Nearly half of UK businesses use AI. Almost none of them have an agent.

There is a gap between what UK businesses call AI and what an AI agent actually is, and almost everybody selling into that gap is being vague about it on purpose. Official figures say 41% of UK businesses use AI. Look at what they use it for and it is mostly a person typing into a chat box. Here is the real difference, three worked examples with the sums shown, and the four questions that decide whether it is worth building.

Start with the real number, not the pitch

The Department for Science, Innovation and Technology publishes the UK Business Data Survey. The 2026 edition ran from October 2025 to January 2026 across 4,450 businesses, and it found that 41% of UK businesses handling digitised data used AI-based technologies.

Split by size, adoption is not the story people expect. Large businesses are at 82%. Medium are at 58%, small at 51%, micro at 41%, and sole traders at 40%. The gap between a small business and a large one is real but it is not a chasm.

Then look at what they are doing with it. The most common use was researching information, at 28%, described in the survey as being in place of a traditional search engine. Second was summarising or collecting in-house information, or drafting reports and correspondence, at 21%.

Read that again, because it is the whole point of this article. The dominant use of AI in UK business is a person opening a chat box and typing into it. That is genuinely useful. It is not an agent, and the distance between the two is where both the value and the risk are sitting.

What an agent actually is

An assistant answers when spoken to. You ask, it produces something, you decide what to do with it. Every output passes through a person before it touches anything real.

An agent takes a sequence of actions across your systems without a person approving each step. It reads the enquiry, checks the calendar, writes the record, sends the confirmation. Nobody signs off the middle.

The difference is not intelligence. Both are running much the same models. The difference is permission to act, and that single distinction is why an agent can save you an afternoon a week and also why it can do an afternoon worth of damage before anybody notices. Anybody selling you one who does not open with that trade-off has not thought about your business.

Three examples, with the sums shown

Abstractions are why this subject stays confusing. Here are three real shapes, for three businesses that exist on every high street in the country. The numbers in them are not our claims about your business. They are the arithmetic you should run on your own, and we have shown our working so you can replace every figure with yours.

The dental practice and the new patient enquiry

The form on the website sends an email. Somebody on reception reads it, decides whether it is a new patient or an existing one, checks whether the treatment they mentioned is one the practice offers, looks at the diary, replies with two or three appointment options, and adds a note to the practice management system.

Say that happens twelve times a day and takes six minutes each time. That is 72 minutes a day. Call it 46 working weeks after holidays and bank holidays, and you are at roughly 275 hours a year of somebody standing at a desk retyping things that were already in an email.

An agent version reads the enquiry, classifies it, checks real availability, drafts the reply with the actual open slots in it, and writes the record. A human still presses send. That last sentence is the whole design: reception spends fifteen seconds approving rather than six minutes assembling, and no patient ever receives something nobody read.

That is not a transformation. It is one job, and it is the kind that pays back. It is also worth reading alongside what a dental practice website costs, because the enquiry form feeding this is the part most practices have never looked at properly.

The trades business and the quote that never got followed up

A plumbing or electrical firm sends quotes and then loses track of them. Not through carelessness, but because the follow-up lives in one person head while that person is under a floor.

The job here is smaller than it sounds: watch which quotes have gone out, notice which have had no response after a set number of days, draft a short follow-up in the owner voice, and put it in front of them to send. It also notices when a quote was accepted and nobody started the paperwork.

The arithmetic is different for this one, because the saving is not time, it is the quotes that were never chased. You already know what your average job is worth. You almost certainly do not know how many quotes went cold last year, and finding that number out is the honest first step, before anybody builds anything.

The accountancy practice and the document chase

January is a document chase. Which clients have sent what, who has been asked twice, who has been asked and replied with the wrong thing.

An agent that reconciles a checklist against what has actually arrived, and drafts the chaser for the ones missing, replaces a job that is pure administration and entirely rule-driven. And it hits the four tests below cleanly: it happens constantly, the rules are writable down, the data is in one system, and the worst case if it gets something wrong is a client being asked for a document they already sent. Embarrassing, not dangerous.

That last point is why this example is the safest of the three, and safety is a real selection criterion rather than a footnote.

Agent washing, and why the word is now nearly meaningless

Gartner has a term for what is happening to the label: agent washing. Rebranding existing products, assistants, robotic process automation, ordinary chatbots, as agentic without any of the underlying capability changing.

In the press release where it coined the phrase, published in June 2025, Gartner estimated that of the thousands of vendors claiming agentic AI, only around 130 were the real thing.

That is not a reason to dismiss the field. It is a reason to ask one question of any supplier: what does it do without a person in the loop, and what happens when it gets that wrong? A product that cannot answer the first half is an assistant with a new badge on it.

The failure numbers, and what survives reading them properly

Two figures get quoted constantly. Both are worth knowing and both are worth knowing the shape of.

Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Worth noting what that is: a prediction from an analyst firm, with a stated basis of a January 2025 poll of 3,412 webinar attendees about investment posture. It is a considered forecast about intentions. It is not a measurement of outcomes, and it should not be quoted as one.

