"AI agency" is the label for at least five different kinds of business, and most owners who type it into Google want the one that builds working automations for a company their size.
That one is usually called an AI automation agency. The others are a marketing agency that uses AI tools, a custom software or machine learning firm, a specialist in AI search visibility, and a course seller teaching people how to start their own agency.
The term gets around 9,900 UK searches a month, and the results page reflects the confusion. You get directory listings, a hardware vendor's glossary entry, a "how to start an AI agency" guide and a handful of service pages, all for the same phrase.
This guide sorts them out. It covers what each type does, what an AI automation agency does week to week, how to judge one, and how the work is usually priced.
Why are the search results for "AI agency" such a mixed bag?
The results are mixed because three different groups type the same two words: buyers, would-be agency owners and people researching a job or a supplier.
Buyers want a supplier. Would-be owners want a business model, which is why a "how to start an AI agency" guide ranks beside service pages. Researchers want a definition, which is how a hardware vendor's glossary page ends up on the first screen.
When we checked the first page of UK results in October 2026, the formats ran from a London service page with a client logo strip to a directory of "top AI companies" to a glossary entry that defined an AI agency as, among other things, a government body.
None of them said what a small business should expect to pay or what happens after the build.
That gap is what this guide tries to fill, so you can decide which of the five you are looking for before you start phoning suppliers.
Which kind of AI agency do you actually need?
It depends on the problem you are trying to fix, and the five types solve very different ones.
| Type | What you get | Best fit | Check before signing |
|---|---|---|---|
| AI automation agency | Workflows, chatbots and agents wired into your existing tools | Teams losing hours to admin, follow-up and data entry | Who owns the accounts and who fixes it when it breaks |
| Marketing agency using AI | Ads, SEO and content, with AI inside the production process | Businesses that want leads, not software | Whether a person edits and signs off everything published |
| Custom AI or software firm | Bespoke models, apps or internal tools | Companies with a specific product or data problem | Whether a ready-made tool already does 80% of it |
| AI visibility specialist | Work to get your brand named in ChatGPT, Gemini and AI Overviews | Firms that lose enquiries to AI answers | What they measure, and how often |
| Agency-starter course | A programme for people who want to sell automation | Individuals, not businesses | Whether the person teaching has delivered client work |
If you run a business and want hours back, the first row is probably the one. If you want more enquiries, you may be after the second or fourth, and our guide to choosing a digital marketing agency in the UK covers the first of those.
AI visibility has its own evidence base. We wrote up what 45 published GEO studies actually found so you can see which claims hold up.
What does an AI automation agency do week to week?
It finds a repeated task in your business, builds a workflow that does most of it, puts a person at the points where a mistake would cost you, and then watches it run.
The first step is the least glamorous. Someone sits with the people doing the job and writes down every step, every exception and every place a decision gets made. Most of the value comes from this, because a badly understood process automated is a badly understood process running faster.
The build is usually a mix of three things:
- a workflow tool such as Make.com, Zapier or n8n that moves data between your systems (we compare them in Make vs Zapier vs n8n for a UK SME)
- a language model that reads, classifies or drafts text inside a step
- a few lines of custom code where the off-the-shelf tools run out
Then comes the part many proposals skip, which is monitoring. Automations fail quietly. A form field gets renamed, an API key expires, a model starts returning a different format. Somebody has to be told when that happens and have the access to fix it.
A good agency can show you its failure log from a live client, with the client's name removed, before it asks for your budget.
Which jobs are worth automating first?
The jobs worth automating first are the ones that repeat daily, follow clear rules and sit between two systems you already pay for.
Typical examples are lead routing, quote follow-up, invoice chasing, appointment reminders, document intake and weekly reporting. We list the five that save the most hours in the five automations that save a small business the most hours, and none of them is a chatbot.
Speed matters more than people expect on the sales side. If an enquiry waits an hour for a reply, a competitor may already have answered, which is why we wrote a separate piece on answering new leads in under two minutes.
For clinics, the equivalent is a missed call. A patient who rings a practice at 6pm and gets voicemail will often ring the next one on the list, so healthcare practices tend to get quick value from automated call-back and booking flows.
Chatbots belong on the list but not at the top. They convert in some settings and annoy visitors in others, and when AI chatbots earn their place on a website is worth reading before you buy one.
If you do not yet know where your own time goes, a one-page AI audit is the cheapest way to find out. It takes an afternoon and needs no software.
Is anyone using AI agents properly yet?
