AI consulting is the fastest-growing line in the UK technology budget and the one buyers are least equipped to challenge. A software quote can be compared against another software quote. A cloud bill has a published price list. An AI consulting proposal, by contrast, can arrive at £40,000 or £400,000 for what reads like the same paragraph of scope — and most buyers have no reference points with which to argue.
This guide supplies those reference points. It sets out 2026 UK AI consulting day rates by supplier tier, the five project shapes that make up almost all commercial artificial intelligence work and what each genuinely costs, a phase-by-phase day count, a full worked budget for one 90-person business, the running costs that begin the day the consultants leave, and a three-year total cost of ownership. Every figure below is either a published market rate, a UK government statistic, a vendor list price, or arithmetic performed on numbers stated in this article.
Two companion guides sit either side of this one. Our Microsoft 365 Copilot cost and ROI guide prices the buy-a-licence route that many businesses should try before commissioning anything bespoke, and the AI governance framework for SMEs covers the policy work that any serious project has to sit inside. If you are pricing bespoke software more broadly, our custom software development cost guide uses the same costing method on a different problem.
Table of contents
- The One-Minute Answer on AI Consulting Cost
- What AI Consulting Actually Buys in 2026
- UK AI Consulting Day Rates: The 2026 Benchmarks
- The Five AI Consulting Project Shapes and What Each Costs
- Where the Days Actually Go in an AI Consulting Engagement
- A Worked AI Consulting Budget: One 90-Person Brokerage
- Four AI Consulting Delivery Routes Compared
- What an AI Project Costs to Run After AI Consulting Ends
- The Three-Year Cost of Ownership Your AI Consulting Quote Omits
- Why the AI Consulting Business Case Fails at 56 Per Cent
- Ten Things That Inflate an AI Consulting Quote
- Ten Things That Cut an AI Consulting Quote Before You Brief Anybody
- How to Read an AI Consulting Proposal
- Red Flags in AI Consulting Proposals
- Compliance and Governance Costs Inside an AI Consulting Budget
- In-House, Contract or AI Consulting: The Real Comparison
- Frequently Asked Questions About AI Consulting Cost
- What This Means for Your Budget
- References and Further Reading
The One-Minute Answer on AI Consulting Cost
If you read one section, read this one. The numbers below are the honest middle of the UK market in 2026, and the rest of the guide shows how each is built.
The headline bands
A first, properly scoped AI project at a UK business of 50 to 250 people costs £45,000 to £140,000 in AI consulting fees and takes three to seven months. A discovery engagement on its own is £8,000 to £18,000 and three to five weeks. A multi-step agent with tool access, approvals and audit trails is £110,000 to £240,000. Below about £25,000 you are buying a prototype, which is a legitimate thing to buy as long as everyone agrees that is what it is.
The day rate that underpins it
Fair market for competent UK AI consulting in 2026 is £950 to £1,400 per day at a small specialist firm, with £1,100 a reasonable midpoint for a named senior consultant. Independents sit at £700 to £950. Mid-tier consultancies and systems integrators charge £1,200 to £1,800. The global firms start around £1,800 and pass £3,000 for partner time.
The line nobody budgets
Running cost. The model licence is trivial; the supervision is not. Budget 30 to 45 per cent of the original build cost, every year, for evaluation reruns, regression testing when models change, monitoring, and the internal owner who does none of those things if you have not named them.
The single biggest variable
Data readiness. Not model choice, not vendor, not sector. In the worked AI consulting budget later in this guide, preparing and plumbing the data takes 22 of 92 consultant days — more than the retrieval and prompt build it feeds. Two businesses with identical use cases can sit £40,000 apart purely on the state of the documents they already hold.
What AI Consulting Actually Buys in 2026
Before pricing anything, be precise about what you are buying, because “AI consulting” covers at least four distinct products in the UK market and they are not interchangeable. Confusing them is the most common reason a quote looks wrong.
