How to Evaluate an AI Consultancy Before You Sign a Contract

The short version: before you sign anything, ask an AI consultancy for a named case study with numbers, a fixed scope tied to a deliverable rather than hours, and a client reference from at least eighteen months ago, not last month. If they can't give you all three without hesitating, walk away. Most of the damage I've seen clients suffer from bad AI hires happened in the two weeks before the contract was signed, not after.

Why this decision gets rushed

I've watched business owners spend more time choosing an office chair than choosing the consultancy they're about to hand £15,000 and three months of momentum to. Part of it is fear of missing out. Every week there's a new headline about a competitor "using AI" and a founder panics and signs the first proposal that lands in their inbox. Part of it is that AI consultancies are still new enough that there's no obvious checklist, no equivalent of checking a solicitor's SRA number or an accountant's ICAEW membership.

There isn't a licensing body for AI consultants. Anyone can print a business card that says "AI Strategist" three weeks after finishing a weekend course. That's not a reason to avoid hiring one, it's a reason to do the checking yourself, before the deposit invoice goes out.

A story that still bothers me

Last year a manufacturing client of mine had already signed with another firm before she called me. She'd paid £45,000 for a "90-day AI transformation" that promised three things: an internal chatbot for staff, an automated quoting tool for sales, and a training day for her team. The contract ran twelve weeks. By week ten, one deliverable existed, a chatbot demo that lived on a test server nobody had access to. No quoting tool. No training day booked. When she chased, the account manager had left the company and nobody replied for eleven days.

What made it worse wasn't the money, it was that the contract had no milestone payments. She'd paid 70% upfront because the sales rep said "that's standard for AI projects, the models need setup time." It isn't standard. It's a way to get paid before the risk of underdelivery lands on the consultancy instead of the client. When I finally got the quoting tool built (in four weeks, for a fraction of the original fee), it turned out the "AI" in the original scope was a rebadged version of a template chatbot builder with no custom training on her actual product catalogue. She'd paid enterprise money for a template.

I tell this story often because it's not rare. It's close to the median experience for a first-time AI hire made under time pressure.

Start with the case study, not the pitch deck

Any consultancy worth hiring should be able to name at least one client (with permission, or anonymised with real numbers) and tell you exactly what changed. Not "improved efficiency." A number. Hours saved per week, cost per lead before and after, response time reduced from four hours to twelve minutes, whatever it is for their work.

Ask these three questions and listen closely to how specific the answers are:

• What was the business doing before you started, in plain terms?
• What exactly did you build or change, and how long did it take?
• What number moved, and how do you know it moved because of your work and not something else happening in the business at the same time?

If the answer to the last question is vague, that's not automatically dishonest. Attribution is hard in marketing and operations work. But a decent consultancy will admit that complexity rather than claim clean credit for every improvement.

Ask for a reference nobody would normally give you

Here's the part most guides on this topic skip, and it's the one I'd push hardest on. A happy client reference tells you almost nothing. Every consultancy has one client who loved them, that's survivorship bias, not evidence. What tells you far more is asking for a client from eighteen months to two years ago, someone whose project has had time to either hold up or quietly fall apart.

Ask that older client one question: "Is the thing they built still running, and is anyone still using it?" I've had consultancies refuse this request outright, which is itself useful information. The ones worth working with usually say something like "sure, call Dave, though fair warning, he'll tell you the first phase was rocky." That kind of honesty from a client, relayed by the consultancy itself, is worth more than five polished testimonials on a website.

If you want a wider set of the questions buyers ask before hiring, I've put together a longer reference guide covering real buyer questions on AI consulting that goes beyond just references and into scope, tools, and data ownership.

Check the badge, then ignore most of it

Plenty of consultancies now list themselves as "Microsoft Certified Partner" or "OpenAI verified" or similar. Here's the uncomfortable bit nobody selling AI services wants to say out loud: those badges mostly mean the firm paid a fee, completed an online course, and hit a sales quota with that vendor. They tell you almost nothing about whether that person can solve your specific problem. I know consultants with three vendor certifications who've never shipped a working automation for a business with under fifty staff, and I know brilliant, scrappy operators with zero badges who've quietly saved clients six figures in labour costs.

Use certifications as a tiebreaker, never as the deciding factor. What predicts good work is whether they can describe, in ordinary language, how a specific AI tool would handle a specific mess in your business, right there in the sales call, before you've paid them a penny.

Understand exactly who does the work

This is the question that catches people out most often. You take a call with a confident, experienced consultant, sign the contract, and then a junior team member you've never spoken to shows up to run the actual project. That's not automatically bad, junior staff can be excellent, but you need to know it's happening.

Ask directly: "Who will be doing the day-to-day work, and can I meet them before I sign?" If the answer is dodged or delayed, treat that as a red flag on par with a vague price.

The pricing conversation you need to have

UK AI consultancy pricing in 2026 tends to sit in fairly consistent bands: freelance specialists around £600 to £1,200 a day, established boutique consultancies around £1,500 to £3,000 a day or £4,000 to £12,000 a month on retainer, and larger firms or fractional AI leadership roles running £8,000 to £20,000 a month depending on scope. If you're being quoted well outside those ranges in either direction, ask why. Too cheap often means templated, low-customisation work. Too expensive without a named senior person attached often means you're paying for brand, not delivery. I've written a fuller breakdown of what drives those numbers in the honest price bands for AI consultants in 2026, worth reading before any negotiation.

