AI Agents for Business: Use Cases, Costs and What to Watch
An AI agent takes actions in your systems rather than only answering questions — raising a ticket, updating a record, scheduling something. That single difference changes the risk profile completely: a chatbot that is wrong says something incorrect, an agent that is wrong does something incorrect. Permissions, confirmation and audit logging are most of the build.
Most businesses asking for an agent need a grounded assistant. That is not a criticism — assistants are useful, cheaper and far lower risk. The agent is warranted when the valuable part is the action rather than the answer, and when the action is repetitive enough to be worth automating.
- Chatbot risk is a wrong answer. Agent risk is a wrong action. Budget accordingly.
- Scope the permissions tightly. An agent should hold the narrowest access that does the job.
- Anything consequential needs human confirmation before execution.
- Log everything. You will need to reconstruct what it did.
Where agents genuinely earn their place
- Triage and routing — reading an incoming request, classifying it, creating the right record and assigning it with context.
- Data reconciliation — comparing records across systems and flagging or fixing the differences under rules.
- Document workflows — extracting from an incoming document, creating the downstream records, routing for approval.
- Internal operations — the repetitive multi-system sequences a person currently does by hand.
Notice what these have in common: they are internal, recoverable, and repetitive. Customer-facing agents that commit the business to something are where the risk concentrates, and they warrant far more caution than the demos suggest.
The controls that make it deployable
| Control | Why |
|---|---|
| Least-privilege access | The agent can only touch what its job requires |
| Confirmation on consequential actions | A person approves anything that spends money or contacts a customer |
| Full audit log | What it did, when, on whose behalf, and why |
| Reversibility | Prefer actions that can be undone over ones that cannot |
| Rate and scope limits | A loop should not be able to process ten thousand records |
| An evaluation set | So a prompt or model change can be measured, not guessed at |
What they cost
Typically £30,000 and upwards to build, against £10,000–£30,000 for a grounded assistant. The difference is not the model work — that part is similar. It is the permissions model, the confirmation flows, the audit trail and the testing, which together are most of the project and all of the reason it is safe to run.
Frequently asked questions
What is the difference between an agent and a chatbot?
A chatbot answers. An agent acts — calling APIs, updating systems, completing multi-step tasks. The second needs permissions and an audit trail; the first mostly needs grounding and a refusal path.
Are agents reliable enough for production?
For bounded, internal, recoverable tasks with confirmation on anything consequential, yes. For open-ended autonomy over systems that matter, not yet — and anyone selling that is selling a demo rather than a deployment.
Should it run without human approval?
For low-stakes reversible actions, yes, or there is little point. For anything that spends money, contacts a customer or cannot be undone, no — the confirmation step is cheap and the alternative is not.
How do we start safely?
Pick an internal task where a mistake is annoying rather than costly, run it with confirmation on every action for a few weeks, read the logs, and only then relax the controls on the steps that have proved themselves.
We build these as AI agent projects, usually alongside conventional automation doing the deterministic parts.
