The short answer
You need an AI agent when there is a piece of work in your business that arrives in different shapes every time, needs judgement to route, touches more than one system, and is currently done by a person who would rather be doing something else. Customer enquiries, onboarding, order exceptions, scheduling and internal requests are the usual candidates.
You do not need one when the work is the same every time (that is automation, and it is cheaper), when the work is mostly conversation with no action at the end (that is a chatbot, or a person), or when nobody can say what a good outcome looks like. An agent with no definition of done is a liability with an API key.
Chatbot, automation, agent: the three things being sold under one name
A chatbot answers. You ask it something, it replies, and the conversation is the whole product. It may be excellent at replying; it does not do anything in your systems afterwards. Most “AI on our website” is this.
Automation acts, but only along a path somebody drew in advance. When an invoice arrives, extract the total, create the bill, notify accounts. It is reliable precisely because it does not think, and when the invoice looks unusual it either fails or does the wrong thing confidently.
An agent sits between the two. It is given an outcome (“resolve this customer’s delivery question”), a set of tools it may use (look up the order, check the courier, draft a reply, issue a credit up to a limit), and rules about what it may do alone and what needs a person. It works out the steps for each case, which is what makes it useful on work that varies, and it is also what makes it need a design.
What an agent is made of
Strip the marketing away and every agent has the same five parts. The model, which reasons and writes. The knowledge it may draw on, which should be your documents and systems rather than the open internet. The tools it may call, which are the connections to your CRM, email, calendar, accounting or ERP. The rules, which say what it may do without asking. And the record, which logs every step so someone can see afterwards what happened and why.
The part suppliers talk about is the model. The parts that decide whether the agent works in your business are the other four. An agent with a brilliant model and no knowledge of your policies will answer confidently and wrongly. An agent with no rules will issue a refund it should not. An agent with no record cannot be trusted with anything that matters, because when it goes wrong nobody can find out how.
The approval step is the whole design
The question that decides whether an agent is safe is not how clever the model is. It is: which actions may it take on its own, and which must wait for a person? A well-designed agent drafts a reply and sends it if the question was routine; it drafts a credit note and waits if money is involved; it escalates immediately if the customer is angry or the request is outside its knowledge.
That line is drawn by you, in writing, before the agent is built, and it should be drawn conservatively at first and widened as the record shows the agent is reliable. Businesses that skip this step either end up with an agent that does nothing useful because it asks about everything, or one that does something expensive because it asked about nothing.
Five questions that tell you whether you need one
Four or five yeses and an agent is probably the right shape. Two or three and you are more likely looking at an automation or a knowledge assistant, both of which are simpler and cheaper. Fewer than two and the honest advice is to fix the process before adding software to it.
- Is there a stream of work that varies case by case but follows recognisable patterns? If every case is unique, a person is cheaper. If every case is identical, automation is cheaper.
- Does the work touch more than one system? An agent earns its keep by reading from one place and acting in another.
- Can you write down what “done” looks like for a typical case, and what “stop and ask” looks like?
- Is the information the agent would need digital and reasonably findable? An agent over paper files is a scanning project first.
- Is there a person who will own it: check the record weekly, adjust the rules, and answer when it escalates? An agent nobody owns drifts.
What it costs to get wrong
The failures we see are rarely the model being wrong. They are an agent that had access to more information than the person asking it (so it revealed something it should not have), an agent that could act with no limit (so it did), and an agent with no record (so nobody found out until a customer did). All three are design failures, and all three are avoidable by deciding, before the build, what the agent may know, what it may do, and what it must log.
That is the difference between an agent bought from a demo and an agent that was designed. The demo shows you the model. The design shows you the boundaries.
Where to go from here
If the five questions came out mostly yes, the AI Agents page on this site describes how we design one: the knowledge it draws on, the tools it may use, the approval rules, and the record, before a line of it is built. If they came out mostly no, take the AI Opportunity Score below; it will usually point you at the automation or the knowledge assistant that fits better.
Published 12 September 2026 · Be AI