Every AI agent quote I have seen this year leads with a monthly seat price. Twenty pounds, fifty pounds, a hundred and fifty for the enterprise tier.
That number is real, and it is almost never the number that decides whether the thing pays for itself.
Budget season starts in earnest next month, and a lot of UK small businesses will be looking at an AI agent line item for the first time. Here is the full cost picture, in four parts, and the questions that get each one out of a vendor before you sign.
Cost one: the subscription, which everybody quotes
The sticker price. Per seat, per month, sometimes per agent.
It is the easiest cost to compare and the least useful, because seat pricing tells you nothing about whether the agent can do your job. Two products at the same monthly price can differ by a factor of ten in what they will actually handle without a human stepping in.
Ask: what happens to the price at double our volume, and is there a usage component on top of the seat fee.
Cost two: usage, which appears in month three
This is the one that surprises people.
Most AI agents are built on a large language model that charges by the token, which is roughly a fragment of a word, and the charge applies to everything going in and everything coming out. A vendor either passes that through to you or absorbs it into a plan with a cap.
Two things drive it up faster than anyone expects. Context, because an agent that reads your whole knowledge base before answering is paying for that reading on every single query. And retries, because an agent that gets it wrong and has another go has just doubled the cost of that task.
Published model pricing, such as Anthropic’s rate card ↗, is public and worth ten minutes of your time even if you never touch an API. It tells you what the raw ingredient costs, which tells you whether a vendor’s flat fee is generous or a trap that closes when your volume rises.
Ask: is usage included, capped, or billed on top, and what does a typical month look like for a business our size.
Cost three: integration, which is the real number
An agent that cannot see your data is a chatbot.
The value in an AI agent comes almost entirely from connecting it to the systems where your business actually lives: your bookings, your customer records, your inbox, your accounts. That connection is where the cost sits, and it is the part that varies from nothing to a project.
- If your tools are mainstream and have good APIs, connection is often configuration rather than development.
- If you run anything bespoke, legacy, or held together with spreadsheets, you are paying for integration work, and the estimate should be in days rather than minutes.
- If your data is messy, and most small business data is, you are paying to clean it before any of this works. This cost is almost never in the quote.
We wrote about the wider version of this problem in the hidden costs of vibe coding and in what it really costs to own custom software. The pattern is identical: the build is the visible cost and the connection is the expensive one.
Ask: which of our existing systems does this connect to out of the box, and what does connecting the rest involve.
Cost four: supervision, which never goes away
An AI agent that runs unsupervised is either doing something trivial or about to embarrass you.
Somebody has to check the output, at least at first. Somebody has to decide what the agent is allowed to do on its own and what it must hand to a human. Somebody has to notice when it starts getting a particular kind of request wrong. That is a real, ongoing cost in staff time, and it is the one that makes the difference between an agent that saves five hours a week and one that creates three.
The supervision cost falls over time if the agent is well scoped, and rises forever if it is not. An agent asked to handle one narrow, high-volume task gets checked less and less. An agent asked to “handle customer enquiries” needs checking indefinitely, because the range of things a customer might say has no edge.
There is a compliance dimension here too. If the agent touches personal data, you carry the accountability for what it does with it, not the vendor. The ICO’s guidance on AI and data protection ↗ is the starting point, and our post on giving an AI agent access to business data safely covers the practical side.
Ask: what does the agent do when it is not confident, and who reviews what.
The comparison that actually matters
Put the four costs against the thing the agent replaces, not against another vendor.
| What you pay | What it replaces | |
|---|---|---|
| Subscription | Fixed, predictable | Nothing on its own |
| Usage | Scales with volume | The marginal cost of the task |
| Integration | Mostly upfront | The copying and pasting between systems |
| Supervision | Ongoing staff time | The judgement the task needed anyway |
If a task takes a member of staff two hours a week and the agent handles 80 percent of it with 20 minutes of checking, the maths is easy and the sticker price is irrelevant. If the task takes twenty minutes a week, no agent is worth integrating, at any price.
That is the honest filter: volume, repetition, and a clear boundary. Everything that clears all three is worth pricing. Everything else is a demo you enjoyed.
Where to start if the budget is small
Start with one task that a person currently does more than twenty times a week, that follows the same shape every time, and where a wrong answer is embarrassing rather than expensive.
Email triage is the usual winner, which is why we wrote about AI agents for small business email separately. Document extraction is the other. Both are high volume, well bounded, and easy to supervise while you build confidence.
Do not start with anything customer facing until you have watched an agent work for a month. And be sceptical of anything sold as replacing a role rather than a task; we picked that apart in what AI agents can and cannot do for small business.
The one line summary
The monthly fee is not the cost. Usage, integration and supervision are the cost, and a vendor who will not quantify all three is quoting you the easy number and leaving you the hard ones.
If you want a straight answer about whether an agent fits a specific job in your business, get in touch and we will tell you if it does not.