Four products under one name
| Product | What you get | Typical UK price | Buy it when |
|---|---|---|---|
| Advisory | Opportunity assessment, prioritised roadmap, business case, no code | £8,000–£25,000 | You have candidates but no ranking |
| Delivery | A working system in your estate, integrated and evaluated | £45,000–£240,000 | One use case is already chosen and sponsored |
| Capability build | Your staff trained, your platform, your patterns, handover | £60,000–£180,000 | You intend to do this repeatedly in-house |
| Fractional leadership | Two to four days a month of senior direction and vendor challenge | £2,400–£5,600/month | You have builders but nobody to point them |
A proposal that mixes all four without separating the prices is not a proposal; it is a hope. Ask for them as four lines. Most credible AI consulting firms will do this without complaint, because they price that way internally anyway.
Where AI consulting differs from ordinary technology consulting
General technology consulting is largely a design and coordination exercise: the components exist, the integration is known, the risk is in sequencing. AI consulting carries an additional risk that traditional projects do not — you cannot know from the specification whether the thing will work well enough. A search index either returns the row or it does not. A language system returns something plausible every single time, including when it is wrong.
That single property drives most of the cost difference. It is why a serious AI consulting engagement contains an evaluation phase, a labelled test set, and an agreed accuracy threshold before anybody discusses rollout — and why an engagement without those things is cheaper for a reason.
It is not the same as rules-based automation
If the task is deterministic — move this field to that system, raise this ticket when that inbox receives that subject line — then you want business process automation, not a language model, and it will cost a third as much and never hallucinate. Some of the best value an AI consulting partner delivers in week one is telling you that four of your six candidate use cases are ordinary automation problems wearing a fashionable hat.
What AI consulting cannot do for you
No engagement can supply organisational willingness. It cannot make account handlers trust an output they were not consulted about, it cannot decide your risk appetite for a wrong answer, and it cannot own the outcome after handover. Those three things stay with you, and every project that failed in the research below failed on one of them rather than on model quality.
UK AI Consulting Day Rates: The 2026 Benchmarks
Two independent reference points bound this market: what individual specialists charge on the open contract market, and what firms charge for a delivered engagement. The gap between them is where most of the confusion about AI consulting cost lives.
The contract market floor
ITJobsWatch samples advertised UK contract rates. In the six months to 20 August 2026, the median daily rate for an AI Engineer contract role was £588, from 275 daily rates quoted, up 11.90 per cent year on year from £525 — with vacancies rising from 123 to 432 over the same comparison. The 10th percentile was £415, the 25th £500, the 75th £688 and the 90th £870.
For AI Consultant contract roles over the same six months, the median was £575 from 73 rates quoted, up 0.66 per cent from £571 a year earlier, with the 10th percentile at £463, the 25th at £531, the 75th at £700 and the 90th at £734.
On the permanent side, UK roles citing machine learning had a median advertised salary of £75,000 from 1,421 salaries quoted in the same six-month window — down 6.25 per cent from £80,000 — with the 10th percentile at £46,250, the 25th at £55,000, the 75th at £97,500 and the 90th at £118,750. Outside London the median was £65,000.
Why a consultancy charges twice the contract rate
The £500-plus gap between the £588 contract median and a £1,100 specialist AI consulting rate is not pure margin. It buys a second pair of eyes on every design decision, a delivery method that survives one person leaving, professional indemnity cover, an evaluation practice you would otherwise have to invent, and — critically — someone who has already made the expensive mistakes on somebody else’s project. Whether that is worth £512 a day is a real question with a real answer, and the answer changes with how much of the risk you can absorb yourself.
The full rate card
| Supplier tier | Day rate band | Midpoint used here | Best for |
|---|---|---|---|
| Independent consultant | £700–£950 | £825 | Advisory, review, second opinion |
| Small specialist AI consulting firm | £950–£1,400 | £1,100 | First delivery, 50–250 staff |
| Mid-tier consultancy or integrator | £1,200–£1,800 | £1,500 | Multi-system estates, regulated data |
| Global firm | £1,800–£3,000+ | £2,400 | Board-level programmes, group scale |
| AI Engineer, contract (ITJobsWatch) | £415–£870 | £588 | Augmenting a team you already lead |
| AI Consultant, contract (ITJobsWatch) | £463–£734 | £575 | Filling one named gap |
The market context behind the rates
Demand is real rather than rhetorical. The Office for National Statistics reported on 20 July 2026 that self-reported AI use among UK businesses with 10 or more employees had risen from around 12 per cent in late 2023 to around 35 per cent by June 2026, reaching 49 per cent among businesses with 250 or more staff and 28 per cent among those with fewer than 10. Adoption by industry ranged from 58 per cent in information and communication down to 13 per cent in construction.
