Buyer's guide

How do you choose an AI consultancy? Nine questions to ask before you sign.

Ask who does the work and who owns the result. Those two answers decide whether you end up with a working system or a folder of slides. Below are the nine questions we would ask, each with a good answer, a bad answer, and ours.

  • Nine questions, answered about us as well as about them
  • A twelve-row scorecard you can print and take to three firms
  • The row we fail, written down rather than left out
NineQuestions, each one answered about ourselves
£2,000Our published diagnostic fee, 14 days
£98,214.72One issued job, 63 measured lines, published in full
16 September 2026The date every claim on this page was checked

The short answer

Two answers carry the decision.

Ask who does the work and who owns the result. At AGMM the two founders do it, and the audit's four outputs, the constraint map, the opportunity map, the economic cases and the ranked proposal, are yours whether you build or not. Every other question here tests whether those two answers survive a contract.

If you have ten minutes before a call, take the nine headings below and nothing else. A supplier asked these before answers each in a sentence. One who has not takes a paragraph to say very little.

The problem with buyer's guides

Why is this hard to answer honestly?

Because everyone writing the guide is also selling the service, and the criteria always favour the author's own model.

We read the pages that rank for this question and measured them. All list criteria a buyer should apply. Not one applies those criteria to itself, and several compare costs without a single figure on the page. A criterion you would fail is awkward to publish, so it quietly does not make the list.

We are in the same position, so each of the nine answers below ends with ours, and the scorecard scores us against all twelve criteria including the one we fail. For market context rather than our answer, the price index is a separate page.

Question 01

Who will actually do the work, and will they be in the room?

Why it matters

The person who sells is often not the person who builds. Speed, judgement and the questions asked in week one follow whoever writes the code. If you cannot name that person before you sign, you are buying a process.

A good answer sounds like

Names, what each has built, and a written commitment that they attend the working sessions. A firm using subcontractors says which parts.

A bad answer sounds like

"Our team." A named senior who appears at the pitch and is gone by week two. Any answer deferred until after signature.

Our answer

Samuel Wall and Adam Gillett do the work. The format is exactly this: Three sixty-minute working sessions with the forward deployed AI engineers. There is no account manager between you and the people building the thing.

Question 02

What do I own at the end if I never build?

Why it matters

A diagnostic you cannot take elsewhere is a sales document with a price on it. If you read the findings, decide against building and walk away, what is left in your hands?

A good answer sounds like

A named list of artefacts you can open without their software, yours whether you proceed or not. On a build, source code assigned on delivery.

A bad answer sounds like

"A strategy" with no artefact list. Findings that live in their portal. Ownership left to settle later.

Our answer

Four outputs, yours in every case: a business-wide constraint map, an AI opportunity map, the economic cases with their assumptions written down, and a ranked next step with a fixed-price proposal.

Question 03

Is the price fixed, a day rate, or a retainer, and what does each do to my risk?

Why it matters

The shape of the price decides who carries the overrun. A fixed fee puts it on the supplier, who then scopes hard before starting. A day rate puts it on you, and the incentive runs the wrong way. A retainer buys availability, not an outcome.

A good answer sounds like

The shape named plainly, what changes the number, and what happens when the work runs long.

A bad answer sounds like

A range with no scope at either end. A day rate presented as a price. A retainer before any diagnosis.

Our answer

£2,000 fixed for the diagnostic, 14 days. If it runs longer, it runs longer at our cost. Builds are scoped and priced after the diagnostic, never before.

Question 04

Is the diagnostic fee credited against a build?

Why it matters

This question exposes the incentive. If the fee comes back the moment you sign a build, the diagnostic only pays for itself by recommending one, and every finding has a direction to travel before anyone sees your data.

A good answer sounds like

A straight no with the reason, or a yes that admits what the credit does to the incentive.

A bad answer sounds like

A free audit. It is a sales demonstration, and it has to end in a proposal, because that is how the hours are recovered.

Our answer

No. A free audit is a sales call with a document attached. It has to arrive at a recommendation to build something, because that is the only way it ever pays for itself. Ours is paid, and the fee is not deducted from the build, so we carry no financial reason to recommend building anything. That answer costs us work, and it is why "do not build" is an outcome we can afford to reach.

Question 05

What is not included?

Why it matters

The exclusion list tells you more than the inclusion list, because anybody can write the inclusion list. Most proposals have none, so the boundary gets found during delivery, in an argument.

A good answer sounds like

A written list of what sits outside the fee, and who pays for licences, data clean-up and changes.

A bad answer sounds like

Everything included, nothing excluded. A proposal where "included" appears fourteen times and "not" twice.

Our answer

The diagnostic is not a build, and no production software is written during it. Licences, data clean-up and integration work are named in the findings and priced in the proposal. We do not sell brand work, advertising, or open-ended strategy retainers.

