Most lists of agentic AI examples describe what agents could do someday, interesting but really useful today. This one sticks to what companies have actually deployed and what they reported. We looked for agentic systems with a named company, a specific task and a published result, then checked each claim against its source. Six cleared that bar: Klarna, Salesforce, Amazon, Microsoft, Walmart and Booking.com.
The results are uneven, and that is the useful part. One deployment was partly reversed a year after launch. One was tested in a randomized controlled trial. One has no published results yet. Read together, these agentic AI examples show where autonomous AI is earning its keep in business today, and where the claims run ahead of the evidence.
Table of Contents
What Counts as Agentic AI
MIT Sloan defines agentic AI as autonomous software that perceives, reasons and acts in digital environments to pursue goals on someone’s behalf (MIT Sloan). The difference from a chatbot is action. A generative AI assistant answers a prompt. An agent plans several steps, calls other software and tools, and completes a task with limited supervision.
That distinction matters because the label is applied loosely. Gartner has warned about “agent washing,” where vendors rebrand chatbots, robotic process automation and AI assistants as agents without adding real autonomy. By Gartner’s estimate, only about 130 of the thousands of vendors marketing agentic AI offer the genuine article (New Electronics).
For this article, a deployment had to do more than generate text. It had to take or trigger an action inside a business process: resolving a support case, changing code, sorting a security queue, or preparing a reply that a person then sends.
If you want the technical background first, our explainer on how generative AI and large language models work covers the models underneath. The Model, The Agent, The Harness explains where a model ends and an agent begins.
Agentic AI Examples at a Glance
| Company | What the agent does | Reported result | Caveat |
|---|---|---|---|
| Klarna | Handles customer service chats | Two-thirds of chats in its first month | Company later hired human agents again to restore quality |
| Salesforce | Handles customer support conversations | About half of customer conversations | Support staff cut from about 9,000 to about 5,000 |
| Amazon | Upgrades Java applications | 4,500 developer-years of work saved | Figures come from Amazon’s CEO |
| Microsoft | Triages user-reported phishing emails | Up to 6.5 times the productivity in a trial | Trial run by Microsoft’s own researchers |
| Walmart | Unifies tools for customers, sellers, staff and developers | No results published yet | Most agents were still rolling out at announcement |
| Booking.com | Suggests replies to guest messages for partners | 70% boost in user satisfaction | Covers only a slice of daily messages |
1. Klarna: Customer Service at Scale, Then a Partial Reversal
Klarna, the Swedish payments company, gave the market its best-known early agentic AI example. In February 2024 it said its AI assistant, built with OpenAI, had held 2.3 million conversations in its first month. That was two-thirds of Klarna’s customer service chats, or the workload of about 700 full-time agents (The Pragmatic Engineer).
Klarna also reported that customer satisfaction matched human agents, repeat inquiries fell 25%, and average resolution time dropped from 11 minutes to under two. The assistant ran in 23 markets and more than 35 languages, and Klarna projected a $40 million profit improvement for 2024.
Critics pointed out that most of this work was first-line support: routine questions that older automation had handled at other companies for years. Complex cases still went to people. Automating that tier is valuable, but it is not the same as replacing a support operation.
The bigger lesson came in May 2025. Klarna began hiring human customer service staff again, and CEO Sebastian Siemiatkowski acknowledged that putting cost ahead of everything else had hurt service quality (Maginative). The company described a flexible, remote model for the new agents, which Siemiatkowski compared to working for Uber.
The takeaway: the volume numbers were real, but faster resolution is not the same as better service. Measure both before cutting staff.
2. Salesforce: Agents Handling Half of Support Conversations
Salesforce sells its Agentforce platform to other companies, and it also runs its own customer support on agents. In a podcast interview reported in early September 2025, CEO Marc Benioff said Salesforce had reduced its support staff from about 9,000 people to about 5,000, because AI agents now handle about half of its customer conversations (The Register).
Benioff also said agents were now following up on sales leads that the company had previously lacked the staff to call back. He described the change as rebalancing people from support toward sales.
This is one of the clearest agentic AI examples tied to a workforce decision. It is also a vendor describing its own product, so the figures are company claims. Salesforce has not published an independent assessment of how well the agents resolve cases.
