Guides

Agentic AI Examples: 22 Real Deployments (and What They Actually Do)

Agentic AI plans a step, calls a tool, and acts. Twenty-two named deployments are below, from Claude Code, Codex, and Devin to Klarna, checkout, computer use, and a security investigation agent. Each row has a date and a source. Vendor numbers are labeled.

Anthropic logoAnthropicClaude logoClaudeOpenAI logoOpenAIGitHub logoGitHubGitHub Copilot logoGitHub CopilotCursor logoCursorDevin logoDevinGoogle logoGoogleMicrosoft logoMicrosoftAmazon logoAmazonElevenLabs logoElevenLabsSEC logoSEC
Editorial policy

Key takeaways

  • Agentic AI plans a step, calls a tool, acts on the result, and loops. A chatbot stops at the reply. A fixed automation follows a script a person wrote in advance.
  • Twenty-two named deployments are in the table, checked November 16, 2026. Coding agents, Klarna’s assistant, four Salesforce Agentforce customers, three checkout systems, computer use, Deep Research, a security investigator, and one healthcare voice agent.
  • Vendor case studies are labeled vendor-reported. Klarna’s Form 20-F prints two different full-time-equivalent estimates. Do not add them.
  • The public failure record is specific. OpenAI logoOpenAI’s help page says ChatGPT logoChatGPT agent and Operator are gone. OSWorld 2.0’s best 500-step binary score, read November 9, is 20.6 percent.

Agentic AI plans a step, calls a tool, and acts on the result until a stop rule fires. The strongest named examples on November 16, 2026 are Claude Code, OpenAI Codex, Devin, GitHub Copilot’s cloud agent, Klarna’s assistant, and the checkout agents from Google, Microsoft, and Amazon.

This page is a dated list of 22 deployments, not a tour of product categories. Each row has a primary source. A number that comes from a vendor case study is marked vendor-reported. A page that prints no success rate stays blank. The year those products do, or do not, have a launch day is AI agents in the next 12 months. The workshop around the model is an AI agent harness.

