What UK Manufacturers Pay for AI in 2026: A £28.98 Model Bill and a £575 Day

A quote landed on my desk in June for an AI system to read job sheets. Six figures, four pages, one line of pricing: "Discovery, build and integration — fixed price." No unit costs. No per-document rate. No run cost. Nothing a finance director could challenge, because there was nothing there to challenge.

That is normal, and it is why buyers cannot tell a fair price from a bad one. Search "what should AI cost" and you get vendor marketing, US dollars and reports priced at £3,000 a copy. Almost nobody publishes what the components cost in sterling, on a date you can check.

So I did. I pulled live list prices from Microsoft's public retail pricing API, Amazon's offer files for the London region, Google's rate card, and OpenAI's and Anthropic's developer pricing pages. I pulled 974 Innovate UK awards from the UKRI Gateway to Research API and verified 25 line by line. I pulled UK contract day-rate medians with sample sizes. Then I built one worked example — 10,000 job sheets a month — and costed it nineteen ways.

The finding is not that AI is expensive. It is that the model bill is trivial and the delivery bill is everything, and almost every quote you receive prices those two things as one undifferentiated lump.

The price list nobody publishes in sterling

Almost every mid-market AI project I see involves getting structured data out of unstructured documents — job sheets, delivery notes, purchase orders, engineer reports. There are three serious managed services for this: Azure AI Document Intelligence, AWS Textract and Google Document AI. All three publish list prices. None publishes a UK-denominated comparison, and two do not publish in sterling at all.

Here is that comparison. Azure prices are its native GBP list for UK South. AWS and Google are converted at $1 = £0.74811, the official rate for August 2026 published by HMRC's Trade Tariff service (£1 = $1.3367).

Read the top and bottom of that table together. £1.12 per 1,000 pages at one end, £151.52 at the other — a 135x spread, inside a category buyers think of as one product. Nothing about the document changes. What changes is which meter the integrator wired you to.

Two traps sit in there. Azure's Query Pages meter is 133 times its own Read meter on the identical page. And Google's custom processor carries a standing charge of roughly £328 a year per deployed version, billed by the hour whether you send it a page or not — the only true fixed cost in the table.

Ten thousand job sheets a month

Unit prices only mean something attached to a volume. Here is one workload, costed every way I could price it.

The takeaway is uncomfortable for anyone selling extraction by the page: a general-purpose language model does this job for £28.98 a month, and the most-recommended managed alternative charges £374.06 for the same 10,000 documents — 12.9 times more.

That does not make the managed meters useless. They return page geometry, per-field confidence scores and table structure that a raw model call does not, and they behave deterministically. If you need bounding boxes to drive a human review screen, you are buying something real. But know you are paying a 13x premium and decide deliberately, rather than discovering it in month nine. It is the same trap I described in job sheet parsing from paper and PDF into your ERP: the technology choice gets made by whoever writes the first integration, and nobody revisits it.

Microsoft's private exchange rate

Here is the finding I did not expect, and have not seen published anywhere.

I pulled Azure list prices in GBP and in USD for the same meters, from the same public API, on the same day. Every sterling price is the dollar price multiplied by exactly 0.7576. Not approximately — exactly, on every meter I checked, from a fraction of a penny per million tokens to £136 per million.

The spread is small — 1.27% over HMRC's official rate. On a £28.98 monthly bill that is 37 pence. On a £250,000 annual Azure commitment it is £3,175 that no line item mentions.

The more important point is the shape, not the size. Your sterling price is pegged, not floating. If the pound falls ten per cent tomorrow, your GBP list price does not move — until Microsoft resets the peg, at which point it moves in one step, with no notice. AWS and Google avoid this by billing you in dollars and letting your bank take the spread. Neither approach is wrong. But if you are modelling five years of run cost in a board paper, "our cloud bill is fixed in sterling" is a claim with a hidden expiry date on it.

The bill that actually matters

Now the other side of the ledger. ITJobsWatch publishes UK contract day-rate medians from job advertisements, with sample sizes and a stated rolling window — the only free UK source I found giving a median, a sample count and a percentile spread together.

Take the machine learning median: £575 a day. That single day costs 19.8 months of the £28.98 model bill. A twenty-day month of one contractor costs £11,500 — thirty-three years of model calls.

Now build something. Assume a deliberately conservative 40 contractor days to take one document workflow from nothing to production: discovery, prompt and schema design, the ERP integration, an exception-handling screen, testing and handover. At the machine learning median that is £23,000. Against it, a full year of model calls is £347.76.

What it costs to keep running

The model is not the run cost either. Something has to host the thing.

That is the smallest stack I would put a production document pipeline on: one App Service instance to run the application, one managed PostgreSQL server for the extracted records and audit trail, and 64 GB of storage. £138.37 a month of infrastructure against £28.98 a month of model — the hosting is 4.8 times the model bill, and the model is 17% of what it costs to keep the lights on. Total run cost: £167.35 a month, £2,008 a year.

Put the halves together. A £23,000 build plus £2,008 of year-one running gives £25,008. The model is 1.4% of that. The thing everyone asks about in the first meeting — "but won't the API costs run away?" — is the smallest line in the budget by an order of magnitude.