The takeaway: when agents absorb a large share of routine work, the realistic outcome is often redeployment as much as reduction. Plan where people will move, not only how many you need.
3. Amazon: An Agent That Upgrades Java Applications
Software maintenance is a strong fit for agents because the work is tedious, rule-bound and easy to test. In August 2024, Amazon CEO Andy Jassy said the company had used Amazon Q, its AI coding assistant, to upgrade its Java applications to newer versions (Simon Willison).
According to Jassy, upgrading a single application had typically taken about 50 developer-days, and the agent cut that to a few hours. He estimated the total savings at 4,500 developer-years of work. Developers shipped 79% of the agent’s auto-generated code reviews without changes, and Amazon upgraded more than half of its production Java systems within six months.
Two details make this example credible. The task had a clear finish line: upgraded code either builds and passes its tests or it does not. And developers reviewed the agent’s changes before shipping them, so the 79% figure doubles as a rough measure of the agent’s accuracy.
The takeaway: the strongest early agentic use cases produce output you can verify, with a human review step already built into the workflow.
4. Microsoft: A Phishing Triage Agent Tested in a Randomized Trial
Security teams receive large volumes of emails that employees report as suspicious, and most turn out to be harmless. Microsoft built a Phishing Triage Agent into Security Copilot to sort them (Microsoft Learn).
What makes this example unusual is the evidence behind it. In November 2025, Microsoft researchers published a randomized controlled trial with 167 professional security analysts (arXiv). Analysts triaged emails drawn from 93 real user reports, 11 of them malicious. One group worked without the agent, one saw the agent’s verdicts, and one worked from a queue the agent had prioritized without being told.
Analysts working with the agent found up to 6.5 times as many true threats per minute, or 3.1 times under less favorable assumptions. Accuracy, measured by F1 score, improved by up to 77%. Most of the gain, 83%, came from the agent putting likely threats at the front of the queue rather than from its explanations.
Analysts who could see the agent’s verdicts also spent 53% more time on the malicious emails and checked its calls rather than accepting them. The caveat is that Microsoft studied its own product. Even so, a randomized trial is far stronger evidence than the before-and-after figures most companies publish.
The takeaway: an agent does not have to replace a decision to add value. Sorting work so people see the riskiest items first can deliver most of the benefit.
5. Walmart: Super Agents for Customers, Sellers, Staff and Developers
In July 2025, Walmart announced four AI “super agents,” each meant to consolidate dozens of separate tools into one interface (TechInformed):
- Sparky, for customers, already live in the Walmart app with product suggestions and review summaries, with reordering and event planning to follow
- Marty, for suppliers and sellers, covering onboarding, order management and ad campaigns
- An associate agent, giving employees HR tasks and real-time sales data in one place
- A developer agent, a platform for building and testing new AI tools
TechInformed’s report on the announcement included no performance results. It belongs on this list because it shows a different strategy. Rather than deploying one agent per task, Walmart is building a small number of agents organized around who uses them. MIT Sloan also cites Walmart’s use of agents for personalized shopping and merchandise planning.
The takeaway: for large organizations, the hard part may be consolidation, not capability. A few agents with clear owners are easier to govern than dozens of disconnected pilots.
6. Booking.com: Suggested Replies for Partner Messages
Booking.com built an AI system that suggests replies to guest messages for the hotels and property owners on its platform. B2BNN’s case study of the Booking.com AI agent found a deliberately conservative design.
The agent first looks for an existing reply template, drawing on only its top eight matches. It writes a custom reply only when it has enough information, and it opts out when it is not confident. Some categories, including refund requests, are off limits. The case study reports a 70% boost in user satisfaction, but the system helps with only tens of thousands of the roughly 250,000 messages partners receive each day.
That limited reach is a design choice as much as a weakness. Booking.com kept the scope narrow enough to evaluate continuously, using human annotation, automated scoring and production monitoring to catch problems.
The takeaway: restraint scales better than ambition. An agent that knows when to step aside is safer to expand later.
What These Agentic AI Examples Have in Common
Read side by side, the six deployments share a few patterns. This is our analysis, not a claim any of the companies made.
- Bounded tasks. Each agent works on a specific, repeated job: a support chat, a code upgrade, a suspicious email, a guest message. None of them runs open-ended.
