What should I ask an AI agency before signing a contract?

Ask what problem the work will solve in your numbers, what happens when the AI is wrong, who owns the code, prompts and data afterwards, what the system costs to run as well as build, where your data goes, and what the smallest first version would be. An agency that answers all of these in plain English, in writing, is worth shortlisting.

What are the red flags when choosing an AI consultant in the UK?

Guaranteed returns quoted before discovery, marketing led by model names rather than outcomes, a chatbot offered as the answer to every problem, proprietary platforms you cannot export from, accounts held in the agency's name, and pressure to sign a long retainer before any small project has proven the relationship. Any one of these should slow the conversation down.

Should a small business buy off-the-shelf AI tools or commission a custom build?

Do both, in order. Subscribe to cheap generic tools for tasks like drafting and meeting notes first. Commission a build only for the specific process where off-the-shelf tools clearly do not fit, such as a bespoke booking flow or tender pipeline. Government adoption research shows most UK firms buy external AI software rather than developing in-house, and that is usually the right starting point.

Who should own the code, prompts and data after an AI project ends?

You should. A well-structured engagement leaves the client owning the code, the prompts, the documentation and every third-party account, with the data exportable at any time. Put ownership in the contract before work starts. If an agency treats prompts as trade secrets or holds accounts in its own name, you are buying lock-in rather than a system.

How do I know whether an AI agency has really built what it claims?

Ask to see a comparable system running live, even anonymised, and ask the team to talk about failure modes and edge cases. Builders who have shipped real systems talk readily about the unglamorous parts: error handling, deduplication, attribution, escalation to humans. Silence on what went wrong in past projects usually means nothing real was shipped.

Why do so many AI projects fail?

RAND research puts the failure rate at more than 80% by some estimates, twice that of ordinary IT projects. The leading root cause is misunderstanding or miscommunicating which problem the AI is meant to solve, not the technology itself. Choosing an agency that insists on diagnosing your process before proposing anything removes the most common cause of failure before any money is spent.

How to choose an AI agency or consultant in the UK

An honest decision guide for UK SMBs: the ten questions to ask an AI agency before signing, the red flags that should end the conversation, and when to subscribe, commission a build or hire in-house instead.

Last updated 10 June 20268 min read

Choose a UK AI agency by asking it to show you a working system, not a slide deck. A good agency starts with a specific business problem, names the process it will change, explains who owns the prompts, the data and the code when the contract ends, and prices the work against a measurable outcome. Avoid any firm that leads with model names, promises results before it has seen your data, or cannot tell you what happens when the AI gets something wrong. Most UK small and mid-sized businesses do not need a transformation programme. They need one well-built system around one process, proven, then extended.

"Half the AI projects we get asked to quote should not exist, and a decent agency will tell you so. Ask to see a working system, not a slide deck, and watch what happens to the conversation."

Dean Cookson, founder, Operosus

Why does choosing well matter more than choosing fast?

Because the market has filled with sellers faster than it has filled with builders. AI adoption among UK SMEs reached 54% in 2026, up from 35% the year before, according to the British Chambers of Commerce's Future of Work research with Atos. As Patrick Milnes, the BCC's Head of Policy for People and Work, put it: "AI has rapidly moved from the margins of business to the mainstream." (The adoption numbers behind this guide are kept sourced and current in our UK small business AI statistics table.)

Mainstream adoption has not made buying easier. The government's own AI adoption research, based on 3,500 business interviews conducted by IFF Research and Technopolis Group, found that 71% of non-adopters see no identified need for AI and 60% cite limited skills. That is the gap an agency is supposed to close: turning "we know we should be doing something" into a system that does something.

And the cost of choosing badly is well documented. RAND's research into failed AI programmes found that by some estimates more than 80% of AI projects fail, twice the failure rate of IT projects that do not involve AI. The leading root cause was not the technology. It was stakeholders misunderstanding, or miscommunicating, which problem the AI was meant to solve. That is a vendor-selection failure as much as a technical one, and it is avoidable at the point of purchase.

What should an AI agency actually do for you?