Be clear about what that £23,000 is: my assumption, not a client figure. Forty days is one workflow with a clean data source and a cooperative ERP. The real UK evidence says it is usually more.

What a whole project costs

There is one public dataset of real UK AI project costs, and almost nobody reads it correctly. UKRI's Gateway to Research publishes every Innovate UK award, and for many of them the total project cost alongside the grant.

I pulled 974 Innovate UK projects from its API, filtered to genuine AI-and-manufacturing applications, and verified 25 against three independent surfaces each: the search API, the per-project API and the public HTML page. Zero mismatches.

Three things fall out of that table, and the first is a correction.

The Gateway to Research API field fund.valuePounds is the grant, not the project cost. GtR's own pages carry three separate labels — "Funded Value", "Project Cost" and "Grant Offer" — and the API exposes the first. On the largest project above, the five participants' grant offers sum to exactly £689,603, the Funded Value, while their project costs sum to £888,870. If you have ever quoted an Innovate UK award value as "the size of the project", it was £200,000 short.

Eight of the 25 sit on exactly 70.0% grant-to-cost. That is the published intervention rate for micro and small organisations on feasibility studies and industrial research, showing up in real award data. Live competition 2512, opened 1 July 2026, states it plainly: "up to 70% if you are a micro or small organisation; up to 60% if you are a medium sized organisation; up to 50% if you are a large organisation", falling to 45/35/25 for experimental development nearer to market. So for any award at 70%, divide the grant by 0.7 and you recover the true project cost — and know the company funded the other 30% itself.

And the common shape is far smaller than the headlines suggest. Fourteen of 25 awards are under £100,000 of grant, median total cost £54,654 — a single SME, around six months, roughly £45,000 to £50,000 all in. Ten are 100% funded, reflecting competition types that fund the whole thing; the current equivalent, competition 2513 (opened 7 July 2026), funds at 100% within a £25,000 to £50,000 band that matches the eight smallest projects almost exactly.

Set that against my 40-day, £23,000 build. The funded evidence says a real, scoped, delivered AI manufacturing project usually costs two to four times that. I would rather show both numbers than pretend the tidy one is typical.

Grants, tax, and the relief you probably will not get

Two things reduce the net cost, and one of them is oversold to the point of being a liability.

The genuine one is grant funding. Made Smarter offers matched funding for digital technology adoption — typically up to 50% of project cost, capped at £20,000 in the North West programme, with real variation between regions. Take a £60,000 project, sitting between the SME median (£54,654) and the all-25 median (£105,902): a £20,000 grant leaves £40,000 of cash out, and if that spend is revenue expenditure and you are a main-rate corporation tax payer with profits to absorb it, relief on the unfunded £40,000 is worth about £10,000. Net cost £30,000 — half of list. I have applied tax relief only to the unfunded portion because the grant is normally taxable income in its own right; the two broadly cancel. The full funding picture is in the Made Smarter guide and the 2026 UK AI grants directory.

The oversold one is R&D tax relief.

The stake is real. Under the merged R&D scheme the credit is 20% of qualifying expenditure, but it is taxable, so a main-rate payer nets about 15.0% and a loss-maker about 16.2%. Loss-making, R&D-intensive companies spending 30% or more of total expenditure on R&D can claim under ERIS instead, which GOV.UK describes as worth around £27 per £100 of qualifying spend. On £40,000 of unfunded project cost, 15% is £6,000 — worth having if you are entitled to it, and a number a claims firm will happily promise you on work that is not.

CIRD81960 does leave a door open, and it is worth knowing where. Integration work can qualify where "a competent professional in the field can't readily deduce how the separate components should be combined to have the intended function". If your project genuinely clears that bar, document it as you go. If it does not, budget as though the relief does not exist, because it probably does not.

Where the money actually gets wasted

Nothing in this post costs much. What costs money is the sequence you do it in.

Methodology, and what I could not verify

Every price here was read from a vendor's own machine-readable price feed or published rate card on 27 July 2026, not from a comparison site. Azure figures come from the public retail prices API, queried per meter for UK South in both GBP and USD. AWS figures come from the public offer file for AmazonTextract in eu-west-2, Google's from the published Document AI rate card, model prices cross-checked against the vendors' own developer pricing pages. Grant data comes from the UKRI Gateway to Research API, triple-verified per project. Day rates come from ITJobsWatch contract pages, each carrying its own sample size.

What I could not verify, and left out. I went looking for Crown Commercial Service day rates to corroborate ITJobsWatch and abandoned them: the G-Cloud 14 SFIA rate cards I found are 2024-vintage, supplier-declared ceiling rates, not market medians with samples, and printing them alongside ITJobsWatch would imply a comparability that does not exist. I also dropped every "average AI project costs £X" figure in circulation — each one I chased traced back to a vendor survey with no published sample, no UK breakdown, or no methodology. The Innovate UK data replaces them with 25 awards you can check yourself.

