AI Agents Examples: 12 Running in Production

Most lists of AI agent examples are lists of demos. The agent books a flight, orders a pizza, plans a holiday, and the video ends before anyone asks what happened on the fortieth run.

These are the ones I actually see in production, grouped by whether they survive contact with a real business. I have included what they cost, where each one breaks, and four categories that keep failing so you can stop pitching them internally.

First, the distinction that makes the rest of this readable. A chatbot answers. An agent acts. An agent decides which tool to call, calls it, reads the result, and decides what to do next, in a loop, until the task is done or it gives up. That loop is where the value is and where all the difficulty is, because a wrong answer is embarrassing while a wrong action writes to your database.

The Ones That Work

1. Inbound call qualification

The most reliably valuable agent in small and mid-sized business, and not a close race. It answers on the first ring at any hour, asks the qualifying questions a human would, books into a real calendar, and writes a transcript plus structured fields to the CRM.

Why it works: the conversation is bounded, the success criterion is unambiguous, and the counterfactual is voicemail. Cost: $0.05 to $0.15 per minute on a platform you configure, or around $799 per month managed. Where it breaks: anything outside the qualification script. Scope it to hand off rather than improvise. We build these as AI receptionists and voice agents.

2. Lead follow-up sequencing

An agent that decides who to contact next, on which channel, with what message, based on how the lead behaved. Not a drip campaign with variables, an actual decision about whether this person should be contacted at all today.

Why it works: high volume, low stakes per action, and mistakes are cheap. Where it breaks: without a hard cap on contact frequency it will over-contact, because nothing in the loop feels awkward about the fourth message. Set the ceiling in code, not in the prompt.

3. Inbox triage and routing

Reads inbound email, works out intent, extracts the entities that matter, routes to the right queue, and drafts a reply for a human to approve.

Why it works: email is unstructured, which is exactly what models are for, and the human approval step absorbs the error rate. Where it breaks: teams remove the approval step around month three because it seems to be always right. That is when it sends something it should not have.

4. CRM hygiene

Deduplication, enrichment, normalising the fields humans typed inconsistently, and flagging records that contradict each other.

Why it works: nobody wants the job, it is high volume, and errors are correctable. Where it breaks: deduplication itself should not be a model decision. Match deterministically, then use the agent for the ambiguous remainder.

5. Call scoring and coaching

Every call scored against a rubric, with the coaching moments surfaced. A manager reviews maybe 5% of calls by hand. An agent reviews all of them.

Why it works: the rubric makes the output checkable, which is rare. Where it breaks: a generic rubric produces generic advice. It has to be built from your own closers. That is what we do in revenue intelligence.

6. Document extraction

Pulling structured fields out of contracts, invoices, statements, and applications.

Why it works: the input is genuinely unstructured and the output is verifiable against the source. Where it breaks: numbers. Have the agent extract and cite the location, then validate arithmetic in code. Never let a model do the maths.

7. Internal knowledge answering

An agent that answers "how do we handle X" from your own documentation, procedures, and past tickets.

Why it works: it replaces interrupting a colleague, and being 90% right is fine when the alternative is waiting two hours. Where it breaks: stale source documents. The agent will confidently quote a policy you changed last year, and it has no way to know.

8. Reporting and analysis

Scheduled agents that pull numbers, notice what changed, and write the summary a person would have written.

Why it works: the query is deterministic and only the narrative is generated. Where it breaks: letting it compute rather than query. The SQL should be fixed; the agent explains the result.

9. Research and enrichment

Given a company or a person, gather what is publicly known and structure it against the fields your team actually uses.

Why it works: it is genuinely tedious, and partial results still beat none. Where it breaks: confident fabrication of specifics. Require a source URL for every field and drop anything unsourced.

10. Dispo and buyer matching

In real estate specifically: scoring a contract against a buyer list and routing it to the ten most likely to close rather than blasting all four hundred.

Why it works: the feedback loop is fast and the outcome is measurable. Where it breaks: a stale buyer list, which no model can detect. We cover this in AI for wholesalers.

11. Onboarding and intake

Walking a new customer through the information you need, chasing what is missing, and assembling it into one record.

Why it works: it is a checklist with conversation on top, and the completion criterion is explicit. Where it breaks: edge cases in the checklist itself. If your intake has exceptions, write them down before automating.

12. Coding agents

Worth naming because it is the one category where the agent's work is checkable by machine before a human sees it: tests either pass or they do not.

Why it works: an unusually tight verification loop. Where it breaks: anything without tests, which turns it back into a plausible-output problem.

AI Agent Use Cases That Keep Failing

Equally useful, and less often written down.

Anything with an irreversible action and no confirmation step. Sending money, deleting records, contacting a customer about a complaint. Not because agents are unusually bad at these, but because the error rate never reaches zero and the cost of the tail case is unbounded.

Negotiation. It reads well in a demo and behaves badly against a counterparty who is paying attention.

Judgment calls with a legal consequence. Eligibility, compliance, anything a regulator might ask you to justify. "The model decided" is not a defensible answer.

Fully autonomous outbound to cold lists. Beyond the regulatory exposure, the results are poor and it is the fastest way to make your domain and phone numbers worthless. We removed this from our own offering in 2026 for exactly those reasons.

What Separates the Two Lists

Look back at the working examples and one property is present in all of them: you can tell whether the agent was right. A call either got qualified or it did not. A field either matches the document or it does not. A test passes or fails.

The failures share the opposite property. There is no cheap check, so errors accumulate silently and you find them in a quarterly review instead of in the moment.

That gives you a usable filter before scoping anything:

Can a person, or better a piece of code, tell within seconds whether the agent did the right thing?

If yes, it is a candidate. If no, you are not building an agent, you are building a liability with good demo footage.

If You Want to Try One

Start with the boring one. In practice that is nearly always first response to inbound, because volume is high, the cost of being slow is invisible, and the check is trivial: did the lead get a reply in under a minute.

If you want a second pair of eyes on which of these fits your operation, that is what a strategy session is for, and it includes telling you when the honest answer is that none of them do. Our approach to building them is on the AI agent development page.