Search for AI agent examples and you mostly get demos: an agent that books a holiday, an agent that orders a pizza, an agent that browses the web while someone narrates excitedly over it. Impressive on stage, absent from production.
The agents that actually run inside businesses look different. They are narrower, more boring, and considerably more valuable. Below are ten that reach production regularly, grouped by the pattern that makes them work — plus an honest note on what each one is genuinely worth.
What Separates a Production Agent From a Demo?
Three traits, consistently.
The task is bounded. "Handle customer emails" is not a task; it is a department. "Read an incoming email, classify it into one of five categories, extract four named fields, and file it" is a task. Bounded tasks have success criteria you can actually evaluate against.
The actions are reversible. Filing a record, drafting a reply, adding a CRM note, flagging a document for review — all cheap to undo. Sending an email to a customer, issuing a refund, or updating a production database are not. Production agents accumulate irreversible powers slowly, and usually behind an approval step.
Something expensive has a human on it. Not every step. Just the ones where being wrong costs real money or credibility. Teams that put a human on everything end up with a slower version of the manual process; teams that put a human on nothing end up with an incident.
An agent that fails these three tests can still demo beautifully. It just will not survive contact with a real workload.
Document and Back-Office Agents
This is where the boring money is. Document work is high-volume, low-variance, and genuinely tedious, which is exactly the profile automation suits.
1. Invoice triage and extraction
The agent watches an inbox or a folder, identifies invoices, extracts supplier, amount, date, PO number and line items, matches them against purchase orders, and either files the match or flags the discrepancy for a human.
Why it works: the output is structured and checkable. If the agent extracts £4,200 and the invoice says £4,200, it is right, and you can verify that automatically against thousands of historical documents before you ever go live.
Honest limitation: supplier invoice layouts vary wildly, and the long tail is genuinely hard. Expect strong performance on your top 50 suppliers and a review queue for the rest.
2. Contract and policy review
The agent reads a contract against a checklist — payment terms, liability caps, auto-renewal clauses, data-processing language — and produces a summary flagging anything outside your standard positions.
Why it works: it never replaces the lawyer, it front-loads them. Review time drops because the reader arrives knowing where to look.
Honest limitation: this is advisory only. An agent that approves contracts is a liability; one that annotates them is a genuine time saver.
3. Expense and receipt processing
Photograph a receipt, and the agent extracts merchant, amount, date and category, checks it against policy, and either files it or asks a question.
Why it works: policy rules are explicit, and the amounts are small enough that occasional errors are cheap.
Customer-Facing Agents
Higher value, higher risk. These need more guardrails.
4. Support ticket triage and routing
The agent reads an incoming ticket, classifies urgency and topic, attaches relevant account context, routes it to the right queue, and drafts a suggested first reply.
Why it works: routing is a decision with a knowable right answer, and you have thousands of historical tickets to evaluate against. The drafted reply is a suggestion, not a send.
Honest limitation: the moment you let it send unsupervised, the economics change completely — one confidently wrong answer to an angry customer costs more than the triage saved.
5. Order and account lookup
A customer asks where their order is; the agent looks it up across the order system and the courier API and answers with a real status.
Why it works: it is read-only, and the answer is verifiable.
6. Refund and returns processing
The agent checks the order, applies the returns policy, and either processes a refund within a threshold or escalates.
Why it works: the policy is explicit and the threshold bounds the damage. A £15 auto-refund limit means the worst case is £15.
Honest limitation: this only works where policy is genuinely deterministic. If your refund decisions involve judgement about customer history, keep the human.
Sales and Research Agents
7. Lead research and CRM enrichment
Given a company name, the agent gathers public information — size, sector, tech stack, recent news, likely fit — and writes it into the CRM record before a rep makes contact.
Why it works: it is read-only, and the alternative is a salesperson doing twenty minutes of tab-hopping per lead.
Honest limitation: it will occasionally confidently attribute a news story to the wrong company of a similar name. Cite sources in the output so the rep can check.
8. Meeting notes to CRM actions
The agent takes a call transcript, extracts commitments and next steps, and creates the follow-up tasks.
Why it works: the transcript is the ground truth, and the output is a task list a human immediately reviews.
Internal Knowledge Agents
9. Policy and documentation lookup
An agent grounded in your handbook, runbooks and internal wiki answers "how do I request equipment" or "what is our incident escalation path" with a citation to the source document.
Why it works: the citation is the whole design. An answer without a link is a claim; an answer with one is a shortcut to the document.
Honest limitation: it inherits your documentation's quality exactly. If the wiki is three years stale, the agent confidently repeats three-year-old policy.
10. Onboarding assistant
A variant of the above pointed at new starters, answering the questions that otherwise consume a manager's first fortnight.
Why it works: the questions are repetitive and the answers are documented. This is also where the agent pattern meets our wider work — an onboarding assistant with a face is an interactive AI avatar, and the same grounding rules apply.
The Pattern Behind All Ten
Look across the list and the shape is consistent: high volume, low variance, checkable output, cheap mistakes. None of them require the agent to be creative. All of them require it to be reliable.
That is the inversion most teams get wrong. The instinct is to point an agent at the most interesting problem in the business. The agents that reach production get pointed at the most repetitive one, where success is measurable and failure is recoverable.
If you are choosing a first agent, the useful question is not "what would be most impressive?" It is: what work happens hundreds of times a month, has a right answer, and costs little to get wrong occasionally? That is your candidate.
Where Teams Get Stuck
The failure is almost never the model. In the builds we have seen stall, the cause is one of three things:
- Missing context. The agent cannot see the system that holds the answer, so it guesses. Fix the retrieval, not the prompt.
- No success criteria. Nobody defined what "correct" means, so nobody can tell whether it is working, so it never leaves pilot.
- An action it should never have had. The agent could do something irreversible, did it once, and the project lost organisational trust permanently.
That last one is worth dwelling on. Trust is the real budget in agent projects, and it is spent in a single incident. Start read-only. Add write access one action at a time, each behind an approval step until it has earned its way out.
We build these as fixed-fee engagements: a scoping call, a working prototype against your real data in two to three weeks, then the guardrails and evaluation that turn it into something you can rely on. If you want to talk through which workflow in your business is the right first candidate, book a scoping call or read more about our AI automation practice.