The 22 deployments

Download CSV

Example Company Category What the agent does autonomously Human-in-the-loop point Reported result (sourced) Date verified or announced
Claude Code Anthropic Coding Proposes a step, then a harness runs an allowed tool, such as a file edit or a command, and returns the result The loop stops when the model makes no further tool call, an error hits, or a person must approve No task-success rate on the loop doc Agent-loop docs, read November 8, 2026
Codex OpenAI Coding Coding agent across ChatGPT, the CLI, an IDE extension, and the web. Cloud tasks can keep running after the laptop closes. The product page describes pull-request review A person reviews the pull request. Desktop Computer Use is a separate surface No success rate on the product page recorded November 6, 2026 Product page recorded November 6, 2026
Devin Cognition Coding Writes, runs, and tests code. A session has a shell, an IDE, and a browser You can watch edits and take over the editor without leaving the session Cognition’s homepage calls it the first autonomous software engineer. That is the company’s line, not a measured rate Docs read November 9, 2026
Copilot cloud agent GitHub Coding and security fixes Researches a repo, plans, edits files, and runs tests and linters in a cloud sandbox or a GitHub Actions runner. Can be assigned code-scanning alerts You review the diff before a pull request, unless the prompt asked for one. GitHub says the output can include vulnerabilities No cost-per-task on the page. Actions-based sessions stop at 59 minutes Docs read November 16, 2026. Paid cloud experience from October 28, 2026
Cursor Cloud Agents Anysphere Coding Clones the repo, works on a branch, builds and tests, can drive a desktop and browser, and can call MCP servers Pushes the branch for handoff. You can take the remote desktop, then hand it back No success rate. Free and Plus do not include Cloud Agents Docs read November 16, 2026. Formerly Background Agents
Klarna AI assistant Klarna Customer service and finance Launched February 2024 with OpenAI. The filing says it handles customer service, refunds, and returns Customers can still reach a person. The risk factor says humans are involved in training and monitoring Company-reported: 80% of customer-service chats in 2025, from Klarna’s chat logs. See the two FTE estimates below Form 20-F for the year ended December 31, 2025, read November 16, 2026
Klarna US voice agent Klarna, built with ElevenLabs Voice First line of US phone support. Answers, gathers context, and escalates Hands over when the query is complex or sensitive. ElevenLabs says customers keep a path to a person Vendor-reported, ElevenLabs: resolutions up to 10 times faster for queries the agent handles. Not a line in the 20-F passage above ElevenLabs post, February 11, 2026
Rivet, Crocs Crocs on Salesforce Agentforce Shopping and service An orchestrator coordinates shopper and service agents for discovery, FAQs, order tracking, returns, and conversational checkout The story describes governance and a phased launch. It does not print an approval on every reply Vendor-reported: 3 agents in 5 months, voice in 7 weeks. No resolution rate on the page Salesforce story, October 27, 2026
Nestlé consumer agent Nestlé on Agentforce Customer service Answers common consumer questions on ingredients, recipes, products, and careers The story says trusted handoffs when a person is needed Vendor-reported: 40% autonomous resolution in the pilot, 4.8/5 CSAT on agent-led interactions Salesforce story, October 23, 2026
Southwest help assistant Southwest Airlines on Agentforce Customer service Help Center and app assistant. Answers FAQs and offers next steps on flight changes Brings a live rep in when more help is needed Vendor-reported: 45% case resolution across more than 2 million interactions, and $6 million in projected annual savings Salesforce story, October 13, 2026
Sam Trustpilot on Agentforce Customer service Authenticates the user, answers common questions, and routes the rest Complex cases go to specialists Vendor-reported: 28% of support cases deflected, and a 99% smaller backlog within six months Salesforce story, August 12, 2026
L’Oréal advisor assist L’Oréal on Agentforce Customer service, copilot Summarizes a case, drafts an email, checks intent, and generates a promo code inside Service Cloud An advisor still handles the case. The 64% figure applies when the draft needs no edits Vendor-reported: handling time from about 7 minutes to about 2.5 minutes when the draft is unchanged. More than 300 advisors Salesforce story, October 20, 2026
Google direct checkout, Canada Google Shopping In AI Mode and the Gemini app, checkout can complete for named Canadian retailers The shopper is in a Google surface. The retailer list is the boundary Named at launch: lululemon, select Shopify merchants, Shoppers Drug Mart, and Simons. No conversion rate in the post Google Canada post, November 4, 2026
Copilot Checkout Microsoft Shopping Completes an eligible purchase inside Copilot, on the merchant’s systems English-language merchants, US buyers, US dollars, as re-read November 14, 2026 No fee or conversion rate restated here. The shopping report holds the longer record Microsoft Agentic Commerce page, re-read November 14, 2026
Alexa for Shopping auto-buy Amazon Shopping Can complete a purchase, including from a text, when the customer’s price is reached, using the default payment method The price condition is the gate. The November 6 post does not print a cap Amazon says more than 400 million customers used Alexa for Shopping in the prior 12 months, and that agentic-feature users grew more than eight times since the start of the year. Those are Amazon’s phrases Amazon post, fetched November 12, 2026. Texting called widely available in the November 13 update recorded on the shopping report
ChatGPT Wallet OpenAI Shopping and payments A dot pays with a network token after you approve Face ID or Touch ID on that purchase. Ordinary ChatGPT shopping, re-read November 14, still finishes on the merchant’s site No volume figure. OpenAI’s merchant page, fetched November 12, says it is moving away from standalone Instant Checkout Wallet announced November 13, 2026
ChatGPT desktop Computer Use OpenAI Computer use With ChatGPT Work and Codex, looks at a screen and requests clicks and keys on macOS and Windows Linux is not named. The person is at the machine No success rate on the guide Guide read November 9, 2026
Claude computer use in Cowork Anthropic Computer use Beta desktop control on macOS and Windows for Pro, Max, and Ultra. The API toolset is a separate surface The help page says Team and Enterprise plans do not have computer use. You can stop the run No success rate on the help page. The API toolset is not available in Claude Managed Agents Help page read November 9, 2026
Chrome auto browse Google Browser Browses and completes steps in Chrome for a signed-in preview user The January 28, 2026 post introduced a US preview for AI Pro and Ultra No task-success rate in the announcement. This site’s November 9 report was still watching whether the country list had moved Announced January 28, 2026
Deep research OpenAI Research Builds a plan, searches the public web and uploaded files, and reads connected apps. Returns a cited report You can edit the plan before it starts and interrupt the run. OpenAI says it uses read actions only, not app write actions No quality score on the help page. The o3-deep-research model shut down July 23, 2026. The product now runs on later models Help page read November 16, 2026
Attack Investigation Agent Microsoft Security Starts from an alert or incident, correlates Defender signals, and writes an investigation report with a verdict, a timeline, and recommended actions Microsoft says to review the findings before remediation or before you send them on No precision figure on the page read Microsoft Learn, read November 16, 2026
AI Front Door Hippocratic AI Healthcare voice A voice agent for inbound patient access. The release says it resolves needs in one conversation instead of deflecting them, and escalates Clinical safety is the company’s claim to review, not a number this page can grade The release names Wellspan and Cincinnati Children’s as initial deployment partners. It does not print a resolution rate Press release, April 16, 2026