Question 06

What will you show me that I can check myself?

Why it matters

A case study is written by the seller and cannot be audited by you. A checkable artefact exists at an address, and a stranger can open it.

A good answer sounds like

A live URL you can open now. A public record. A client who will take your call.

A bad answer sounds like

A wall of logos. "A UK manufacturer saw a forty per cent reduction." Return multiples with no arithmetic.

Our answer

One issued job on a system we built, published line by line on the product itself: £98,214.72 across 63 measured lines on a build-over in Stockport. Where we fall short is in the scorecard below, marked as a fail.

Question 07

What happens if the answer is "do not build"?

Why it matters

A firm whose only product is a build has one answer available, whatever it finds. You are asking whether telling you the truth is survivable for them.

A good answer sounds like

What the client keeps when the answer is no, and a fee that still works when it happens.

A bad answer sounds like

"That has never happened." Which claims nobody who ever called them had a problem software could not fix.

Our answer

You keep the four outputs and the measurements underneath them, and we stop. We will not publish a percentage for how often that happens, because we have not run enough diagnostics for one to mean anything.

Question 08

Do I need an AI consultancy, an AI engineer, or a software agency?

Why it matters

These three cost different amounts and carry risk differently. The usual comparison is a salary against a project fee, which means nothing until somebody says what each buys and who pays when it runs long.

A good answer sounds like

A supplier who tells you which of the three your situation calls for, including when it is not them, and asks what the problem costs first.

A bad answer sounds like

Any answer that arrives before the problem is described. If the recommendation never changes, it is a product, not advice.

Our answer

Ours is the third column, and only when you cannot yet say which problem is worth money. If the specification is written, you want hands. If the workload never runs out, hire.

Three ways to buy the same capability, compared on cost shape and risk
 An in-house hireA software agency or contractorA fixed diagnostic, then a scoped build
What you are buyingA person, and their whole week, indefinitelyDelivery against a specification you already wroteA measured answer to which problem is worth money, then a build
Cost shapeA salary plus employer National Insurance, pension and recruitment. The example below uses £70,000, our own illustration, not a survey figureA project fee. The UK starting prices we index run from about £2,000 to a £150,000 ceiling, each read from the firm's own page on a dated check£2,000 fixed for 14 days, then a fixed-price proposal for the build
Time to the first checkable evidenceA hiring round, a notice period, then ramp-upAs fast as the specification allows, no faster14 days, to four outputs you keep
Who carries the overrunYou do, monthly, whether or not the work is thereWhatever the pricing shape you agreed decidesWe do, on the diagnostic. On a build, whatever the proposal says
What you own at the endEverything, plus a dependency on one personWhatever the contract assigns, worth reading twiceThe four outputs either way, and the system if you build
When it genuinely winsThe work is continuous enough to keep one person busy for a yearThe specification exists, is right, and you need handsNobody can say which problem costs the most, so specifying is guesswork

The arithmetic, and it is our own assumption rather than a survey: take the salary you would actually offer, add employer National Insurance, pension and any recruitment fee, then the months before the person is productive. Compare that twelve-month figure against a £2,000 diagnostic plus a scoped build.

Question 09

What are the red flags?

Why it matters

These are patterns, not firms, and you will meet several. None is proof of bad faith alone. Two together usually mean the engagement is the product.

A good answer sounds like

A supplier who names their weak spots before you find them, with an answer ready for each flag below.

A bad answer sounds like

Defensiveness about any of the five. Each is cheap to ask and cheap to answer.

Our answer

We fail one test on our own scorecard, published below rather than left out. For the rest, the page says where to check each claim.

Five patterns worth naming

  • The free audit that is a sales demonstration. Unpaid discovery is recovered from the work it recommends. Ask what it produces if you buy nothing.
  • Time and materials with no defined outcome. An open-ended day rate pays more the longer the problem lasts. Ask what the finished state is, and who pays past it.
  • A proposal that arrives before a diagnosis. A solution written before anyone measured the constraint is a product they already sell. Ask what they measured, and how.
  • Guaranteed savings nobody can measure. A guarantee is only as good as its baseline. Ask what the current figure is, and who agrees it.
  • A sales team standing between you and the engineers. Most of your questions are technical ones in commercial clothes. Ask to speak to whoever writes the code, before you sign.

The scorecard

A scorecard you can use, with our own score filled in.

Twelve criteria, the answer we would accept, and how AGMM does against each. It prints on one page. Take it to three firms.

