What is an AI agent, in practice?
The technical definition is settled and widely shared. Google Cloud, on its reference page last updated 2 April 2026, describes an agent as software that pursues goals and completes tasks on a user's behalf, with reasoning, planning, memory and a degree of autonomy. Its own comparison table is the useful part: an assistant recommends and the user decides, while an agent acts autonomously and proactively. The Institute for AI Policy and Strategy, in its April 2025 field guide, uses a sharper wording: systems that achieve goals in the world with little or no explicit human instruction about how.
Those definitions are correct, and they tell you almost nothing about what will happen in your company. The useful translation fits in one sentence: an AI agent is software you hand a task described in writing, with limited access to your tools, permission to prepare, and a boundary past which it has to ask.
Three things separate an agent that earns its place from one that adds noise. The task: repetitive, described precisely, with a result somebody can check. The material: an agent only knows what it can read, and an agent wired into disorganised information produces disorganised work. The boundary: what it does alone, what it submits, what it never touches.
What an AI agent actually does
These are the five jobs where the difference shows up fast, because they repeat and because the output can be checked in seconds.
It sorts and qualifies what comes in. An enquiry arriving by form, email or phone gets read, matched to the right customer, tagged by subject and urgency, and filed where your team will look for it. It is the least impressive task on this list and the most profitable one: it gets done badly when everyone is busy, which is exactly when it matters.
It prepares documents from your own rules. A quote, a standard reply, a meeting brief, a summary. The agent collects the pieces from your tools, applies your rules, and puts the document in front of you with its sources attached. You read it, correct it, send it. That is the principle behind our enquiry and quote agents: the proposal is ready, the signature stays human.
It follows up on a schedule nobody keeps. The quote sent on a Friday, the missing document asked for once, the appointment still unconfirmed. An agent does not forget and does not find a third follow-up awkward. That is also the work a prospecting agent does earlier, on first contact.
It watches and warns. Stock running down, a deadline approaching, a customer silent for three weeks, an invoice past due. The agent decides nothing: it puts the alert in front of the right person, with the context.
It answers questions whose answer already exists inside your company. Opening hours, lead times, the returns procedure, the status of an order. Here too the agent only knows what it was given to read, and the quality of its answers depends first on what you tidied up before plugging anything in.
What an AI agent does not do
This section is missing from almost every page that ranks for this question, and it is the one that prevents unpleasant surprises.
It does not decide for you. An agent can propose a discount. It cannot judge that a customer deserves a gesture because they have been loyal for twelve years and have had a bad one. That decision commits your company. It stays yours.
It does not know your trade. A language model has read a great deal of general text and nothing at all about your business. Your prices, your exceptions, your difficult accounts, the reason you never deliver on a certain day: all of it has to be written down somewhere for the agent to take it into account. Writing the rules is the real cost of a project, not the technical wiring.
It does not replace the relationship. A customer who calls because they are unhappy wants a person. A well-set agent has the file ready before you pick up; it does not take the call for you.
It does not supervise itself. An agent running with nobody reading its output drifts quietly. Somebody on your side has to review a sample each day, at least for the first few weeks, and know how to switch it off.
It does not fix a disorganised company. If three people answer the same enquiries without talking to each other, the agent will do the same thing faster. Automation amplifies what is already there.
| The task | What the agent does alone | What stays with a person |
|---|---|---|
| Incoming enquiry | Read, file, match to the customer, flag urgency | Deciding what happens next when the case breaks the rule |
| Quote | Gather the parts, apply the rules, prepare the document | The price, the commercial gesture, the send |
| Follow-up | Keep to the schedule, prepare and send the agreed message | Choosing when to stop following up |
| Alert | Spot the gap, put the information up with its context | The call and the action |
| Common question | Answer from what is written down | Writing and maintaining what counts as the reference |
How to spot a promise that will not hold
The most common signal is a productivity percentage on a home page. A productivity gain cannot be read as a single global figure: it is observed on one specific task, in a company whose size and trade you know, with a measurement before and a measurement after. Until those three things are given, the percentage tells you nothing about what you will get. Ask for them. We publish no figure of that kind, because we do not yet have a before-and-after measurement we can show.
Three more signals should slow you down. The phrase "it handles that on its own": nobody should offer to automate an action that commits your company without a check. No written list of access rights: an agent is an account that acts, and you need to know what it can read and write. And the project that starts with five processes at once: take one, make it work, look at what it produces, and only then add a second.
Where to start in a twenty-person company
Take the task that comes back most often, that nobody enjoys, and whose result can be checked at a glance. Write down what it requires, exceptions included. Give the agent the minimum access it needs. Set the boundary between what it sends alone and what it submits. Read its output for two weeks, then widen the remit or stop.
That is the method we apply, and we run this way ourselves: our agency is a team of agents, each with a written role definition listing its permitted tools and what it is not allowed to do, a board where every piece of work passes with its source attached, and a human sign-off on everything that leaves the building. We do not sell technology we do not use. The flip side is that we also know what it costs in rule-writing time at the start, and we say so before we begin. The full approach is on our AI agents for small and medium businesses page.
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot answers inside a conversation and stops there. An agent goes and gets information from your tools and carries out actions: filing an enquiry, preparing a document, sending a follow-up. What separates them is access to tools and permission to act, not the quality of the text they produce.
Can an AI agent work unsupervised?
On tasks with no external consequence, yes, once it is running properly: filing, preparing, alerting. On anything that reaches a customer or commits a price, a deadline or a promise, no. The constraint is managerial before it is technical, and it has to be written down before go-live.
How long before there is a result?
It depends entirely on the state of your information. A well-bounded task, with rules already written and tools you can reach, goes live quickly. A task whose rules exist in nobody's head needs that writing work first, and the writing is what takes the time. We do not promise a timeline before looking.
Do you need someone technical in-house?
No, but you need someone on your side who knows the task, reads what the agent produces in the first weeks, and knows how to stop it. That is a real load, a few minutes a day at the start, and it cannot be handed to a supplier.
Sources: "What are AI agents? Definition, examples, and types", Google Cloud, page last updated 2 April 2026 (cloud.google.com) · "AI Agent Governance: A Field Guide", Jam Kraprayoon, Institute for AI Policy and Strategy, 17 April 2025 (iaps.ai) · "AI agents for small business: what they can actually do (and where to start)", Sharon Hafuta, Wix, 19 August 2026 (wix.com) · Search volumes quoted in the metadata come from Ahrefs (United States and United Kingdom), measured 28 September 2026. The practices described as ours are those of the Équipage IA team charter. No result figure and no customer case is quoted in this article.