AI implementation cost

The short answer, before the detail

An AI implementation arrives as three bills, not one. There is what a supplier charges to build the thing. There is what it costs to run after launch, every month, for as long as you use it. And there is what it costs you in your own people's time, which nobody invoices you for and which is the line most budgets leave out entirely.

For a UK business of five to two hundred people, the published figures for the first bill run from a few hundred pounds for one small automation to the high tens of thousands for something custom. The second bill is the one the market disagrees about most, and the third is the one that decides whether the first two were worth paying. The rest of this page takes them in that order, with every figure quoted from the page that published it.

The lines of the budget, and what the UK market publishes for each

Ten pages were read for this, on 4 November 2026, and the figures below are quoted from their own published text. Prices change and these are not quotes for your work; they are the going rates to have in front of you before you talk to anybody.

We do not link to those pages, because a price quoted here ages the day they change it. Read the current figure on their own site before you compare.

The two UK percentages that are not the same number

Ask the field what the running of it costs after launch and you get two percentages that are quoted as if they answered the same question. They do not, and the gap between them is where budgets go wrong.

PrimeWise, writing for a CIO building a board case, says that ongoing operational costs — licensing, token consumption and retainers — "account for approximately 60% of total" cost of ownership. Halo Tech Lab, writing for a UK SME, says to budget "10-15% of the initial build cost annually" for maintenance, and spells it out: a £10,000 integration should have roughly £1,000 to £1,500 a year set aside.

Read quickly, those look like one number given twice. They are not even measuring the same thing. PrimeWise's 60% is every ongoing cost together — licences, token usage and the retainer — as a share of what the whole thing costs over its life. Halo Tech Lab's 10 to 15% is the maintenance line only, per year, and its own page prices usage and storage separately beside it: £50 to £200 a month in API costs for a chatbot handling 1,000 conversations, £20 to £100 a month for storage. Neither firm is contradicting the other. But a business that reads the second number as its all-in running cost has budgeted a tenth of the build and left the usage, the licences and the retainer out, and nothing on either page says which of the two you are being quoted.

Which shape you end up in is a different question, and it is the one worth asking. The 60% shape is what you get when the build is bought as a project and the running of it is bought separately afterwards — a retainer, plus licences, plus usage, each with its own invoice and its own annual increase. The 10-15% shape is what you get when nothing much changes: a single automation, few connections, no model to retrain. Most real implementations sit nearer the first than the second, because the thing that drives the cost is not the agent, it is your software changing underneath it. A redesigned invoice, a renamed field in the CRM, a supplier who switches to sending PDFs — any of those can stop an agent while every dashboard still looks healthy, and somebody has to be paid to notice and repair it.

So the honest version of the question is not "what percentage?" but "who is on the hook when it breaks, and is that already paid for?" AI integration services is the page about those joins and the repairs they need.

The part of the bill nobody sends you

Nearly every page in this field prices what a supplier charges rather than what the work takes out of you. One of the ten does both: Halo Tech Lab tells a team of ten to budget 2 to 3 days of reduced productivity during rollout, and names the internal champion's ongoing time after it. It is the exception, and it stops at those two lines. The rest of what follows is unpriced everywhere, and in the implementations that go wrong, your own side is where the cost was.

Three people decide whether an implementation lands, and usually none of them was in the room when it was scoped. An administrator has to grant the access, which needs somebody above them to agree it. A manager has to change how the work is done, because an agent that handles the easy half of a job has not saved anybody time until the old process stops. And the first genuinely awkward exception has to land on a named person's desk rather than nobody's. LSI's guide makes the same point from the enterprise end, that the hidden costs sit outside the technology and the operating model is the main investment. It is as true at eight people as at eight hundred.

In hours, for a small business, this is not enormous — a day or two of one senior person's attention spread over the first few weeks, then ten minutes a day of someone reading what happened. It is the fact that it is unbudgeted that hurts, because unbudgeted time comes out of whoever has least slack, and that is usually the person the whole thing depends on. Put it in the plan with a name against it and the implementation stops being fragile.

Working out whether the number is worth paying

The arithmetic is simpler than the field makes it look, and you can do it before you speak to anybody.

Take one task you would hand over. Count the hours a week it really takes, including the chasing and the checking. Multiply by a fully loaded hourly cost — AI Workforce's own illustration uses 5 to 10 hours a week of manual invoice work at £20 to £30 an hour, which it values at £400 to £1,200 a month of released capacity. That is your ceiling. Anything that costs more than that per month has to earn the difference somewhere else: fewer missed enquiries, faster payment, work that happens at eleven at night instead of waiting until nine.

Then apply the discount nobody applies. Released capacity is only worth money if something is actually done with it. An hour freed from invoice entry is worth the hourly rate only if that hour goes to work that earns; if it goes to the general pressure of the week, the saving is real but it will not show up in your accounts. The claims in this field assume the first case. Softrobo, writing about UK SMEs, puts successful implementations at £3.70 returned for every pound invested and cost savings of 20 to 40%. Treat a figure like that as what a good case looks like rather than what an average case returns, because it is a supplier's own number and no supplier publishes the ones that failed.

The useful test is narrower and you can check it yourself in a month: did the person whose task it was stop doing that task, and did anything else go wrong because of it?

What a flat monthly fee does to the shape of all this

Most of what is above is a consequence of one pricing habit: the build is sold as a project and the running of it is sold separately. That is why "ongoing is sixty per cent of the total" is a sentence anyone needs to write, why the second agent is a fresh negotiation, and why the repair after your CRM changes is a conversation about who pays.

Ours is the other shape. One flat monthly fee for unlimited agents, month to month. Nothing extra per call, minute or message. Stop any month. The practical difference is not the headline figure, it is the second year: when the question "could it also do this?" costs nothing to ask, it gets asked, and the agents end up covering the jobs you found after the first one was live rather than the jobs you could think of before it.

The figure itself is one we give on the call rather than publish, once we know what the agents have to do and what they have to reach — and it comes with the guarantee at the foot of this page, which is a promise about the agents working rather than about the calendar.

Where this page stops, and which page carries on

This page is the budget. What an AI agent costs is the price list beside it: what UK agencies charge to build one agent, what the off-the-shelf tools charge per month, and what moves the figure. If you are comparing quotes, read that one. If you are building a case for spending the money, this one.

AI implementation services is what an implementation actually involves, stage by stage, which is the same subject without the money. Choosing an AI implementation partner is the nine questions to put to whoever is quoting. AI agent development is the build itself, and AI consultancy in the UK is the version of this that ends in advice rather than a working agent — priced by the day, which is a different kind of bill again.

The thing on this page that is not a claim

The agent in the panel here is a working one. Ring it, interrupt it, ask it something it will not be expecting. An implementation costs what it costs; what it produces is the only part you can test before you spend anything, and that is the part we would rather be judged on.

Bring us the job

Tell us the task, or the part of somebody's week you would like back. On the call we will say which of it an agent can hold, which stays with a person, what access we would need from you and roughly how many hours of whose time, which part of it worries us, and the price.

Unlimited agents. One flat monthly fee. Fully functioning and delivering value in 30 days guaranteed, or you pay nothing more until your agents are.