Depth, however, has barely moved. The average number of AI technologies used per adopting business rose only from around 1.4 to around 1.6 over the same period, and just 10 per cent of adopting businesses described their use as extensive. That gap between breadth and depth is precisely the gap an AI consulting engagement is bought to close — and it is why so many quotes are for a first real project rather than a fifth.
The Five AI Consulting Project Shapes and What Each Costs
Almost every commercial engagement in this market is one of five shapes. Identifying yours before you request proposals is worth more than any negotiation tactic, because it lets you compare like with like.
The price bands by shape
| Project shape | Consultant days | UK fixed-price band | Elapsed time | Main cost driver |
|---|---|---|---|---|
| Discovery and opportunity assessment | 8–15 | £8,000–£18,000 | 3–5 weeks | Number of business units interviewed |
| Single-workflow assistant over your documents | 45–75 | £45,000–£95,000 | 3–5 months | Document quality and volume |
| Extraction pipeline into a system of record | 60–110 | £65,000–£140,000 | 4–7 months | Number of document types and edge cases |
| Multi-step agent with tool access and approvals | 90–180 | £110,000–£240,000 | 6–11 months | Number of systems it may act upon |
| Predictive model on your own data | 55–100 | £60,000–£130,000 | 4–8 months | History depth and label quality |
| Governance and policy programme | 20–40 | £22,000–£48,000 | 2–4 months | Regulatory overlay and group structure |
| Fractional AI leadership | 2–4 days/month | £2,400–£5,600/month | Ongoing | Seniority and number of workstreams |
Shape one: the assistant over your own documents
The most common first project, and the one most AI consulting proposals in the UK are written for. A retrieval system indexes documents you already own, and a language model answers questions or drafts output grounded in what it retrieves. The technique leans on natural language processing that has been production-grade for several years; the difficulty is never the model, it is that your 40,000 documents include six formats, three naming conventions, two decades of inconsistency and a subset nobody is allowed to see.
Shape two: extraction into a system of record
Read an invoice, a policy schedule, a lab report or a claim form and write structured fields into a database. This shape has the clearest business case in the market because the saving is countable, and the highest hidden cost, because the last 8 per cent of edge cases takes as long as the first 92. Price the exception path explicitly or it will be priced for you later.
Shape three: the multi-step agent
The system does not merely answer; it acts — reading a case, checking a system, drafting a response, requesting approval. Costs rise sharply here because every system it touches needs credentials, permissions, an audit trail and a rollback story. Gartner predicted in June 2025 that over 40 per cent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. That is a budgeting fact, not a reason to abstain: it means the approval and control design is the project, and pricing it as a trimmable extra is how the 40 per cent happens.
Shape four: prediction on your own data
Churn, demand, credit, failure, no-show. This is classical ML model development rather than generative work, and it is priced by the state of your history. Three years of clean, labelled outcomes makes it a 55-day job. Two years of partially migrated records with a system change in the middle makes it a 100-day job, and most of those extra days are data management and analytics work rather than modelling.
Shape five: the governance programme
Policy, risk classification, approval routes, an inventory of what is in use, supplier assessment and staff guidance. It creates no working software and it is the cheapest insurance in this market. Sequence it alongside the first delivery, not after the third.
Where the Days Actually Go in an AI Consulting Engagement
This is the section that changes how proposals read. The instinctive assumption is that most of the money buys model work. It does not.
Data readiness is the biggest single line
Twenty-two of ninety-two days — just under a quarter of the engagement — go on locating, cleaning, de-duplicating, permission-mapping and chunking the content the system will read. This is unglamorous, invisible to sponsors, and the phase most often cut to win a competitive AI consulting tender. It is also the phase whose omission produces a demo that dazzles in March and is abandoned in September.
Evaluation is smaller than you expect and more important than everything
Nine days builds a labelled test set, an automated scoring harness and an agreed accuracy threshold. Without it you have no way to answer the only question that matters after go-live: is it still as good as it was? When a model version changes underneath you — and it will — the harness is the difference between a two-day regression check and a three-week argument.
