How much does an AI agent team cost in the UK? Build prices, running costs and the 14.7% waste we found

When we audited where the computing time on our own research programme was going, we expected to find a few small leaks. Instead we found that nearly a seventh of all the computing time was going on work that nothing ever used.

That audit changed how we price, and how we build. It is why this post starts with the uncomfortable part: AI costs are easy to underestimate and easy to waste, unless you can see them.

So here is a straight answer to the question we hear second most often, right after "is my data safe?"

The two costs, in plain terms

Think of an agent team like a specialist committee you are setting up for the first time.

Building it is the one-off work of deciding who sits on the committee, writing its rules, giving it access to the right documents, and setting up how it records its decisions. That is design and engineering work, and we quote it as a fixed price.

Running it is what each meeting costs. For an agent team that is mostly model usage: the fees the AI provider charges for the reading, reasoning and writing the agents do. It is paid per use, so a quiet month costs less than a busy one.

What building costs

Package	Length	Price	What you get
Discovery Sprint	2 weeks	from £4,500	One decision-heavy workflow mapped, the team and its rules designed, and a working session on your own question
Agent Team Build	4 to 6 weeks	from £18,000	The team built in your own environment, the council with recorded minutes, cost controls and session caps, the tamper-evident audit trail, handover and training
Run and Improve	Monthly	from £3,500 a month	We operate and tune the team, run deep research sessions on your questions, and report monthly on decisions made, accuracy and spend

Most clients start with the Sprint, for a simple reason: it ends with real minutes on a real decision, so you can judge the result before spending more.

What running costs

A short council session on a business question, with five or six agents researching, debating and voting, typically costs a few pounds in model usage.

A deep research session costs more: a survey of a whole field, several position papers, critics assigned to each paper and a final review. We quote those individually and set a cap before they start.

Three things decide the running cost of any session:

1. How many agents take part. Not every question needs every specialist.
2. How much they read. A session over two reports costs far less than one over two hundred.
3. Which models they use. Not every role needs the most expensive model, which brings us to the most useful lever you have.

Choose the model by the job

A finance agent checking a complicated investment case may need strong reasoning. A note-taker that tidies the minutes does not. Paying top-tier prices for every role is the most common way teams overspend.

We choose a model per role, and show you the cost of each choice before anything goes live. You can always move a role up a tier if its work is not good enough. You rarely need to.

Why every session needs a cap

Agents that can search, read and debate can also keep going long after they have stopped adding value, and run up a bill while nobody is watching.

That is why every session in a governed team has a spending cap set before it starts. When the cap is reached, the session stops and the chair writes up what it has. The minutes then show the actual spend, per session and per agent.

You never get a surprise invoice. And you can see, in black and white, which agents earn their cost.

The waste we found in our own work

Back to that audit. On our own research programme, we traced every piece of work the system did and asked one question of each: does anything downstream actually use this?

Removing the waste made the measured run 1.67 times faster, and the results were byte-for-byte identical. Nothing was lost except cost and time.

You cannot cut waste you cannot see. A cost line in every set of minutes is how you see it.

The same thinking applies to a business agent team. When every session reports its spend per agent, the agents that do not pull their weight become obvious, and you can drop them or give them a cheaper model.

The cost nobody puts on the invoice

Here is the comparison most buyers never make. Take that pricing review that keeps coming back.

Say six senior people spend ninety minutes on it, and it reopens three times a year because nobody can remember why the last decision was made. That is twenty-seven hours of senior time a year on one recurring question, before anyone counts the preparation, the follow-up emails or the cost of a slow decision.

Put your own salaries into that sum. For most firms, it is far larger than the running cost of an agent team that does the evidence gathering, records the reasoning and stops the question reopening in the first place.

Hidden costs to plan for

To be fair to your budget, plan for these as well:

• Your people's time in the Sprint: a few hours to explain the decision and review the rules.
• Access to your systems: if the agents need to read from internal tools, your IT team will need to grant it.
• Change: a new way of reaching decisions works best when one person owns it.

We name these up front in every proposal, so they do not surprise you later.

The honest summary

• Building is fixed and quoted before you start.
• Running is small per session, capped in advance and reported in the minutes.
• The real saving is usually time: fewer reopened decisions, shorter meetings and less rework.

If you want a number for your own case, a thirty-minute call is enough for a rough estimate, and the Discovery Sprint turns it into a firm one.