyclability, recycled content, restricted substances, format bans), never from the language model. The same packaging always gets the same grade. The agent's job is to call the engine correctly, explain the result and cite the article behind every finding.
Around that core: quotes the model takes from the regulation are checked word for word against the source text, and output below the match threshold is rejected (this check was added after the model invented regulation text during development). Every answer labels what is evidence and what is interpretation. Brand names and SKU codes are replaced with pseudonyms before a request reaches the model provider. The tracked EU acts are checked on EUR-Lex every day, and if a document cannot be fetched, the agent says so instead of answering from stale text. Frequent tasks run as commands that skip the model entirely, so they answer instantly and cost no tokens.
- ScanDaily check
Scheduled check of the tracked EU acts on EUR-Lex, 5 August 2026.
- FoundNew text
A new consolidated version of the EU food-contact plastics regulation.
- ProcessSummaries
Text downloaded and split; the model drafts summaries with quotes.
- Verify10 of 10 quotes
Every quote matched word for word in the source; none rejected.
- Done3 min 18 s
Knowledge graph rebuilt. No person involved.
The team is as clear about the limits: the agent is an internal edition used by Carbonorm's own staff, its grades are a preliminary assessment, not legal advice, and it covers a defined set of EU acts, not all EU law. Lesson: when a wrong answer has a legal cost, keep the decision in code and give the model the explaining. (Disclosure: Carbonorm is a Semroi client; see the case study.)
9. Semroi: the agent that edits this website
Since 31 October 2026, content changes on semroi.net go through an agent: Claude Code, connected to our Payload CMS through Payload's MCP server. It reads and writes pages, posts and case studies in English, Turkish and Romanian. It works as a CMS user, so every write passes the same validation, drafts and version history as an edit by a person. This article was written by it, as a draft, and published by a person. Here is that run:
- ReadContext
Fetched the published posts in this category and the block format of one of them.
- ResearchSources
Read the pages ranking for "AI agent examples" and every source cited here.
- WriteDraft
Created the post as a draft: not public, the address returned 404.
- ReviewA person
An editor compared it with the pages that rank for the topic and sent it back for more depth. This is the second draft.
- PublishCheck
After approval: publish, clear the page cache, request every language version and check status and head tags.
The rules it works under: it writes drafts, a person publishes, and it reads every document back after saving it. We have not measured time saved yet. Lesson: start with the narrowest permissions and widen them as the agent proves itself, the way Ramp widened its action list.
10. Air Canada: what happens without a line
In 2022 a passenger asked Air Canada's website chatbot about bereavement fares and was told he could claim the discount after travelling. The airline's policy said otherwise and it refused the refund, arguing the chatbot was "a separate legal entity". On 14 February 2024 the British Columbia Civil Resolution Tribunal rejected that and awarded CA$812.02: "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."
This was a chatbot, not an agent, but the rule applies even more when software acts. Lesson: ground answers in your own policy documents (see RAG) and send anything that commits money to a rule or a person.
What the working examples have in common
- One trigger, one end state: a chat, a request, a report. "Done" can be checked.
- A short tool list: read access first, write access only where it is needed.
- A line on autonomy: what the agent does alone, what it proposes, what it never does (Ramp, Shopify, IBM).
- The decision that matters stays in code or with a person: Carbonorm's grade, Shopify's confirmation, our publish button.
- An evaluation set before launch: real cases with expected outcomes, rerun on every change (Uber, Ramp).
- A visible handoff: customers and staff can reach a person without fighting the bot (Klarna, Salesforce).
- One business number: escalations, containment, time to resolution, error rate.
AI agents for business: 6 patterns a smaller company can run
Most deployments above come from large companies with their own engineering teams. The same shape works at a smaller scale when the job is narrow. These are patterns we scope with clients, not published results; the last column is what to measure.