Integration is a systems problem, not an AI problem
Fourteen days on connectors, authentication, rate limits, error handling and the unglamorous business of writing results back where staff already work. Any AI consulting partner who treats integration as an afterthought has built you a demo with a login page.
Change and training decide whether any of it pays
Ten days. The MIT NANDA study The GenAI Divide: State of AI in Business 2025 reviewed more than 300 enterprise initiatives alongside 52 executive interviews and 153 surveys, and found that 95 per cent of organisations were getting no measurable return on an estimated $30–40 billion of enterprise investment, with only around 5 per cent of custom tools reaching production. The reported causes were adoption and workflow integration, not model capability. More than half of budgets went to visible functions such as sales and marketing despite better returns in back-office work.
A Worked AI Consulting Budget: One 90-Person Brokerage
Abstract bands only get you so far. Here is one complete engagement, priced line by line, using the £1,100 midpoint for a small specialist firm.
The scenario
A 90-person UK commercial insurance brokerage with 11 account handlers processes roughly 1,900 renewals a year. Renewal-pack preparation currently takes about 3.5 hours per renewal. The document store holds 42,000 PDFs across a broking platform and Microsoft 365. The goal is to cut preparation to 1 hour using an assistant that reads the prior year’s documents, drafts the summary and answers handler questions with citations.
The build cost
| Phase | Days | Cost at £1,100/day | Share of engagement |
|---|---|---|---|
| Discovery and use-case selection | 8 | £8,800 | 8.7% |
| Data readiness and document pipeline | 22 | £24,200 | 23.9% |
| Retrieval and prompt build | 16 | £17,600 | 17.4% |
| Evaluation harness and test set | 9 | £9,900 | 9.8% |
| Integration with broking platform and Microsoft 365 | 14 | £15,400 | 15.2% |
| Security, impact assessment and controls | 6 | £6,600 | 6.5% |
| Change, training and rollout | 10 | £11,000 | 10.9% |
| Hypercare, first six weeks | 7 | £7,700 | 7.6% |
| Total | 92 | £101,200 | 100% |
The value on the other side of the ledger
At 1,900 renewals and 2.5 hours saved each, the project releases 4,750 hours a year. At a loaded cost of £34 an hour, that is £161,500 of released capacity annually. Hold that number lightly — released capacity is not banked cash, and the section on the 56 per cent threshold below is the most important arithmetic in this guide.
Where this budget would move
Add a second use case and the discovery, evaluation and change phases barely move while the build and data phases roughly repeat: call it £55,000 to £70,000 more, not another £101,200. Move from a specialist firm at £1,100 to a mid-tier at £1,500 and the same 92 days becomes £138,000. Take it to a global firm at £2,400 and it becomes £220,800 — for identical scope, which is the strongest argument in this guide for matching supplier tier to problem size.
Four AI Consulting Delivery Routes Compared
The same outcome can be bought four ways. The right answer depends on how much capability you intend to keep.
| Route | Year-one cost | You own | Main risk |
|---|---|---|---|
| A — Buy a vendor product Licence £2,400/month plus 14 configuration days | £44,200 | Nothing; you rent | Fits perhaps 70% of your workflow |
| B — Assisted build 92 days at £1,100, time and materials | £101,200 | Code, prompts, evaluations, data | Scope creep sits with you |
| C — Fixed-price outcome Same scope, schedule risk transferred | £118,000 | Code, prompts, evaluations, data | Change control becomes adversarial |
| D — Embedded fractional team 2 days/week for 39 weeks at £1,250 | £97,500 | Code plus real internal capability | Needs your people to show up |
Why the cheapest route is not automatically wrong
Route A is less than half the cost of anything bespoke and should be tried first for any workflow that is not genuinely specific to you. The honest test is the 70 per cent question: if a product covers 70 per cent of the workflow and the remaining 30 per cent is preference rather than obligation, buy the product and change the preference. If the 30 per cent is regulatory, contractual or the actual source of your margin, no configuration will close it and Route A becomes the most expensive option in the table — because you pay £44,200 and still need Route B.
