- Documented agent deployments at Klarna, Intercom, GitHub, and Uber, with the primary-source numbers behind each
- The four-part pattern behind every production success, and how the failures invert it
- Why GitHub designed distrust into its coding agent, and what that teaches about running agents safely
- Four properties that identify your own first agent use case
Every vendor deck has an agent demo. Very few companies will show you an agent that’s been running in production for a year, because the demo and the deployment are different animals. The demo optimizes for wonder; the deployment optimizes for a metric someone’s job depends on.
This post collects examples of the second kind: named companies, public numbers, systems that survived contact with real customers and real codebases. The pattern behind them matters more than any single example, so we’ll end there.
What counts as an agentic AI example?
The bar for “agentic” is behavior. A system qualifies when it pursues a goal through multi-step actions in real systems, rather than producing content for a human to act on. We drew that line in detail in agentic AI vs generative AI; the short test is whether the work is finished when the AI stops.
Production agent: an agentic AI system serving real users or real workloads on an ongoing basis, with its performance measured against a business metric, as opposed to a pilot, demo, or benchmark run.
Every example below is documented by the company itself, with figures they published. Company-reported numbers deserve healthy skepticism, but they beat the alternative: example lists whose “deployments” trace back to nothing.
Customer service: the most proven category
Klarna gave the category its landmark numbers. In February 2024, one month after launching its OpenAI-powered assistant, Klarna reported 2.3 million conversations handled, two-thirds of all its customer service chats, doing the work of about 700 full-time agents. Repeat inquiries dropped 25%, resolution times fell from 11 minutes to under 2, and the company estimated a $40 million profit improvement for the year. Whatever you think of the framing, the shape of the result is the point: a bounded workflow (support conversations), deep integration (order systems, refunds, policies), and a hard metric (resolution).
Intercom’s Fin shows the same category as a product. Intercom reports Fin resolving an average of 76% of support conversations across more than 12,000 customers. A resolution here means the customer’s problem is handled end to end, with escalation to humans as the designed fallback rather than the failure state.
Software development: agents that ship code
GitHub’s Copilot coding agent, announced in May 2025, made the agent workflow concrete for every developer: assign an issue to Copilot, and it boots an environment, explores the codebase, makes changes, and pushes commits to a draft pull request, then iterates on review feedback like any teammate. The guardrails are as instructive as the capability: it cannot push to default branches, the developer who requested the work cannot approve it, and its internet access is restricted to a trusted list. GitHub designed distrust into the system, which is exactly right.
DevHawk’s factory is our own entry in this category, and the reason this post exists. Seven specialist agents cover the whole development job: product management, engineering, code review, security, QA automation, DevOps, and site reliability. While one writes code, another reviews it, another runs the tests, another watches production. Client roadmaps move up to 10x faster at roughly 80% less than the old way. We wrote about why whole-cycle agents beat single-seat tools in AI for software development.
Internal operations: the quiet workhorses
Uber’s Genie is the best-documented internal example. Genie is an on-call copilot that answers thousands of engineering support queries in Slack, citing Uber’s internal documentation. In May 2025, Uber’s engineering team published the results of moving it from a standard retrieval pipeline to an agentic architecture, with LLM-powered agents refining queries and selecting sources: acceptable answers up by a relative 27%, incorrect advice down by a relative 60%. Two details make it exemplary. The scope is narrow (answer on-call questions, cite sources), and the improvement was measured, published, and attributed to a specific architecture change.
Internal deployments like this rarely make headlines, which is exactly why they’re worth studying. Nobody demos a Slack bot that answers security questions. Plenty of companies bank the hours it saves.
The pattern behind every success
Line the examples up and the common structure is hard to miss:
Narrow, bounded scope. Support conversations. Assigned issues. On-call questions. Each one is a workflow with edges; the “AI employee” pitch is nowhere in sight. The viral demos chase generality, and the production systems ruthlessly avoid it.
Deep integration with real systems. Klarna’s assistant is wired into order management and refund flows. Copilot’s agent works inside the repository, the CI pipeline, and the review process. The intelligence is commodity; the integration is the moat.
Human judgment gates where mistakes are expensive. Fin escalates. Copilot’s PRs require an independent human approval. The agent does the work; a person owns the outcome.
A metric that was being measured before the agent arrived. Resolution rate, time to merge, answer quality. If nobody can say what number the agent moved, nobody can say whether it worked.
The failures invert the same list. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls: broad scope, shallow integration, no gates, no metric.
How to spot your own first use case
Look for a workflow, not a wow. The strongest candidates share four properties: it repeats (daily or weekly, not quarterly), it’s bounded (a clear start, finish, and definition of done), it runs through digital systems an agent can integrate with, and it has a number attached that someone already tracks.
In most companies that points at unglamorous territory: support triage, code review queues, report assembly, data reconciliation, on-call questions. That’s the method working as intended. Every example above started exactly there.
If you want a second pair of eyes on where an agent would pay off first in your business, that’s what our AI audit is for: a two-week review of your team, stack, and workflows, ending in a written, vendor-neutral playbook. You never have to hire us to act on it.
Frequently asked questions
What is the best example of agentic AI in production?
Klarna’s customer service assistant remains the most cited: 2.3 million conversations in its first month, two-thirds of the company’s support chats, resolution times down from 11 minutes to under 2, and an estimated $40 million annual profit improvement, per Klarna’s February 2024 announcement. It’s the clearest public case of an agent owning a workflow end to end at scale.
Are AI agents really being used, or is it hype?
Both, which is why example quality matters. Named, documented production deployments exist at Klarna, Uber, GitHub, and across Intercom’s customer base, with published metrics. At the same time, Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027. The technology is real; the failure rate is a scoping problem.
What do failed agentic AI projects have in common?
They invert the success pattern: scope too broad to bound, integrations too shallow to act on real systems, no human gates where errors are costly, and no pre-existing metric to judge results against. Gartner’s cited reasons for project cancellations (escalating costs, unclear business value, inadequate risk controls) map to the same three gaps.
How do companies keep production agents safe?
By designing distrust into the system. GitHub’s coding agent can’t push to a default branch and can’t have its work approved by the person who requested it. Fin escalates to humans by design. The general rules: scope permissions to the workflow, log every action, gate irreversible steps behind human review, and measure quality continuously.
What’s the difference between an AI agent example and an AI tool example?
A tool assists a human doing the work: an engineer accepts an autocomplete suggestion. An agent does the work and submits it for judgment: a draft pull request appears, tests included. Every example in this post is an agent by that test, and it’s the test worth applying to any vendor claiming their product is “agentic.”
- Klarna. "Klarna AI assistant handles two-thirds of customer service chats in its first month." February 27, 2024. https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/
- Intercom. "Fin. The #1 AI agent for customer service." Accessed August 2026. https://fin.ai/
- GitHub. "GitHub Copilot: Meet the new coding agent." May 19, 2025. https://github.blog/news-insights/product-news/github-copilot-meet-the-new-coding-agent/
- Uber Engineering. "Enhanced Agentic-RAG: What If Chatbots Could Deliver Near-Human Precision?" May 29, 2025. https://www.uber.com/us/en/blog/enhanced-agentic-rag/
- Gartner. "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027." June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027