Strip away the branding and a competent AI agency does four things:

Diagnoses before it prescribes. It maps how a process runs today, where time and money leak, and where judgement genuinely matters, before proposing anything.
Builds systems, not demos. A real system handles the unhappy paths: malformed form submissions, duplicate records, a model returning nonsense, an API being down. A demo handles the happy path in a screen share.
Keeps humans where humans belong. Good builds put AI on the repetitive middle of a process and keep people on the judgement at either end, approving, correcting and handling exceptions.
Hands you the keys. You should end the engagement owning the accounts, the data, the prompts and the documentation, with the option to walk away.

If a prospective agency cannot describe its last three projects in those terms, it is selling software with a services badge on it.

What questions should you ask before you sign?

Ask all of these, in writing if needed. The answers matter less than whether the agency can answer them at all.

What problem will this solve, in our numbers? A serious answer references your volumes: enquiries per week, tenders per quarter, invoices outstanding. A vague answer references "efficiency".
Can we see a comparable system running? Not a case study PDF. A screen share of a live build, even anonymised.
What happens when the AI is wrong? Every system built on a language model will sometimes be wrong. You want to hear about review steps, confidence thresholds, escalation to a human and logging, not reassurance.
Who owns what when we part ways? Code, prompts, data, third-party accounts, documentation. Get it in the contract.
What will this cost to run, not just to build? Model usage, software subscriptions, hosting and maintenance continue after the invoice is paid.
Where does our data go? Which providers see it, under what terms, and whether anything is used for model training. If you handle personal data, ask how the build meets UK GDPR and who the processors are.
What is the smallest version of this you would build first? Good agencies want a narrow first slice they can prove. Bad ones want the full scope signed up front.
How will we measure success? Agree the metric before the build starts: hours saved, response time, conversion rate, error rate. If it cannot be measured, it cannot be defended at renewal.
Who actually does the work? Named people, not "our delivery team". Ask who you will speak to when something breaks at 9am on a Monday.
What would make you advise us not to do this? An honest agency has turned work away and can tell you why. A pure sales operation has never met a project it didn't like.

What are the red flags?

Walk away, or at least slow down, when you see these:

Model-name marketing. Leading with "powered by GPT-5" or "built on Claude" instead of the business outcome. The model is a component, not the product.
Guaranteed numbers before discovery. Specific promised returns, multiples or percentage savings quoted before anyone has looked at your data are invented. Treat them as a character reference.
No questions about your process. If the first meeting is all about their platform and not about how your business actually works, the proposal was written before you arrived.
Chatbot-first thinking. A chatbot is occasionally the right answer. When it is the only answer an agency ever gives, you are buying their template, not a solution to your problem.
Lock-in by design. Proprietary platforms you cannot export from, accounts held in the agency's name, prompts treated as trade secrets you may not see.
No talk of failure modes. Any builder who has shipped real AI systems has war stories about edge cases. Silence on the subject means they have not shipped, or will not admit what happened when they did.
Pressure to sign a long retainer up front. The economics of a good build mean the agency should be happy to prove itself on a small first project.

Should you buy a subscription, commission a build, or hire in-house?

These are the three realistic routes for a UK SMB, and most businesses end up blending them. Notably, most UK firms already buy rather than build: in the government's adoption research, 71% of businesses using natural language processing bought external software, against far smaller numbers developing in-house.

	Off-the-shelf subscription	Agency-built system	In-house hire
Best for	Generic tasks: meeting notes, drafting, transcription	A specific process the off-the-shelf tools do not fit	Continuous AI work across many processes
Cost shape	Low monthly fee per seat	Project fee plus modest running costs	Salary, tools and management overhead
Time to value	Days	A first working slice, then iteration	Months, including recruitment
Fit to your process	You adapt to the tool	The system is shaped to your process	Depends entirely on who you hire
Ownership	None, you rent it	Should be full, check the contract	Full, but knowledge leaves with the person
Main risk	Shallow value, tool sprawl	Picking the wrong agency	One person carrying everything

A sensible sequence for most SMBs: subscribe to the cheap generic tools immediately, commission a build for the one process where a generic tool clearly does not fit, and only hire once you have several live systems that need permanent ownership. Hiring first is the expensive way to discover what you should have bought.