Why fixed price costs 16.6 per cent more
Route C is £16,800 above Route B for identical scope, a premium of 16.6 per cent. That is the market price of transferring schedule and estimation risk. It is good value when your requirement is genuinely stable and poor value when it is not, because every fixed-price AI consulting contract converts discovery into a change request — and discovery is the thing you most want to be free to act on in a first project.
What an AI Project Costs to Run After AI Consulting Ends
The build is a project. The system is a liability. Most business cases in this market model the first and ignore the second.
Inference is almost never the expensive part
Published list prices at the time of writing put a mid-tier frontier model at roughly $3 per million input tokens and $15 per million output tokens, and a top-tier model at about $5 and $25 respectively. Smaller models sit near $1 and $5. Batch processing typically halves those rates, and prompt caching cuts the cost of the repeated context that dominates a document assistant.
Work it through for the brokerage. Each renewal triggers about three assistant calls, so 1,900 renewals produce 5,700 calls a year. At roughly 45,000 input tokens and 2,000 output tokens per call, that is 256.5 million input tokens and 11.4 million output tokens annually. On the mid-tier model that is $769.50 of input plus $171.00 of output, or $940.50 a year. On the top-tier model it is $1,282.50 plus $285.00, or $1,567.50. At $1.34 to the pound, the mid-tier figure is about £700 a year and the top-tier about £1,170.
Set that against £101,200 of build. The model bill is roughly 0.7 per cent of the first-year cost. Any AI consulting conversation that spends an hour on model pricing and ten minutes on evaluation has its priorities inverted.
What actually costs money after go-live
| Running cost line | Basis | Year-one cost |
|---|---|---|
| Model inference | 5,700 calls, mid-tier list pricing | £700 |
| Hosting, vector store, observability | £280 per month | £3,360 |
| Monitoring and evaluation reruns | 1 day per month at £1,100 | £13,200 |
| Model-change regression testing | 4 events per year, 2 days each | £8,800 |
| Internal product owner | 0.4 FTE of a £58,000 role | £23,200 |
| Total running cost, year one | £49,260 |
Running cost is £49,260 against a £101,200 build — 48.7 per cent of the build in the first year, and the two largest lines are people, not compute. This is the number most AI consulting proposals leave to the reader to discover.
Why the model changing underneath you is a budget line
Providers deprecate versions, adjust behaviour and release successors. A system tuned against one version can shift on another in ways no unit test catches — a summary that grows longer, a refusal that appears where it did not before, a format that drifts. Four regression events a year at two days each is not pessimistic; it is the reason the evaluation harness earns its nine days back within the first twelve months.
The Three-Year Cost of Ownership Your AI Consulting Quote Omits
| Year | Build or enhancement | Running cost | Total |
|---|---|---|---|
| Year 1 | £101,200 (92 days) | £49,260 | £150,460 |
| Year 2 | £13,200 (12 enhancement days) | £49,260 | £62,460 |
| Year 3 | £8,800 (8 enhancement days) | £49,260 | £58,060 |
| Three-year total | £123,200 | £147,780 | £270,980 |
Two observations follow. First, the AI consulting fee is only 37.3 per cent of the three-year bill: £101,200 of £270,980. Second, the run rate does not decay — years two and three are 42 and 39 per cent of year one, because supervision does not get cheaper just because the system is older.
Why the AI Consulting Business Case Fails at 56 Per Cent
Here is the arithmetic that ought to sit on the first page of every proposal in this market, and almost never does.
The break-even calculation
The project releases £161,500 of capacity a year, or £484,500 over three years. It costs £270,980 over the same period. Dividing one by the other, the project needs 56 per cent of the theoretical time saving to convert into real, banked value before it breaks even — 270,980 divided by 484,500 is 0.559.
Why that number is uncomfortable
Released capacity converts only when something specific happens to it. Either headcount does not grow when volume does, or the freed hours go to work that generates fee income, or a service level improves in a way that wins retained business. If none of those is named and owned, the hours simply diffuse — and at a 50 per cent realisation rate this project returns £242,250 against £270,980 of cost and loses £28,730. At 70 per cent it returns £339,150 and makes £68,170.
The distance between a loss and a healthy return here is not technical at all. It is whether somebody was made accountable for converting hours into money on the day the project was approved.