- Team: HotelTrigger: Guest message before arrivalTools: Reservation system, extras catalogue, policiesAgent does alone: Answers, room notes, offers published extrasNeeds a person: Refunds, rate changes, complaintsMeasure: Reply time, extras revenue per stay
- Team: Online storeTrigger: "Where is my order?" or a return requestTools: Order system, carrier tracking, returns policyAgent does alone: Order status, return label within policyNeeds a person: Exceptions, refunds above a limitMeasure: Tickets closed without a person, repeat contacts
- Team: B2B salesTrigger: Inbound form or emailTools: CRM, company data, calendarAgent does alone: Qualifies, enriches, proposes meeting slotsNeeds a person: Pricing and discountsMeasure: Time to first reply, meetings booked
- Team: FinanceTrigger: Supplier invoice arrives by emailTools: Accounting system, purchase ordersAgent does alone: Matches to the order, flags mismatchesNeeds a person: Approving paymentMeasure: Invoices per hour, exception rate
- Team: MarketingTrigger: Weekly report is dueTools: Ads, GA4 and Search Console dataAgent does alone: Pulls the data, drafts the summaryNeeds a person: Budget changesMeasure: Hours saved, errors found in review
- Team: Internal help deskTrigger: Staff questionTools: Company documents, ticketingAgent does alone: Answers with the source, opens a ticketNeeds a person: Policy exceptionsMeasure: Answers with the right source
| Team | Trigger | Tools | Agent does alone | Needs a person | Measure |
|---|---|---|---|---|---|
| Hotel | Guest message before arrival | Reservation system, extras catalogue, policies | Answers, room notes, offers published extras | Refunds, rate changes, complaints | Reply time, extras revenue per stay |
| Online store | "Where is my order?" or a return request | Order system, carrier tracking, returns policy | Order status, return label within policy | Exceptions, refunds above a limit | Tickets closed without a person, repeat contacts |
| B2B sales | Inbound form or email | CRM, company data, calendar | Qualifies, enriches, proposes meeting slots | Pricing and discounts | Time to first reply, meetings booked |
| Finance | Supplier invoice arrives by email | Accounting system, purchase orders | Matches to the order, flags mismatches | Approving payment | Invoices per hour, exception rate |
| Marketing | Weekly report is due | Ads, GA4 and Search Console data | Pulls the data, drafts the summary | Budget changes | Hours saved, errors found in review |
| Internal help desk | Staff question | Company documents, ticketing | Answers with the source, opens a ticket | Policy exceptions | Answers with the right source |
For the hotel case in detail, see what a hotel AI agent does. Most of these agents reach your systems through connectors; MCP servers are now the standard way to build them, and the way our own CMS agent works.
If a vendor's "agent" cannot call a tool, take an action or hand over to a person, it is a chatbot with a new name. Ask to see a run log for one real request: the tool calls, in order, and where it stopped.
How to pick your first agent
- List tasks that repeat every week, have a clear end state and touch two or three systems at most.
- Cross out any task where one mistake costs more than a day of doing it by hand, unless the decision can stay in code, as at Carbonorm.
- Check whether a fixed workflow does the job. If the steps never change, you do not need an agent.
- Write 30–50 real cases with the expected outcome before you choose a tool.
- Pilot with a person approving every action, then widen what the agent may do alone.
Gartner expects over 40% of agentic AI projects to be cancelled by the end of 2027, mostly for cost and unclear value. A narrow first task with a number attached is the cheapest protection. The full steps, architecture and cost are in how to build an AI agent.
Frequently asked questions
What is an example of an AI agent?
A customer service agent that reads an order, checks the returns policy and issues a return label without a person, like Klarna's assistant. Others: IBM's AskHR, which handles vacation requests and employment letters; Ramp's agent that corrects merchant records; and Shopify Sidekick, which prepares store changes for the merchant to confirm.
Is ChatGPT an AI agent?
When it only answers, ChatGPT is an assistant. When it searches the web, runs code or uses connected apps to finish a task, it is acting as an agent. The test is whether it decides on and takes actions with tools.
What are AI agents used for in business?
In LangChain's 2026 survey of 1,300+ professionals, customer service was the most common use case (26.5%), followed by research and data analysis (24.4%). In companies with 10,000+ employees, internal productivity came first (26.8%).
Can an AI agent be used where mistakes have legal consequences?
Yes, if the model does not make the decision. CarbonormAI grades packaging with a rule engine coded from the regulation and uses the model only to call that engine, explain the result and cite the article. Its output is still a preliminary assessment, not legal advice.
How much does an AI agent cost to run?
Work it out per task: model calls per task times the tokens and price of each call. Anthropic measured agents at about 4× the tokens of a chat and multi-agent systems at about 15×. Commands that skip the model, as at Carbonorm, cost nothing per run. Our build guide has a worked example.
Do small businesses need AI agents?
Only for a task that repeats often, has a clear end state and uses data in systems the agent can reach. If the steps never change, a workflow tool is cheaper. If the task is rare, keep it manual.
Is a company liable for what its AI agent says?
In Moffatt v. Air Canada (2024) a Canadian tribunal held the airline liable for its chatbot's wrong answer and rejected the argument that the bot was a separate entity. Treat agent output as your own statement.
Sources
- Klarna: AI assistant handles two-thirds of customer service chats in its first month (27 Feb 2024)
- Klarna changes its AI tune and again recruits humans for customer service (CX Dive via Yahoo Finance, 2025)
- CNBC: Salesforce will hire 2,000 people to sell AI products, Benioff says (17 Dec 2024)
- Yahoo Finance: AI-driven job displacement may already be here (Salesforce support headcount, 2025)
- IBM case study: AskHR
- IBM's CHRO weighs in on lessons learned from an early AI rollout (Yahoo Finance)
- Ramp Engineering: How Ramp fixes merchant matches with AI
- Uber: Unlocking financial insights with Finch
- Shopify: Agentic commerce for every developer, the Spring '26 Edition
- Anthropic: How we built our multi-agent research system
- Carbonorm: PPWR and CBAM compliance
- BBC: Airline held liable for its chatbot giving passenger bad advice (23 Feb 2024)
- American Bar Association: BC tribunal confirms companies remain liable for information provided by AI chatbot
- LangChain: State of Agent Engineering 2026
- Gartner: over 40% of agentic AI projects will be canceled by end of 2027 (25 Jun 2025)
- AI agents
- AI agent examples
