How to choose an AI consultancy in the UK: the questions, the prices and a scoring rubric.

The UK now has hundreds of firms calling themselves AI consultancies, and they are two different trades sharing a name. Advisers produce documents: readiness assessments, roadmaps, policies, training. Builders produce systems that run: an agent that answers the phone, a pipeline that reads documents, a workflow that updates three systems. Both are legitimate. The mistake is buying one when you needed the other, and the commonest form of it is paying for a roadmap from a firm that cannot build what it recommends, then discovering that the builders you take it to want to start again.

The test is production. Ask to see something the firm built that is running today, with the numbers: how many tasks a month, what it costs to run, what happens when it fails. An adviser will show you a slide. A builder will show you a dashboard, or a call recording, or a log, and will talk about voicemail rates and per-message costs, because those are the things that hurt them last month. The Office for National Statistics found in June 2026 that only 10 per cent of AI-adopting businesses with ten or more employees use it extensively; the rest are using a tool rather than running a system, and the firms that can move you from the first group to the second are fewer than the directory suggests.

There is a respectable middle: firms that advise first and build second, with the advice priced so that it is credited against the build. Ask whether the audit fee comes off the project. If it does, the firm expects to build; if it does not, the audit is the product. Neither is dishonest, but the credited model tells you the firm is prepared to be held to its own recommendation, which is the thing a roadmap on its own can never promise.

Five firms that publish their prices, read on their own pages on 14 September 2026, plus our own. They span every pricing model in the market: fixed fees, credited discovery, monthly retainers, day rates and hourly work. None of them is wrong; each suits a different kind of buyer, and the table is here so you can see which is which before a sales call frames it for you.

All figures exclude VAT and were read on the firms’ own pages on 14 September 2026. Where a firm describes the wider market rather than its own price, the table says so. Prices change; the date at the top of this page is the date of the last check.

Helium42’s April 2026 survey of the UK market found fixed-price engagements accounting for 45 to 55 per cent of SME work, retainers running £3,000 to £15,000 a month for on-call advisory, hourly work at £150 to £400, and outcome-based pricing adopted by only 12 to 18 per cent. It also reported average overruns of 15 to 25 per cent under time and materials. Set against that, the measured market is cheaper than any consultancy’s rate card: ITJobsWatch puts the median UK AI consultant contract at £575 a day in the six months to 14 September 2026, with nine in ten below £734, from 96 postings, and London at the same median.

The reading is straightforward. A fixed fee is right for anything with a definable end: an audit, a pilot, a build. A retainer is right for a standing need, such as a fractional AI lead or the monitoring of systems already live, and the published range of £750 to £6,000 a month covers both. A day rate is right for research and remediation and should be capped. A quote that offers only a retainer for work that plainly has an end is asking you to fund the firm’s cash flow. A quote that offers a fixed price for genuine research is guessing, and the guess is in the premium.

Use the measured rate to test scope rather than price. A £5,000 audit is about nine days of the median contract rate; if the proposal implies three weeks of senior time for that money, either the firm is efficient or the audit is thinner than it sounds. Ask which. The same test works upward: a £40,000 build is about seventy days at the median, and a proposal should be able to say roughly where those days go.

These are the questions to ask before the price, because the answers change what the price is for. Each has a checkable form: a clause you can read, a number you can test at your own volume, or a document the firm either has or does not. Ask them in the first meeting rather than the last, because a firm that has good answers will be glad of the chance to give them, and a firm that does not will spend the remaining meetings steering you back to the demo.

A regulated buyer does not get a new rulebook for AI; the old one applies to the new tool, and the regulators have said so in terms. The Solicitors Regulation Authority’s Risk Outlook on AI in the legal market tells firms to "Tell clients when you will be using AI with their case, and how it will operate", warns that you cannot "trust an AI system to judge its own accuracy", and is explicit that firms "must remain responsible" for their activities whoever supplies the technology. In August 2026 the SRA reported 42 reports of potential AI misuse in the year to July 2026 and restated the position: "Individuals remain responsible for the work they produce and the advice they provide, regardless of whether AI is used, or not."

The ICO’s rules apply to every buyer. Its list of processing that requires a data protection impact assessment includes "processing involving the use of innovative technologies, or the novel application of existing technologies (including AI)" when combined with other risk factors, decisions about access to a product or service based on automated decision-making, and large-scale profiling. Most customer-facing AI systems touch at least one of those. A consultancy that has not asked whether a DPIA is needed has not read the guidance it is about to make you rely on.