What to do with the number
Put the break-even realisation rate in the business case, name the person who owns it, and review it quarterly against the evaluation harness output. An AI consulting engagement that helps you write that sentence has already earned a meaningful share of its fee, whatever the model turns out to do.
Ten Things That Inflate an AI Consulting Quote
| # | What inflates it | Typical added days | Cost at £1,100 |
|---|---|---|---|
| 1 | No named decision-maker for accuracy thresholds | 6–12 | £6,600–£13,200 |
| 2 | Documents scattered across more than three systems | 8–15 | £8,800–£16,500 |
| 3 | Permissions that must be mirrored per document | 5–10 | £5,500–£11,000 |
| 4 | Scanned or image-only source material | 6–14 | £6,600–£15,400 |
| 5 | A line-of-business system with no usable API | 8–20 | £8,800–£22,000 |
| 6 | Three or more stakeholder groups with veto rights | 5–12 | £5,500–£13,200 |
| 7 | Regulated data requiring a full impact assessment | 4–9 | £4,400–£9,900 |
| 8 | A requirement to run models in your own tenancy | 10–25 | £11,000–£27,500 |
| 9 | Multilingual content or multiple jurisdictions | 6–15 | £6,600–£16,500 |
| 10 | “While we are in there” additions after kick-off | unbounded | unbounded |
Item eight deserves special attention. Running models inside your own cloud tenancy is a legitimate requirement for some regulated workloads and an expensive reflex for most others. Ask what specific obligation drives it before accepting £27,500 of AI consulting time and a permanent infrastructure bill.
Ten Things That Cut an AI Consulting Quote Before You Brief Anybody
Every day of preparation you do yourself is a day you do not buy at £1,100. None of the following requires technical skill.
Preparation that removes days
| # | Do this first | Days removed | Saving at £1,100 |
|---|---|---|---|
| 1 | Write the decision the system supports, in one sentence | 2–4 | £2,200–£4,400 |
| 2 | Name one accountable business owner, not a committee | 3–6 | £3,300–£6,600 |
| 3 | Collect 40 real examples with known correct answers | 4–7 | £4,400–£7,700 |
| 4 | Consolidate the documents into one repository | 5–10 | £5,500–£11,000 |
| 5 | Agree what an acceptable error rate is, in advance | 2–5 | £2,200–£5,500 |
| 6 | Confirm API access to the target system in writing | 3–8 | £3,300–£8,800 |
| 7 | Complete your data protection impact assessment | 2–4 | £2,200–£4,400 |
| 8 | Publish an internal policy on acceptable AI use | 2–3 | £2,200–£3,300 |
| 9 | Baseline the current process with real timings | 2–4 | £2,200–£4,400 |
| 10 | Decide who converts released hours into value | 1–3 | £1,100–£3,300 |
Taken together those remove 26 to 54 days — £28,600 to £59,400 at the same rate. On the 92-day worked example that is between a quarter and well over half the engagement, achieved entirely with internal effort that costs you nothing on an AI consulting invoice.
The single highest-value item on that list
Number three. Forty real cases with agreed correct answers is the raw material of the evaluation harness, and it is the one artefact that no consultant can produce for you because it encodes your judgement, not theirs. Businesses that arrive with it consistently get better proposals, because it converts a vague scope into a measurable one.
How to Read an AI Consulting Proposal
The seven questions worth asking
- How many days are in data preparation, and what happens if the documents are worse than expected? A proposal with fewer than 15 per cent of days in data work has either seen your data or has not thought about it.
- What is the evaluation method, and who signs off the threshold? If the answer is “we will demo it to you”, there is no method.
- Which model, and what happens when it is deprecated? You want a named regression process, not reassurance.
- Who owns the prompts, the code, the evaluations and the fine-tuned artefacts? All four should be yours in writing.
- What does month four look like if we stop? A good AI consulting partner has an honest answer that does not involve you being stranded.
- What is the running cost, itemised, for the first twelve months? Compare it against the 48.7 per cent figure in this guide.
- What would you refuse to build for us? The most informative question in the list.
