How to choose an AI consultancy in the UK (and when to build in-house)

Choosing a UK AI consultancy comes down to scoring six things honestly: shipped track record, security certifications, sector fit, pricing transparency, whether they deliver or only advise, and how candid they are about the cost of doing it in-house instead. Ask to see one AI workflow they run in their own business. The firm that shows you working software, not a slide deck, is usually the one worth hiring.

This guide gives you a scorecard to do that, the questions most consultancies quietly avoid, and an honest read on when you should skip the consultancy and build a team yourself. Most advice on this topic comes from US or offshore firms using US salary data and US assumptions. The UK market is its own thing: a smaller talent pool, IR35, and day rates that swing with London and regional cost. Everything below is UK-specific.

How do I choose an AI consultancy in the UK?

Score every firm on six dimensions and weight them by what your project actually needs. A regulated business weights certifications and sector fit heavily; a fast proof of concept weights delivery and pricing transparency. Use the scorecard, then make the decision deliberately rather than on the strength of the pitch.

The point of scoring is to separate the firm that can talk about AI from the firm that has shipped it. Most can do the first. Fewer can do the second, and fewer still will be straight with you about when you would be better off not hiring them at all.

The AI consultancy evaluation scorecard

A firm that scores well on the first two rows and badly on the last two is a common trap: technically credible, commercially opaque, and quietly planning to keep you dependent. Weight delivery and pricing transparency higher than you might expect, because those are the rows that decide whether the project finishes on budget and whether you can leave when you want to.

What should I ask an AI vendor before signing?

Ask the questions that are hard to answer with marketing language. The single most useful one: show me one AI workflow you have shipped in your own business. A firm that automates other people’s work but runs none of it internally is selling a thing it does not use, and that gap tends to show up in the delivery.

These are the questions most consultancies would rather you did not ask, and the reason each one matters.

• Show me one AI workflow you run in your own business. Tests whether they practise what they sell, not just whether they can build a demo.
• Who actually does the build? Many firms put senior names on the proposal and hand delivery to juniors or subcontractors. You want the pitch team and the build team to be the same people.
• What did your last three handovers look like? A specific answer means they have done it; a vague one means your knowledge transfer is an afterthought.
• Which security certifications do you hold, and where is our data processed? For UK data you want ISO 27001 and Cyber Essentials and a clear answer on residency, not a US framework cited out of habit.
• What happens if this does not work? A firm willing to scope a small first step, and to tell you when AI is the wrong tool, is more useful than one that says yes to everything.

If the answers to those five are clean and specific, the rest of the decision is usually about fit rather than capability. If they are evasive on any of them, that is your answer.

When does in-house actually win?

In-house wins when AI is the product you sell rather than a tool that supports your operations, when you need daily model iteration, when you already have the technical leadership to manage specialists, and when your data is in a state someone can actually build on. Trying to force these situations through a consultancy creates more friction than it removes.

There are five conditions where building a team is clearly the right move.

• AI is your core product. You need staff who live inside the problem domain every day, not a partner who leaves when the statement of work ends.
• You need continuous model iteration. Retraining on new data and adjusting outputs against user feedback in near real time does not fit a project-shaped engagement.
• You are hiring at scale. If you plan to bring on five or more AI engineers within two years, start now, because assembling a team that size in a competitive UK market takes considerably longer than the hiring plan usually assumes.
• You have the management capacity. An AI team without an AI-literate technical lead produces specialist work that nobody senior can evaluate, which shows up as delayed projects rather than a line in the salary budget.
• Your data and infrastructure are ready. If your data is scattered across spreadsheets and legacy databases with no API access, your first hire spends their opening months on data engineering, which is expensive work to buy at an AI engineer’s salary.

The honest version of the cost comparison matters here, because the salary line is not the real cost.

What an in-house AI hire actually costs in the UK

We have deliberately left figures off that table. Salaries and day rates move with seniority, specialism, and whether you are hiring into London, and a number broad enough to be honest is too broad to plan a budget against. For what does move the figure, see our AI development cost guide. The headline is that neither route is cheaper in the abstract. A hire builds ongoing capability that compounds; a consultancy delivers a bounded project faster. Confusing the two is how companies end up disappointed with either choice.

When a consultancy wins

A consultancy is the better answer when the problem is bounded, when you need proof before you can justify headcount, and when the skill is too niche to keep busy full time. Hiring permanently for a six-month problem means paying for the role long after the work is done.

• The project has a clear scope and end date. A retrieval system over your knowledge base or an AI agent that automates one workflow is a deliverable, not an open-ended research programme.
• You need production-quality output before the board will approve headcount. A consultancy delivers that proof point; a recruitment process delivers a candidate in six months and a system some time after.
• The skill is niche and not needed full time. Computer vision, voice AI, or compliance-grade systems command premium salaries you do not want to carry for the nine months you are not using them.
• Your compliance requirements demand certified processes. A consultancy that already holds ISO 27001 and ISO 9001 certifications has the controls and the evidence trail in place for its side of the work. Certification does not transfer, so your own scope stays your own, but their processes are what your procurement team will ask to see.

How to keep the scoping honest

The worry behind this question is a fair one. A firm that only earns when you buy its platform will find a reason for you to buy its platform. The protection is not to split the work across two suppliers, which mostly buys you a translation problem between them and a second firm arguing with the first one’s assumptions. What protects you is insisting the assessment is builder-neutral: that it weighs buying, building, and doing nothing on the same terms, and puts the reasoning somewhere you can challenge it line by line.

Continuity on the other side of that assessment is worth having rather than avoiding. OpenKit becomes your business’s embedded AI team: the people who ran your audit stay on to deliver your 12 month roadmap, so the assumptions in the report carry into the system instead of being re-derived by a team that was never in the room.

The failure mode we actually see is a different one: a single firm scopes, builds, and owns the whole stack, and the client has nobody internally who understands any of it. That is dependency rather than partnership, and the fix is cheap. Insist on a documented handover as a line item in the statement of work, and make sure at least one of your people learns the system while it is being built.

The hybrid model: consultancy builds, your team scales

Most UK firms that succeed with AI long term do not pick one model and stick with it. They sequence them: a consultancy builds the first system, transfers knowledge deliberately, and the client’s first in-house engineer learns that system while it is being built rather than inheriting it cold.

A good handover is documented architecture decisions, runbooks for common operations, a recorded walkthrough of the code, and at least two weeks of paired working between the consultancy and your engineer. The gold standard is a consultancy that writes itself out of the engagement and leaves your team self-sufficient. Ask any prospective partner what their last three handovers looked like; if they cannot answer specifically, that tells you what kind of relationship you would be buying.

IR35 and the contractor question

If you are weighing an individual AI contractor against a consultancy firm, IR35 status has to be assessed for the contractor route. Since the 2021 off-payroll rules, the responsibility for determining employment status sits with the hiring organisation for medium and large businesses, which adds administrative overhead and financial risk to engaging a sole contractor.

A consultancy engagement is usually written as a B2B service delivering a defined outcome rather than a person supplied to work under your direction, and that structure is what ordinarily keeps it outside the off-payroll rules. Status turns on the contract and on how the work is actually carried out, not on the label at the top of the invoice, so an arrangement that in practice looks like a member of your staff will be treated as one whatever it says. If you do engage an individual contractor, get your legal team involved before the work starts rather than after HMRC asks questions.

For what an AI consulting engagement actually involves, see our guide to AI consulting for UK businesses. For a sense of what OpenKit has shipped, the Rubrical education AI and EMQN healthcare assessment case studies are the closest thing to “show me what you built.”