Source: aiagentsinfo.com, Agentic AI Examples: 22 Real Deployments (and What They Actually Do), updated . How to cite: Agentic AI Examples: 22 Real Deployments (and What They Actually Do). AI Agents Info (aiagentsinfo.com). https://www.aiagentsinfo.com/blog/agentic-ai-examples. Updated November 16, 2026.

Directory pages for the coding rows: Claude Code, Codex, Devin, Cursor. Prices and plan limits stay on Best AI coding agents. Screen control stays on Best computer use agents.

What Klarna’s filing actually prints

Use the Form 20-F, not the February 2024 launch post, if you quote Klarna in 2026. The filing says the assistant launched in February 2024 with OpenAI, and that it manages customer service, refunds, and returns. For the year ended December 31, 2025 it says the assistant handled 80 percent of customer-service chats, according to Klarna’s service chat logs, with no drop in satisfaction on internal surveys.

The same document prints two equivalent-headcount estimates, and they are not the same sentence.

  • One passage says the assistant has had 31 million conversations since launch, resolved queries in two minutes versus 12 minutes for human agents as of September 2024, did the equivalent work of over 700 full-time agents based on the 2024 reduction in conversations those agents handled, and delivered $39 million in cost savings in 2024.
  • A later passage says the 2025 work was the equivalent of over 850 full-time agents, estimated from the average monthly reduction in chat and telephone conversations handled by full-time agents in 2025, and about $59 million in cost savings in 2025.

Both figures are Klarna’s estimates. The methods are not the same year. Do not add 700 and 850, and do not treat $59 million as audited savings. The risk factor still calls the system a customer-service chatbot, and it warns that generative output can be wrong. The same filing says Klarna does not use generative AI for credit underwriting. The assistant is a support and shopping workflow. It is not the credit model.

The US voice line is a second product. ElevenLabs’ February 11, 2026 post says Klarna launched it as the first line of phone support for US customers, that it escalates, and that resolutions are up to 10 times faster for queries the agent handles. “Up to 10 times” is ElevenLabs’ phrase. It is not in the 20-F passages above.

What makes it agentic

A system is agentic on this page when it does five things.

  1. Plan. The model chooses a next step from a goal, not from a script that already named every step.
  2. Tool calling. It requests a tool: a shell, a file edit, a browser action, a refund, a search, a cart, or a security API.
  3. Act. The harness runs that tool on something outside the chat window.
  4. Loop. The result comes back, and the model can take another step.
  5. Guardrails. A permission check, a spend cap, a plan you can edit, or a person who must approve before the action counts.

A chatbot fails 3 and 4. It replies. A fixed automation can act, and it can even loop, but a person wrote the path. The model did not pick the tool. That distinction is the long version of how AI agents work. The protocol many of these tools now speak is MCP. The short definition is the Model Context Protocol glossary entry.