- A measurable finish line. The strongest results come where success is easy to check, such as code that passes its tests or an email correctly flagged as malicious.
- A human checkpoint. Amazon’s developers reviewed code, Microsoft’s analysts verified verdicts, and Booking.com’s partners decide whether to send a suggested reply. Klarna, which leaned hardest on full automation, is the one that pulled back.
- Company-reported numbers. Five of the six results come from the companies themselves. Only Microsoft’s was tested in a controlled study, and that study was run by Microsoft.
The last point deserves weight. Most published agentic AI results are case studies, earnings-call remarks or press releases. They are useful signals, but they are not independent evidence, and they rarely report what failed.
Why Many Agentic AI Projects Fail
Not every deployment ends like Amazon’s. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 (New Electronics). It cites three causes: escalating costs, unclear business value and inadequate risk controls. Gartner also notes that most projects so far are early experiments driven more by hype than by a clear need.
The examples above show what separates the survivors. Klarna’s reversal was about quality, not cost. Booking.com limited its risk by refusing to handle refunds. Amazon’s project worked because its value could be counted in developer-days. If a proposed agent cannot point to a cost, time or quality measure before launch, it is a likely candidate for Gartner’s 40%.
For more on the operational side, our coverage of July’s agent releases looks at containment and cost. Our analysis of where LLMs fail covers the reliability limits agents inherit from the models underneath them.
How to Evaluate an Agentic AI Use Case
Based on the deployments above, these six questions separate promising agentic AI use cases from expensive pilots:
- Is the task repeated and well defined? Agents perform best on work that happens thousands of times in a similar form, like routine support questions or code upgrades.
- Can you verify the output? Prefer tasks where success is checkable: tests pass, a ticket closes, a threat is confirmed.
- What happens when the agent is wrong? Map the cost of an error before launch. Booking.com excluded refunds for a reason.
- Where does a person review the work? Design the human checkpoint before launch, not after the first incident.
- What will you measure besides speed? Klarna’s resolution times improved while service quality slipped. Track satisfaction, repeat contacts and escalations as well.
- Is it really an agent? Ask vendors which actions the system takes on its own. If the answer is that it drafts text for someone else to act on, you are buying an assistant.
None of these questions requires a technical team to answer. They are business questions, and the companies above that answered them well before launch are the ones with results worth publishing.
What Analysts Expect Next
Gartner predicts that by 2029, agentic AI will resolve 80% of common customer service issues without human help, cutting operational costs by 30% (Gartner). It also expects that by 2028, 15% of day-to-day work decisions will be made autonomously by agents, up from none in 2024, and that a third of enterprise software applications will include agentic AI, up from less than 1% in 2024 (New Electronics).
Those forecasts sit alongside the 40% cancellation estimate, and both can be true. Based on the deployments here, the likeliest growth is in narrow, high-volume work like support and security triage, while broad “digital worker” projects struggle to show returns. Klarna’s experience suggests even the customer service forecast will depend on companies measuring quality, not just containment rates.
Agentic AI Examples FAQ
What is a simple example of agentic AI? A phishing triage agent is a clear one. It reads each email employees report, judges whether it is malicious, and reorders the security team’s queue so the riskiest messages are reviewed first. Microsoft’s Phishing Triage Agent does this inside Security Copilot.
What is the difference between agentic AI and generative AI? Generative AI produces content in response to a prompt. Agentic AI uses that ability to plan and take actions across other software, such as updating a record, changing code or routing a case, with limited supervision.
Where is agentic AI being used most? Among the best-documented deployments, customer service, software development and cybersecurity stand out. Retail and travel companies are building agents around customer and partner interactions, but fewer have published results.
Are chatbots agentic AI? Usually not. A chatbot that only answers questions is an assistant. It becomes agentic when it can act on its own, for example by processing a request or escalating a case. Gartner calls the practice of relabeling chatbots as agents “agent washing.”
Does agentic AI replace employees? In some cases it has reduced headcount, as at Salesforce. But Klarna’s experience shows that replacing people too aggressively can hurt quality, and most of the successful deployments here keep a person reviewing the agent’s work.
How should a business start with agentic AI? Pick one repeated, well-defined task with a measurable outcome and a clear human review step. Run it narrowly, measure quality as well as speed, and expand only once the results hold.
