What an AI agent is, and when a UK SME actually needs one.

An assistant answers a question a person asks: ChatGPT, Claude or Copilot in a chat window. A workflow is a fixed sequence of steps triggered by an event: when a form is submitted, create a record and send an email, which n8n, Make or Zapier do well. An agent sits between them: it is triggered by an event, but instead of following a fixed sequence it reads the situation, decides which steps to take, takes them through your systems, checks the outcome and decides again, until the job is done or it hands to a person. The difference that matters to a buyer is decisions: a workflow cannot make one; an agent makes many; an assistant makes them only when asked.

Three conditions, all of them together. The work runs unattended, out of hours or at volume, so a person cannot be in the loop for every case. It happens inside your own systems, the CRM, the inbox, the calendar, the ledger, so it has to read and write there. And it contains decisions a fixed workflow cannot make: which applicant to chase and what to say, whether an invoice matches an order and what to do if not, whether a caller needs a person now. Reading every application, chasing every applicant and booking every interview is an agent. Sending a confirmation email when a form arrives is a workflow. Writing the proposal is an assistant and a person.

If only one or two of the three apply, the cheaper thing is usually right, and our guides on workflow tools and on off-the-shelf software set out how to tell.

At fixed-price UK rates a first agent is an implementation at £4,000 to £15,000, or a custom AI system at £8,000 to £25,000 where several agents share data and hand-offs, with the running cost stated per feature before go-live. The running cost scales with volume, not headcount, and the channels usually outweigh the model: a text message or a minute of telephony costs more than the model call behind it, and a well-routed agent sends most steps to the smallest model. Our agent-cost guide gives worked examples at published prices, and the case study on cost governance describes how we keep the number visible and capped.

An agent that decides can decide wrongly, at volume, unattended. Four controls make it safe. A human decides anything that affects a customer or an individual, with the agent preparing and the person confirming, which is also what the rules on significant automated decisions require. Spending ceilings per feature and per customer, enforced in real time and tested by watching them fire, so a loop cannot run up a bill. An acceptance threshold as a number, measured on real inputs before go-live and monitored after. And an audit trail on every action, so what happened can be reconstructed and, where people are involved, explained. Scoped credentials so the agent cannot reach what it does not need are the fifth, and the one most often skipped.

Design the fallback before the clever part. Roughly one outbound call in eight reaches a human on the first attempt, so the agent has to be built around the seven that do not. Small models do most of the work; the frontier model handles roughly one call in eighty. The exception path, what the agent does when it is unsure, is where the quality lives, and an agent that says “I am not sure, here is what I found” to a person beats one that guesses. And instrument cost from the first request, because an agent is the one kind of software that can spend money while everyone is asleep.