One sentence on why each row qualifies:

  • Claude Code qualifies because the harness runs the tool the model just chose, then feeds the result back.
  • Codex qualifies because a cloud task keeps editing and reviewing after you close the laptop, and the handoff is a pull request rather than a paragraph.
  • Devin qualifies because a session can use a shell, an editor, and a browser, and you can take the editor back mid-run.
  • Copilot’s cloud agent qualifies because it plans, edits, and runs tests in a sandbox, then waits on your review unless you asked for the pull request up front.
  • Cursor Cloud Agents qualify because the VM builds and tests the change, can drive a desktop, and can call MCP servers before the branch is pushed.
  • Klarna’s assistant qualifies because the filing says it completes service tasks, including refunds and returns, rather than only drafting an answer.
  • Klarna’s US voice agent qualifies because it holds the phone call, gathers context, and escalates, which is a loop with a handoff.
  • Crocs’ Rivet qualifies because an orchestrator routes work across shopper and service agents, including returns and conversational checkout. Orchestration is the multi-agent pattern, and the glossary term is orchestration.
  • Nestlé’s agent qualifies because the case study describes autonomous resolution of a consumer question, with a handoff when a person is required.
  • Southwest’s assistant qualifies because it takes the first pass on a traveler’s request and pulls in a rep only when the request needs one.
  • Trustpilot’s Sam qualifies because it authenticates, answers, and routes. Authentication is a tool call, not a paragraph.
  • L’Oréal’s deployment is the edge case. It calls tools that draft and that generate a promo code, but an advisor still owns the send. Treat it as agentic assistance, not an unattended agent.
  • Google’s Canada checkout qualifies because the shopping flow can reach payment for a named retailer, which is an action, not a link list.
  • Copilot Checkout qualifies because Microsoft’s page says the purchase can finish inside Copilot, on the merchant’s systems.
  • Alexa auto-buy qualifies because Amazon says it can place the order when a price is reached, without a fresh click on that order.
  • ChatGPT Wallet qualifies as a guarded agent. The dot can pay, and the payment does not happen until you approve that purchase.
  • ChatGPT desktop Computer Use qualifies because the model asks for a click or a key and the app runs it. Computer use is that loop.
  • Claude computer use qualifies for the same reason, inside the plans the help page names, and not on Team or Enterprise as of November 9, 2026.
  • Chrome auto browse qualifies because the browser takes the steps. It is browser use, the narrower case of computer use.
  • Deep research qualifies because it writes a plan, searches, reads tools, and revises until the report is done. It does not qualify as a system that writes back to your apps.
  • Attack Investigation Agent qualifies because it runs a multi-step investigation from one alert and produces a verdict you are told to review before you act on it.
  • AI Front Door qualifies, on the company’s release, because the voice agent carries an inbound clinical-access call and is deployed with named health systems. The release is not a peer-reviewed outcome study.

What this means for you

Builders. Pick the harness before you pick the demo. Claude Code, Codex, Devin, Copilot, and Cursor Cloud Agents are five different places the loop can run, and they do not share a stop rule. Copilot’s Actions-based session ends at 59 minutes. Cursor’s cloud agent can call MCP and can also run hooks you put in the repo. If you cannot name the tool list, the approval gate, and the eval that fails a bad run, you have a chatbot with extra buttons. The comparison set is Best AI agent harnesses. Cost per task is not the seat price. Token price, retries, and tool JSON sit outside the monthly sticker. The 12-month calendar dates the token steps. It does not convert them into a successful task.

Buyers. Ask which row you are buying. A support agent that deflects 28 percent of cases, vendor-reported, is not a coding agent that opens a pull request, and it is not a checkout agent that can spend. Ask for the human stop in one sentence: who approves a refund, a payment, a pull request, or a security remediation. Then ask for the source of the percentage. Klarna’s 80 percent is a filing plus chat logs. Southwest’s 45 percent is a Salesforce case study. Those are different kinds of evidence. If the seller cannot point at a page and a date, treat the number as a slide.