Reading the day mix
| Phase | Healthy share | Worry below | Worry above |
|---|---|---|---|
| Discovery | 7–12% | 5% | 20% |
| Data readiness | 20–30% | 15% | 40% |
| Build | 15–25% | 10% | 40% |
| Evaluation | 8–12% | 5% | 20% |
| Integration | 12–20% | 8% | 30% |
| Change and training | 8–15% | 5% | 25% |
A build share above 40 per cent usually means the data problem has been assumed away. A discovery share above 20 per cent usually means the supplier is selling thinking because it cannot sell delivery.
Red Flags in AI Consulting Proposals
Eight things that should stop the conversation
- No evaluation phase at all. The single most reliable indicator of a demo-shaped engagement.
- A price with no day count behind it. You cannot compare two of those, which is usually the point.
- Refusal to name the individuals. AI consulting is bought from people, not brands; the CVs matter more here than in most disciplines.
- Intellectual property retained by the supplier. Prompts and evaluation sets are your business logic. Never rent them.
- No running cost estimate. Half your three-year bill, omitted.
- Accuracy claimed as a percentage with no test set named. A number without a denominator is marketing.
- Pilot pricing that only makes sense if you continue. Check what the second phase costs before you sign the first.
- A proposal that never mentions your existing systems by name. It was written for somebody else.
Compliance and Governance Costs Inside an AI Consulting Budget
Budget six to nine days for this on a first project, more if you sell into the European Union.
The UK position
The United Kingdom has no AI Act. Regulation runs through existing regulators within their remits, with the Department for Science, Innovation and Technology publishing policy rather than statute. Its AI Opportunities Action Plan: One Year On, published on 29 January 2026, reported 38 of the plan’s 50 actions met, five AI Growth Zones designated across Great Britain attracting £28.2 billion of investment and more than 15,000 jobs, over one million AI courses delivered against a target of upskilling 10 million workers by 2030, and £240 million committed to the AI Security Institute alongside up to £500 million for a Sovereign AI Unit.
For a UK business commissioning work, the practical obligations are the familiar ones: data protection law, sector regulation, and contractual commitments to your own customers. That is cheaper than a bespoke regime, but it is not nothing, and it is the reason a data protection impact assessment sits in the six-day security and controls phase rather than outside the budget.
The European overlay, if it applies to you
The European Union’s timetable moved in 2026. The Digital Omnibus, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, deferred high-risk obligations for stand-alone Annex III systems to 2 December 2027 and for high-risk AI embedded in already-regulated products to 2 August 2028. What did not move was Article 50: transparency duties covering chatbot disclosure, AI content marking and deepfake labelling took effect on 2 August 2026, alongside the AI Office’s enforcement powers over general-purpose model providers.
If your system serves EU users, the transparency obligations are live now and the high-risk classification question has been deferred rather than cancelled. Ask any prospective AI consulting supplier which of those two statements applies to your use case; if they cannot answer, they have not read the amendment.
Sizing the UK market you are buying in
The Department for Science, Innovation and Technology’s most recent Artificial Intelligence Sector Study counted more than 5,800 AI companies in the UK — an 85 per cent increase over two years — generating £23.9 billion in revenue, contributing £11.8 billion in gross value added and employing 86,139 people, up 33 per cent year on year. That is a crowded, fast-growing supplier market, which is good for price and bad for quality variance. Reference checks matter more here than in mature categories.
In-House, Contract or AI Consulting: The Real Comparison
| Option | Cost of 92 days | Time to start | Capability retained | Risk if they leave |
|---|---|---|---|---|
| Permanent hire at £75,000 median | £29,900 of salary cost, plus recruitment and 3–6 months of ramp | 3–6 months | High, while they stay | Total |
| Contractor at £588/day | £54,096 | 2–5 weeks | Low unless you insist | High |
| Specialist AI consulting firm at £1,100/day | £101,200 | 2–6 weeks | Medium; high with Route D | Low |
| Mid-tier consultancy at £1,500/day | £138,000 | 4–10 weeks | Medium | Low |
The permanent column looks decisive until you read the three columns beside it. Ninety-two working days is roughly 37 per cent of a working year, so £75,000 of salary apportions to about £29,900 — but you cannot hire for 92 days, the ramp consumes a third of the first year, and a single hire carries no second opinion, no cover and no method.