Some are, but the label is being stretched, and the figures from the UK government and from analysts both say to be careful.
The Office for National Statistics reports that about 35% of UK businesses were using at least one AI technology in June 2026, almost three times the share in late 2023.
Take-up differs by size: 28% of businesses with 0 to 9 employees use some form of AI, against 49% of those with 250 or more, according to the ONS analysis of AI in UK businesses.
xychart-beta
title "UK businesses using at least one AI technology, by size, June 2026 (ONS, %)"
x-axis ["0 to 9 employees", "250+ employees"]
y-axis "Per cent of businesses" 0 --> 60
bar [28, 49]
The ONS also found that most use is general: large language models were used by 18% of businesses and visual content tools by 16%. Using ChatGPT to draft an email is AI use. It is not an agent, and it does not need an agency.
On the supply side, Gartner forecast in 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear value and weak risk controls.
The same analyst note described "agent washing", where vendors relabel chatbots, assistants and robotic process automation as agents. Gartner estimated that only about 130 of the thousands of vendors using the word had real agentic capability.
So when a proposal says "AI agents", ask what the agent can actually decide, what it is allowed to do without approval, and what happens when it is wrong. Agentic workflows for a sales team walks through one worked example, from enquiry to proposal, with the approval points marked.
How do you tell a good AI agency from a convincing demo?
A good one shows you something running in production, says what it does not do, and hands you the keys.
Here are the questions we would put to any agency, including ourselves:
- Can I see a workflow that has run for at least three months, with its error log?
- Who owns the automation accounts, API keys and code at the end of the contract?
- At which step does a person review the output before it reaches a customer?
- Where does my customer data go, which model provider sees it, and what is your legal basis for that?
- What would you tell me not to automate?
- What does it cost to run each month, separate from your fee?
The fourth question trips up more agencies than any other. Putting names, emails or case details into a public AI tool is a data protection decision under UK GDPR, and your agency should be able to explain it without hunting for a lawyer.
We set out what you can and cannot feed into ChatGPT about your customers in a separate guide.
The fifth question is a character test. An agency that cannot name a single thing it would leave alone is selling, not advising.
If your firm handles sensitive files, such as accountants and other professional services firms do, put the second and fourth questions in writing before any work starts.
How is AI automation priced?
Most AI automation work is priced as a one-off build fee plus a monthly retainer for monitoring, with the software subscriptions billed separately at cost.
That structure matters because the running costs are not small and they are not the agency's. Workflow tools charge by operation or by run, and language models charge by the amount of text processed. A workflow that handles 50 enquiries a month and one that handles 5,000 have very different bills even if the build was identical.
Ask for the monthly run cost to be estimated at your expected volume, then at three times that volume. If the answer is vague, the agency has not built one of these for a business with real traffic.
There are two other pricing questions worth settling early.
The first is whether you need a custom build at all. For many small firms a monthly subscription does the job, and build vs buy for AI tools explains how to decide.
The second is what a custom build now costs. AI coding tools have changed the maths on bespoke software, and the end of the £20k custom web app quote covers what we think has shifted and what has not.
We quote the build as a fixed fee and the monitoring as a monthly retainer, and we pass the tool subscriptions through at cost.
When you do not need an AI agency
You do not need one if the task happens a few times a month, if the process is not written down, or if a ready-made feature in software you already own covers it.
Many CRMs, accounting packages and booking tools now have their own automation and AI features. Switch those on first. If they handle 80% of the job, the remaining 20% rarely justifies a bespoke build.
You also do not need an agency to produce content at scale. Google's own guidance on creating helpful, reliable content asks whether a page was made primarily for people or for search engines, and a flood of lightly edited AI posts fails that test.
Our view of where AI belongs in a content process, and where it does not, is in why AI content at scale is a liability.
And you do not need one before you have a clear problem. Agencies are good at building to a brief and weaker at inventing your priorities for you.
How Growthcart approaches AI work
Growthcart builds workflow automation, chatbots and custom AI tools for UK businesses, and we start with an audit of where your hours go before we propose anything.
Our own build stack is Next.js, Supabase, Make.com and Claude Code, and we say that up front because it affects who can maintain the work after we leave. We hand over the accounts and the documentation at the end of every project, so the automation does not depend on us.
You can see the full scope on our AI automation services page. If you would like a view on whether your own process is worth automating, or whether you should simply switch on a feature you already pay for, get in touch and tell us what takes the most time in your week.