Merchants. The shopping rows do not behave the same way. Google named Canadian retailers on November 4, 2026. Microsoft’s page, re-read November 14, limits Copilot Checkout to a US dollar sale. Amazon can auto-buy at a price. ChatGPT’s ordinary shopping flow still sends the buyer to the store, and Wallet is a separate approval. The month-by-month record is AI shopping agents, the next 6 months. The directory of products is the AI agent directory.

Where these systems fail, on the record

The interesting deployments are also the ones with a public limit.

OpenAI’s help page, read November 9, 2026, says ChatGPT agent is no longer available and points people to ChatGPT Work and the cloud browser. The same page says Operator was folded into agent mode and that the Operator website is no longer accessible. A 2025 list that still leads with Operator is describing a product OpenAI has taken down.

OSWorld 2.0, read November 9, 2026, prints a 20.6 percent binary completion for Claude Opus 4.8 at 500 steps on 108 long tasks, and a 54.8 percent partial score. Above 163 minutes, binary completion on that plot falls to zero for every model the page keeps. Computer use is a real product. It is not a solved one. The product calendar is browser and computer-use agents.

On November 11, 2026, United Airlines said it is barring unauthorized third-party personal AI agents from booking, ticketing, and distribution. A shopping agent that can click a store is not a shopping agent that can buy a seat.

GitHub’s own docs say Copilot can produce incorrect code, including code with security vulnerabilities, and that session logs do not replace your review. Microsoft’s Attack Investigation Agent page says the report is AI-generated and that you review it before remediation. Deep research does not get write actions on connected apps. Those are the guardrails. They are also the admission that the loop is not the last word.

Klarna’s filing is the cleanest customer-service source in this set, and it still warns that generative tools can be inaccurate, biased, or the wrong fit for the job. An 80 percent chat share is not a claim that the other 20 percent, or the credit decision, should be automated.

What this page leaves out

A model card that says “computer use” is not a deployment. Tencent’s Hy4 announcement and Meta’s Muse docs describe a capability. They are not a named company running the loop on its own customers, so they are not rows. Abridge’s 2026 clinical assistant stays off the table because the clinician directs the query. ServiceNow’s assist meter is a price, not a customer. If a later primary page names the customer, the action, and the date, it can be added. Until then the count stays 22.

Production surveys, including who says they are scaling agents, are a different object from a named deployment. That cross-section is AI agents in production, 2026.

FAQ

What is an example of agentic AI?

Claude Code is one. Anthropic’s agent-loop docs, read November 8, 2026, describe a model that proposes a step, a harness that runs an allowed tool, and a loop that continues until the model stops, an error hits, or a person must approve. Klarna’s assistant is another. The Form 20-F says it handles customer-service chats, including refunds and returns, and that it handled 80 percent of those chats in 2025 according to Klarna’s own chat logs.

Is ChatGPT agentic AI?

A normal ChatGPT reply is generative. It writes text and stops. Deep research, desktop Computer Use, and ChatGPT Wallet are agentic modes: they plan, call tools, and act, and a person can edit the plan, take the screen back, or approve the payment. OpenAI’s help page, read November 9, 2026, says the older ChatGPT agent product is no longer available and that the Operator website is no longer accessible.

What is agentic AI vs generative AI, with examples?

Generative AI produces an output and stops. A chat answer is generative. Agentic AI adds a loop: the model chooses a next action, a harness runs a tool, and the result comes back. Claude Code editing a repository is agentic. A chat reply that only suggests a code change is generative. Anthropic’s November 2, 2026 prospectus draws that line, and how AI agents work walks through it.

What are examples of agentic AI in customer service?

Klarna’s assistant, Crocs’ Rivet, Nestlé’s consumer agent, Southwest’s help assistant, and Trustpilot’s Sam. Klarna’s 80 percent is from a filing and chat logs. The Salesforce percentages are vendor-reported. L’Oréal’s deployment drafts for advisors, and a person still owns the case.

What are examples of agentic AI in healthcare?

Hippocratic AI’s April 16, 2026 release says AI Front Door is being deployed with Wellspan and Cincinnati Children’s as initial partners. It describes a voice agent for inbound access and does not print a resolution rate. A clinician-directed chart assistant is a different product, and this page does not promote it to the same row.