MIT Project NANDA reported that roughly 95% of enterprise generative AI pilots showed no measurable effect on profit and loss. The preliminary findings drew on more than 300 publicly disclosed initiatives, interviews with 52 organisations and survey responses from 153 senior leaders, gathered between January and June 2025. It is explicitly preliminary, it has not been peer reviewed, and it has been criticised on fair grounds: a narrow definition of success, a roughly six month measurement window, and a small interview base.

So treat both as directional rather than precise. What survives every one of those caveats is the direction, and the direction is consistent with what anybody who has shipped software already suspects: the outcome is decided by scope, not by the model. Projects fail because nobody could say what the thing was for, what it would replace, or how anybody would know it had worked.

The four questions

Test any candidate against these before anybody writes a line of code.

Does it happen often enough to matter? Something you do twice a month is not worth automating, however irritating it is. Twelve times a day is a different conversation. This is where the arithmetic above earns its keep: minutes, times frequency, times working weeks. Do it before you get attached to the idea.

Could you write down the rules? If you cannot describe the decision to a new member of staff in a paragraph, an agent will not infer it either. It will produce something confident and wrong, which is worse than producing nothing, because confident and wrong gets acted on.

Is the data somewhere it can actually reach? This is where most attempts quietly die. If the information lives in one person inbox and a spreadsheet on a desktop, the project is a data project wearing an AI costume, and it will cost several times what anybody budgeted. Find this out in week one, not month three.

What is the worst thing it could do unsupervised? If the answer involves money leaving the business, a customer being told something wrong, or personal data going somewhere it should not, the agent needs a person in the loop at that step. Permanently, not just during the trial. Notice that in the dental example above, the human approval was not a limitation of the build. It was the design.

What actually drives the cost

People assume the model is the expensive part. It is usually the cheapest line on the invoice.

The cost is in the joins. Getting clean, reliable access to the system that holds the diary, the one that holds the client records, and the inbox the enquiries arrive in. If those three things talk to each other today, a narrow agent is a modest piece of work. If they do not, you are paying for integration, and the AI is the small bit on the end.

This is why the published cost guides are so alarming and so useless: they are written for enterprises integrating a dozen systems under a governance regime, and quoting those figures at a practice with eight staff tells you nothing except that somebody wants an enterprise budget. Ask any supplier to price the integration separately from the intelligence. If they will not, that is your answer.

The gap nobody in this conversation mentions

Return to that DSIT survey, because it contains one more number and it is the one that should worry you.

Of the UK businesses using AI, only 17% had a formal policy governing it. Among sole traders, 3%.

Set that beside the most common use case: staff pasting information into a chat box to summarise it, draft a reply from it, or work out what it means. In most businesses nobody has ever said which information may go in there. So customer enquiries go in. Supplier contracts go in. Occasionally patient or client details go in.

Under UK GDPR you remain responsible for personal data you are the controller of, including when a member of staff puts it into a service you have never assessed. A one page policy naming which tools are approved, what may never be pasted into them, and who to ask when someone is unsure, costs an afternoon. It is the cheapest thing on this entire list and almost nobody has done it.

If you are going to move from a chat box to an agent, do this first. An agent has broader access than a person with a browser tab, which means the same unanswered question gets a much larger answer. It is the same reasoning behind handling enquiry forms properly rather than bolting something on: the failure is never in the clever part, it is in the plumbing nobody looked at.

When the answer is not an agent

Sometimes the honest recommendation is that you do not need one, and you will not hear it often from people selling them.

If the job runs on a rule rather than a judgement, write the rule. A filter, a form that routes itself, a scheduled report. It is cheaper, it does not surprise you, and it does not need watching. A surprising share of what gets pitched as an agent is an if-statement with a marketing budget.

If the job is genuinely occasional, do it by hand and spend the money on something that happens every day.

If the underlying process is broken, an agent will run the broken process faster. That is worse than doing nothing, because it also removes the friction that was telling you something was wrong.

And if the data is not reachable, fix that first. It is the less interesting project and it is the one that decides whether anything after it works.

What to do this week

Write the policy. One page, three questions answered: which tools are approved, what must never go into them, and who to ask when somebody is unsure. This is the only item on the list that costs nothing and protects you immediately.

Then run the sum. Take the most repetitive thing anybody in your business does. Minutes, times how often, times 46 weeks. If the answer is not tens of hours a year, pick something else.

Then take it through the four questions, and be honest about the fourth one.

Then build only that. Not a platform, not a transformation. One job that works, that a person still checks, and that pays for itself in something you can point at. The evidence from the deployments that worked is that expansion came from something already working. The evidence from the ones that failed is that it came from a slide deck.

If you want a hand working out which job is worth it, or you would rather be told plainly that a rule and a filter would do the same thing for a fraction of the money, that conversation is free. We would rather lose the work than build you something clever that nobody needed.