How can you tell whether an agency's work is real?

Ask them to talk about the boring parts. Real systems are mostly plumbing, and builders who have done the work talk about it without prompting. Three patterns from our own projects show the shape of it:

Drafting with a human gate. Bidwell, our tender-writing product for UK SMBs, uses AI to produce the draft response, but the structure assumes a person reviews and edits before anything is submitted. The value is in the time saved getting to a reviewable draft, not in pretending the AI can be left alone with a contract bid.
Booking flows that fail soft. For Vets at Home, a home-visit veterinary service, the booking system carries each enquiry from form to CRM to confirmation, with source attribution preserved along the way so the business knows which channel produced which booking. Every form is built to degrade gracefully: if a downstream system is unavailable, the enquiry is still captured rather than lost.
Matching as a system, not a feature. For Vivify, a platform that helps schools hire out their facilities, the work centred on connecting would-be hirers with suitable nearby venues, which is a data and workflow problem long before it is an AI problem.

None of these is glamorous. All of them are the difference between a system a business runs on and a demo that dies after the pilot. When you interview agencies, the ones who light up describing attribution, deduplication and escalation paths are the ones who have shipped.

Where to start

Pick one process, not a strategy. The best first project is narrow, frequent and measurable: lead follow-up, tender drafting, invoice chasing, booking admin. Write down how it works today, how long it takes and what it costs in hours per week. That one page makes you a dramatically better buyer, because every agency conversation now starts from your numbers instead of their pitch.

Then speak to two or three agencies and put the same ten questions to each. You are not looking for the most impressive answer. You are looking for the firm that asks you the most questions back, proposes the smallest credible first build, and is relaxed about you owning everything it makes.

Before any conversation, two pieces of homework will sharpen your questions: our breakdown of what AI automation actually costs in the UK, and how we work, which shows what honest engagement terms look like in plain words. Our own numbers are public on the proof page, because we hold ourselves to question two.

If you want a starting point, we offer a short, no-obligation review of one process and an honest answer on whether it is worth automating at all. Sometimes the right answer is a £30-a-month subscription and no agency involved. An agency that will tell you that is usually the one worth keeping.

Frequently asked questions

What should I ask an AI agency before signing a contract?: Ask what problem the work will solve in your numbers, what happens when the AI is wrong, who owns the code, prompts and data afterwards, what the system costs to run as well as build, where your data goes, and what the smallest first version would be. An agency that answers all of these in plain English, in writing, is worth shortlisting.
What are the red flags when choosing an AI consultant in the UK?: Guaranteed returns quoted before discovery, marketing led by model names rather than outcomes, a chatbot offered as the answer to every problem, proprietary platforms you cannot export from, accounts held in the agency's name, and pressure to sign a long retainer before any small project has proven the relationship. Any one of these should slow the conversation down.
Should a small business buy off-the-shelf AI tools or commission a custom build?: Do both, in order. Subscribe to cheap generic tools for tasks like drafting and meeting notes first. Commission a build only for the specific process where off-the-shelf tools clearly do not fit, such as a bespoke booking flow or tender pipeline. Government adoption research shows most UK firms buy external AI software rather than developing in-house, and that is usually the right starting point.
Who should own the code, prompts and data after an AI project ends?: You should. A well-structured engagement leaves the client owning the code, the prompts, the documentation and every third-party account, with the data exportable at any time. Put ownership in the contract before work starts. If an agency treats prompts as trade secrets or holds accounts in its own name, you are buying lock-in rather than a system.
How do I know whether an AI agency has really built what it claims?: Ask to see a comparable system running live, even anonymised, and ask the team to talk about failure modes and edge cases. Builders who have shipped real systems talk readily about the unglamorous parts: error handling, deduplication, attribution, escalation to humans. Silence on what went wrong in past projects usually means nothing real was shipped.
Why do so many AI projects fail?: RAND research puts the failure rate at more than 80% by some estimates, twice that of ordinary IT projects. The leading root cause is misunderstanding or miscommunicating which problem the AI is meant to solve, not the technology itself. Choosing an agency that insists on diagnosing your process before proposing anything removes the most common cause of failure before any money is spent.