What are examples of agentic AI in finance?

Klarna is the example with a filing. The 20-F reports the 80 percent chat share and also says Klarna does not use generative AI for credit underwriting. Google, Microsoft, Amazon, and ChatGPT Wallet are the payments examples, and each one stops at a different approval rule. The commerce term is agentic commerce.

What makes an AI system agentic instead of a chatbot?

It plans, it calls a tool, it acts, it loops, and a guardrail or a person can stop it. A chatbot stops at the text. A fixed workflow can act, but the path was written before the run. The guardrail and the human in the loop are the stop, not a footnote.

Do these agents run without a person?

No row on this page is an unattended company. Coding agents hand you a branch or an editor. Wallet waits for a biometric approval. The security agent’s page tells you to review the report before you remediate. Deep research reads. It does not write to your apps.

Sources

  1. 1Anthropic logoAnthropic, How the agent loop works (read November 8, 2026)
  2. 2OpenAI logoOpenAI Codex product page, as recorded November 6, 2026
  3. 3GitHub logoOpenAI, Codex on GitHub
  4. 4Devin logoCognition, Introducing Devin (read November 9, 2026)
  5. 5Devin logoCognition, Devin session tools (read November 9, 2026)
  6. 6GitHub logoGitHub Docs, Copilot on GitHub.com and GitHub Mobile (read November 16, 2026)
  7. 7Cursor logoCursor, Cloud Agents (read November 16, 2026)
  8. 8SEC logoKlarna Group plc, Form 20-F for the year ended December 31, 2025 (read November 16, 2026)
  9. 9ElevenLabs logoElevenLabs, Klarna voice agent (February 11, 2026)
  10. 10SSalesforce, Crocs and Agentforce (October 27, 2026)
  11. 11SSalesforce, Nestlé and Agentforce (October 23, 2026)
  12. 12SSalesforce, Southwest Airlines and Agentforce (October 13, 2026)
  13. 13SSalesforce, Trustpilot and Agentforce (August 12, 2026)
  14. 14SSalesforce, L’Oréal and Agentforce (October 20, 2026)
  15. 15Google logoGoogle, agentic shopping in Canada (November 4, 2026)
  16. 16Microsoft logoMicrosoft Advertising, Agentic Commerce (re-read November 14, 2026)
  17. 17Amazon logoAmazon, Alexa for Shopping text (fetched November 12, 2026)
  18. 18OpenAI logoOpenAI, Introducing ChatGPT Wallet (November 13, 2026)
  19. 19OpenAI logoOpenAI, Computer Use in the ChatGPT desktop app (read November 9, 2026)
  20. 20OpenAI logoOpenAI Help, ChatGPT agent (read November 9, 2026)
  21. 21Claude logoClaude Help, computer use in Cowork (read November 9, 2026)
  22. 22Google logoGoogle, Chrome auto browse (January 28, 2026)
  23. 23OpenAI logoOpenAI Help, Deep research in ChatGPT (read November 16, 2026)
  24. 24OpenAI logoOpenAI, o3-deep-research model page (shut down July 23, 2026)
  25. 25Microsoft logoMicrosoft Learn, Attack Investigation Agent (read November 16, 2026)
  26. 26HHippocratic AI, AI Front Door press release (April 16, 2026)
  27. 27OOSWorld 2.0 (read November 9, 2026)
  28. 28UUnited Airlines, November 11, 2026 statement

Guides

AgentCore runtime V2 is now the default: P75 under 2s, CPU at $0.1276

New Amazon Bedrock AgentCore runtimes use the next-generation runtime by default in 19 regions as of November 12, 2026. A fully busy 10-minute session at 1 vCPU and 2 GB is $0.0269 on V2 rates and $0.01807 on V1 rates. Set platformVersion and confirm it.

November 13, 2026 · 18 min read

Guides

How Do AI Browser Agents Complete Tasks?

A browser agent reads a page as an accessibility tree, a screenshot, or both, then acts. OSWorld 2.0's best 500-step finish rate, read November 13, 2026, is 20.6%.

November 13, 2026 · 23 min read