Twelve criteria for choosing an AI consultancy, with AGMM scored against all twelve
CriterionWhat a good answer looks likeAGMMScore
A fixed price and scope for the first stepA number and a duration, published before the call£2,000, 14 days, publishedPass
Prices on the website, not on requestYou can price the first step without speaking to anyoneEvery price we charge is publishedPass
The people who write the code are in the sessionsNamed, and present for the whole engagementBoth directors, every sessionPass
You end up with a working system, not documentsSoftware in daily use, not a deckThe diagnostic is documents by design. The build is the systemPartly
Outputs are yours, in a format you can openNamed artefacts, transferable, no portal lockFour outputs, yours whether you build or notPass
The diagnostic fee is not credited against a buildThe recommendation does not have to be "build"Not credited, and it says so on the audit pagePass
An exclusion list in writing, before you payA heading, not a footnotePublished on the audit and pricing pagesPass
An artefact a stranger can open and checkA live URL or a public record, not a logo£98,214.72 across 63 measured lines, in fullPass
Named client references with published metricsNamed clients, agreed figures, contactableWe do not have this. No client has agreed to be named with numbersFail
A registered company you can look upA number that resolves on the public registerCompanies House 16384156, incorporated 13 April 2025Pass
The firm will tell you not to buyA route to "do not build", and a fee that survives itPaid diagnostic, no credit, so the answer can be noPass
What happens on an overrun, agreed up frontIn the contract, not discovered in week sixFixed on the diagnostic. On a build, set in the proposalPartly

The fail is row nine. We cannot show you a named client with agreed figures, because none has agreed to be named, and we will not publish an anonymised stand-in. What we can show is a real issued job, in full, on a system we built. That proves the software works, not that the client is happy.

The honest part

When are we the wrong choice?

Four situations where the answer is somebody else, written down rather than found out in week three.

  • The constraint is a decision, not a system. If the real problem is a pricing call, a staffing call or a process nobody has written down, software on top makes it faster rather than smaller.
  • You already have the specification. If the requirement is written, agreed and right, you need delivery capacity, and paying to diagnose something already diagnosed wastes £2,000.
  • The workload is continuous and predictable. Once there is a full year of work with no gaps in it, an employee is cheaper and the capability stays in the building. Hire.
  • You need named references with published numbers. Some buyers cannot proceed without them, quite reasonably. We do not have them yet.

Frequently asked questions

What buyers ask before they choose.

What questions should I ask an AI consultant before hiring one?
Ask who will actually do the work and whether they attend the sessions. Ask what you own if you never build. Ask whether the price is fixed, a day rate or a retainer, and who carries the overrun. Ask whether the diagnostic fee is credited against a build, because that exposes the incentive. Ask what is not included.
Is an AI consultancy better than hiring in-house?
They win in different situations. An in-house hire wins when the workload is continuous enough to keep somebody busy for a year, because you then own the capability. A consultancy wins when nobody can yet say which problem is worth solving. The common route is a diagnostic first, a scoped build second, and a hire once the workload is proven.
How long should an AI audit take?
Long enough to measure the constraint, short enough that the business has not changed underneath it. Ours runs fourteen days with three sixty-minute working sessions. Anything under a week is unlikely to have touched real data. Anything past six weeks is a consulting engagement with an audit label on it.
What should an AI audit deliver?
Named artefacts you keep, not a presentation. Ours delivers four: a constraint map, an AI opportunity map, economic cases with the assumptions written down, and a ranked next step with a fixed-price proposal. The test for any supplier is whether that list survives a decision not to build.
How much does an AI consultancy cost in the UK?
Most UK firms publish nothing, which is why quoted ranges are so wide. Among those that do, the figures we index run from about £200 a month to a £150,000 project ceiling, with fixed-scope diagnostics around £2,000. Every figure carries the page it came from and the date it was read.
How do I check an AI consultancy is real?
Look the company up on the Companies House register: the number, the incorporation date, the directors and the filing history are public and free. Be clear about what that proves. It shows the entity exists and who is behind it. It is not evidence of quality, and almost every firm will pass it, so treat it as a floor rather than a differentiator.
What are the red flags when hiring an AI consultancy?
A free audit that is really a sales demonstration. Time and materials with no defined finished state. A proposal that arrives before anybody measured anything. Guaranteed savings with no agreed baseline. A sales team between you and the engineers. Two of these together usually mean the engagement is the product.
Who owns the code an AI consultancy writes for you?
Whatever the contract says, which is why it is worth reading before the kick-off. Ask for assignment of the source code and infrastructure configuration on delivery, in writing, and ask what happens to components the supplier reuses elsewhere.

Next step: the free thirty-minute fit call.

Bring the nine questions and ask us all of them, including the one we answer badly. Nothing is sold on the call. If the diagnostic is the right next step, it is £2,000 fixed for 14 days, not deducted from any build.

Book the free fit call

Or read what UK AI consultancies charge before you book anything