The honest reading is that a permanent hire is the cheapest way to own capability and the most expensive way to start. That is why the sequence that works for most UK businesses is an AI consulting engagement first and a hire second, briefed by what the AI consulting work taught you.
When to hire instead
If you can name three further projects with sponsors attached, hire. If you can name one, buy the delivery and keep the option. The failure mode is hiring a single specialist against a single project, then discovering that the project needed a data engineer, an integration developer and someone who has written an evaluation harness before.
Frequently Asked Questions About AI Consulting Cost
How much does AI consulting cost for a small UK business?
For a business under 50 people, a first project typically lands between £25,000 and £70,000, and a discovery engagement between £6,000 and £12,000. Below £25,000 you should expect a prototype rather than a production system, which can be exactly the right purchase for a first step.
Is a fixed price safer than day rates?
It transfers schedule risk at a premium — 16.6 per cent in the worked example above. It is the right choice when scope is genuinely stable. On a first project it rarely is, and fixed price then converts every discovery into a change request.
What proportion of the budget should be model costs?
Small. In the worked example, inference is about 0.7 per cent of first-year cost. If a supplier’s pricing is dominated by model spend, either the workload is unusually heavy or the design needs review.
How long before an AI project pays for itself?
Ask instead what proportion of the theoretical saving must convert into banked value. In this guide’s example that threshold is 56 per cent over three years. Projects fail this test far more often than they fail technically.
Can we do it without consultants?
Yes, if you already have data engineering, integration and evaluation skills and can spare them for three months. Most businesses of 50 to 250 people have one of the three. The genuine argument for AI consulting is not capability, it is not having to learn which mistakes are expensive at your own expense.
What should the first engagement be?
A discovery of 8 to 15 days that ranks candidates by value and feasibility, and produces a labelled test set for the winner. It costs £8,000 to £18,000 and routinely prevents a £100,000 mistake.
Do we need our own infrastructure?
Usually not. Running models in your own tenancy adds 10 to 25 days and a permanent infrastructure bill. Do it when a specific obligation requires it, not as a default posture.
What This Means for Your Budget
An AI consulting engagement is priced by three things: the state of your data, the number of systems it must touch, and how much certainty you require before go-live. Every other variable is secondary.
The practical sequence for a UK business of 50 to 250 people is unglamorous and reliable. Spend £8,000 to £18,000 on discovery and insist on a labelled test set as a deliverable. Try the vendor product route for anything not genuinely specific to you. Buy one delivery of 45 to 110 days from a supplier whose day mix puts a fifth to a third of the effort into data. Budget 30 to 45 per cent of the build cost every year to run it. Write the break-even realisation rate into the business case and give it an owner.
Do that and the AI consulting market becomes a normal market with normal comparisons. Skip it and you are choosing between two proposals you cannot read, which is how the 95 per cent figure gets made.
If you want a second opinion on a proposal already in front of you, our AI strategy and IT consulting teams review AI consulting scopes and day mixes against the benchmarks in this guide, and our data science practice can tell you within a week whether your documents are ready for the project you are being sold.
References and Further Reading
ONS — Artificial Intelligence in UK Businesses: 2023 to 2026
GOV.UK — AI Opportunities Action Plan: One Year On
DSIT — Artificial Intelligence Sector Study 2024
ITJobsWatch — UK AI Engineer Contract Rates
ITJobsWatch — UK AI Consultant Contract Rates
ITJobsWatch — UK Machine Learning Salaries
Machine Learning Times — Gartner: Over 40% of Agentic AI Projects Will Be Cancelled by 2027
ICO — Artificial Intelligence Guidance and Resources
ICO — Data Protection Impact Assessments
NCSC — Guidelines for Secure AI System Development
GOV.UK — Introduction to AI Assurance
NIST — AI Risk Management Framework
European Commission — Regulatory Framework for AI
EU AI Act — Implementation Timeline
Anthropic — Claude Model Pricing
Legislation.gov.uk — Data Protection Act 2018
The Alan Turing Institute — Research Programmes
Crown Commercial Service — Artificial Intelligence Framework
UK Government — AI Opportunities Action Plan Delivery Tracker
GOV.UK — Artificial Intelligence Playbook for the UK Government
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