Whatever the headline day rate, the structure of payment matters more than the total. This is the specific step I'd tell every reader to take before signing:

• Split payment into at least three milestones tied to deliverables, not dates. For example, 30% on discovery and agreed scope document, 40% on working build delivered for testing, 30% on successful handover and training.
• Never pay more than 30% before any work has been shown to you.
• Ask what happens financially if a milestone slips because of the consultancy, not because of you. A fair contract has a clause for that.

Read the contract for these five things

Most people skim the contract for the price and the date, then sign. Here's what matters:

• IP ownership. Who owns the prompts, workflows, automations, and any custom-trained models once the project ends? You want this written as "client owns all deliverables and underlying configuration" not just "client owns the final report."
• Data handling. Where does your customer data go during the project, which AI tools touch it, and is there a data processing agreement covering GDPR if you're UK or EU based? If they can't answer this in one sentence, that's a problem.
• Exit clause. What notice period lets you leave if the work isn't landing, and what do you keep if you do?
• Named individuals. Does the contract name the actual person doing the work, or only the company? Contracts that name individuals tend to get better delivery because there's personal accountability attached.
• Renewal terms. Does the retainer auto-renew, and how much notice do you need to give to stop it? I've seen 60-day auto-renewal notice clauses buried in appendix three that trap clients for another quarter.

Ask what happens when the AI tool changes

This is a question almost nobody thinks to ask, and it should be standard. AI tools shift fast, a workflow built on one model's API pricing or capability in January can look completely different by June. Ask the consultancy: "If the underlying tool changes its pricing or gets deprecated, is updating the system included in this contract, or is that a new project?" Good consultancies have a clear answer already prepared, because they've been burned by this themselves. Vague answers here usually mean you'll be quoted for "urgent maintenance" six months in.

Watch how they talk about failure

Every consultancy that's done real work has a project that didn't go to plan. If someone tells you, without prompting, "our AI projects have a 100% success rate," that's not confidence, it's either inexperience or dishonesty. Ask them directly: "Tell me about a project that didn't work the way you expected." The answer tells you more about how they'll handle your inevitable mid-project problems than any case study will.

Match the consultancy to your actual size and market

A consultancy brilliant at rolling out enterprise AI governance for a 2,000-person company is often the wrong fit for a 12-person agency needing a working lead-qualification workflow by next month, and the reverse is just as true. Check they've worked at your scale recently, not five years ago before their client list grew up. If you're based outside the UK, region matters too, the regulatory and tooling landscape is different enough that I've written separate hiring guides for American founders hiring an AI marketing consultant and for Canadian founders doing the same, because pricing norms and privacy law expectations aren't identical across borders.

If what you need is ongoing senior marketing direction with AI woven through it rather than a single project, it's worth reading about the difference between hiring a project-based consultancy and bringing in a fractional CMO for the AI era, since the evaluation criteria shift once you're hiring for leadership rather than delivery.

A short evaluation process you can run in a week

Here's the process I'd run, condensed to five days:

• Day 1: request a written scope, a case study with numbers, and a reference from 18+ months ago.
• Day 2: call the older reference and ask if the work still runs.
• Day 3: get the pricing structure in writing, milestone by milestone, and check it against typical UK bands.
• Day 4: ask to meet the actual person doing the work, not just the salesperson.
• Day 5: read the contract for IP, data, exit, and renewal clauses before signing.

Five working days feels slow when you're keen to move, but it's roughly the time it takes to avoid a £45,000 mistake like the one I described earlier. If you want a broader productivity framework for judging any business tool or template before you adopt it, not just AI, the same instinct applies in what to look for in a productivity template for your business, the underlying question is always the same: does this fit how we work, or does it just look impressive in a demo?

If, after going through all of this, you'd rather have someone independent sanity-check a proposal before you sign it, that's exactly the kind of second opinion an AI consultant for small business gets asked to do fairly often, and it's usually a cheap hour compared with the cost of the mistake it prevents.

Two pages that follow on from this: hiring an AI consultant (what to ask, what to expect) and how to find an AI consultant worth hiring (what the work looks like week by week).

Your sector is probably covered: what to automate first, by industry.

Related: ai consultant handover checklist.

Related: ai consultant discovery call what to prepare.

Frequently asked questions

What's the biggest red flag when evaluating an AI consultancy?

Asking for more than 30% payment upfront with no milestone attached is the single biggest red flag, followed closely by vague answers about who will do the work and refusal to provide an older client reference.

How much should I expect to pay an AI consultancy in the UK in 2026?

Freelance specialists typically charge £600 to £1,200 a day, boutique consultancies £1,500 to £3,000 a day or £4,000 to £12,000 a month on retainer, and larger firms or fractional AI leadership roles £8,000 to £20,000 a month, depending on scope and seniority.

Should I trust vendor certifications like Microsoft or OpenAI partner status?

Treat certifications as a minor tiebreaker, not proof of skill, since they mostly confirm a fee was paid and a course completed, not that the consultancy has delivered work for a business like yours.

What should be in the contract before I sign with an AI consultancy?

Look for milestone-based payment, clear IP ownership of everything built, a data handling clause covering which tools touch your customer data, a reasonable exit clause, and named individuals responsible for delivery rather than just the company name.

Official documentation

• Anthropic's own documentation
• the NIST AI Risk Management Framework