The practical consequence for a law firm, an accountancy practice, a clinic or a care provider is that the consultancy must be able to answer three things: how the client will be told, how a person supervises the output, and what the audit trail looks like when a regulator asks. Those are design requirements, and they belong in the scope before the price, because they change it.

A forty-person professional services firm wants to know whether AI can take the first pass at inbound client queries and draft the reply for a person to approve. Here is what the published prices say it would cost to find out, and then to do it, at four of the firms in the table. The figures are the firms’ own; the arithmetic and the pairing of discovery with build are ours, and a real quote would depend on the integrations involved.

At Hartz AI, the readiness audit is from £5,000 plus VAT and a roadmap from £7,500, or both from £10,000; the firm does not publish a build price and describes the market for strategy projects at £15,000 to £50,000. At The AI Consultancy, a two-week readiness sprint is from £3,500, a discovery and pilot £15,000 to £35,000, and a production build from £40,000; a day rate of £950 to £1,500 covers anything undefined. At FiveForward, discovery is £1,500 and credited back once you spend £3,500 more, with the automation typically £5,000 to £15,000 fixed. At Augustova, the audit is £2,500 to £6,000 and credited in full, with a first implementation at £4,000 to £15,000.

Two routes to a working system therefore cost £4,000 to £15,000 all in, because the discovery is credited: FiveForward’s £1,500 once follow-on spend passes £3,500, ours in full. One route costs £58,500 to £78,500 or more, being The AI Consultancy’s sprint, pilot and build added together. Hartz AI’s route ends at a roadmap from £10,000 that you take to a builder. None of these is a bad buy for the right client. The £40,000 build includes integration, compliance and training for a firm that wants one supplier to carry it all; the £5,000 to £15,000 build suits a firm with one clear process and its own people to own it. What you cannot do is compare them on the audit price alone, because the audit is the smallest number in every column.

Score each criterion 0, 1 or 2 for every firm you speak to, weight as shown, and treat a zero in any weight-3 row as a stop. The rubric is deliberately blind to how impressive the demo was; demos are the one thing every firm in the market is good at. It is also blind to price, on purpose: score first, then look at the prices of the firms that passed, because a cheap firm that fails a weight-3 row is not cheap.

Maximum 44. A firm scoring under 30, or zero on a weight-3 row, is an adviser you might still use for a document, not a builder to trust with a system.

Build or advise: we build, and we run the same kinds of agents in our own products, which is where the running-cost numbers in our guides come from. Data handling: a processor contract with the Article 28 terms before we touch live data, a named model provider per feature, and no training on your data. Model choice: chosen per feature against cost and accuracy on your inputs, stated in the scope, changeable without a rebuild. Running cost: stated before go-live, by line, and capped in production; ceilings are triggered deliberately before anyone relies on them. Ownership: code, prompts, evaluation sets and every account are yours from day one, assigned in writing. Measurement: an acceptance threshold agreed as a number before the build, measured on a sample of your real inputs including the messy ones.

Prices, fixed and excluding VAT: AI adoption audit £2,500 to £6,000 over one to two weeks, credited in full against a build; first implementation or agent £4,000 to £15,000; custom AI system £8,000 to £25,000; training £1,200 for a half day, £2,000 for a full day, programmes up to £6,000. No day rates, no minimum, no retainer until something is live and worth supporting.

We read the pricing and service pages of The AI Consultancy, Hartz AI, Ronins, FiveForward and Yekta Design Studio on 14 September 2026 and recorded every published figure with its basis; where a firm described the market rather than its own price we say so, and where it published no figure we say that too. Market pricing-model shares and overrun figures are Helium42’s April 2026 survey. Measured day rates are ITJobsWatch’s for the six months to 14 September 2026. Adoption figures are the ONS Business Insights and Conditions Survey as reported in July 2026.

Regulatory quotations are from the SRA’s Risk Outlook and its August 2026 release, and from the ICO’s guidance on processor contracts and impact assessments. The first-party running-cost description is the share already published in our guide on what AI features cost to run; no new internal figures appear here. All prices exclude VAT. The worked example is arithmetic on published prices, not a client account, and we will revise the table as the firms change their pages.