Three limits on what I kept. Company size is not published per Innovate UK award — eight projects landing on exactly 70.0% is strongly consistent with the micro/small rate, but that is my inference from the arithmetic, not a retrieved field. ITJobsWatch samples are advertised rates, not paid rates, on a rolling six-month window that moves daily. And Azure Content Understanding, a newer Microsoft line at £0.7576 per 1,000 pages for basic extraction, undercuts everything in Exhibit 1 but is a different product with a different output, so I left it out of the 135x spread rather than flatter the range.

Everything modelled rather than retrieved is labelled as such: the token profile, the 40-day build, the hosting stack, the £60,000 tax example. None are client results.

What to do on Monday morning

Ask every quote for three numbers. Cost per document processed, monthly run cost at your actual volume, and total build days with a day rate attached. A supplier who cannot produce those three from a fixed-price proposal has not costed the work; they have priced your budget. Exhibits 1, 2 and 4 are enough to check whether the answers are sane.

Count your documents before you value the project. Not an estimate — a count. How many job sheets, delivery notes and purchase orders crossed a desk last month, and how many minutes did each take a human? That is the number the business case turns on, and nobody has it. Our Hidden Waste Audit produces exactly that in about five minutes.

Split build from run in your own budget. Model spend belongs in an opex line you watch weekly; build spend in a capex-style line with days and a rate. Merged, they hide the only thing worth managing — whether the delivery estimate was honest.

Do not start with the model. The technology decision here is worth about £30 a month. The process decision is worth twenty times more, which is the same conclusion I reached comparing AI against hiring another head. If your process is unclear, a better model will not clarify it.

Frequently Asked Questions

What does it cost to run AI document extraction on 10,000 pages a month in the UK?

Between £2.20 and £1,515.21 a month, depending entirely on the route. A general-purpose language model doing structured extraction costs £28.98 a month on gpt-5.4-mini at Azure's UK South list price. Managed per-page services range from £11.22 for plain OCR to £374.06 for AWS Textract Forms and £1,515.21 for Azure's Query Pages meter. All are August 2026 list prices for 10,000 single-page documents at 3,000 input and 350 output tokens each.

Why is Azure's price in pounds not the market exchange rate?

Because it is not an exchange rate. Every Azure GBP list price I checked is the USD price multiplied by exactly 0.7576 — a fixed internal conversion. That is 1.27% above HMRC's official rate for August 2026 ($1 = £0.74811) and 0.94% above the ECB market rate on 24 July 2026. The consequence is that your sterling price is pegged rather than floating: it does not move when the pound moves, then moves all at once when Microsoft resets the peg.

What is the median UK contract day rate for AI and machine learning work in 2026?

£575 a day for machine learning skills (1,026 quoted rates) and £550 for artificial intelligence (2,866 quoted rates), on ITJobsWatch data for the six months to 27 July 2026. Data engineers sit at £500, data scientists at £600, solution architects at £625. The quartile spread on machine learning runs £492 to £695. These are advertised rates, not paid rates.

How much does a complete AI project cost a UK manufacturer?

Across 25 verified Innovate UK AI manufacturing awards, the median total project cost is £105,902 and the median grant £99,040. Among the 14 awards under £100,000 of grant, the median total cost is £54,654 — typically a single SME running about six months. One narrow workflow can be delivered for less: 40 contractor days at £575 is a £23,000 build, but that is modelled, and the funded evidence suggests real projects run two to four times larger.

Does AI integration work qualify for R&D tax relief?

Usually not. HMRC's CIRD81960 states that configuring existing software to a company's own requirements is unlikely to qualify, and Guidelines for Compliance GfC3 gives a worked example of adapting natural language processing using open-source methods, concluding "the costs of this work are not R&D for tax purposes". Being first in your company is not an advance in science or technology. Where work does qualify, the merged scheme credit is 20% gross, about 15.0% net to a main-rate payer.

Is the cheapest route always the right one?

No, and the gap is smaller than it looks. On a £25,000 year-one budget the difference between the cheapest and a mid-priced extraction route is a few hundred pounds. Managed services buy bounding boxes, per-field confidence scores and deterministic behaviour a raw model call does not give you, which is worth paying for if you need a human review screen. What moves the budget is not which meter you pick but how many contractor days go by before anyone looks at a document.

Related Reading

• Made Smarter AI Funding Guide for UK Manufacturers
• UK AI Grants for Manufacturers: The 2026 Directory
• Why AI Projects Fail Without Process Redesign
• Job Sheet Parsing: From Paper and PDF into Your ERP
• AI vs. Hiring Another Sales Rep: The Maths Your CFO Needs to See
• Choosing an AI Consultant: A Guide for UK Manufacturing SMEs
• The Hidden Cost of Manual Processes in UK Manufacturing

Want to know what your own document volume is actually worth? Our Hidden Waste Audit puts a number on the admin hours, double entry and rework in your operation in about five minutes. Or book a 30-minute call and bring a quote — I will tell you which line is the one to argue about.

Sources, all retrieved 27 July 2026: Azure Retail Prices API; AWS Textract price list, eu-west-2; Google Document AI pricing; OpenAI API pricing; Anthropic model overview; HMRC Trade Tariff exchange rates, August 2026; UKRI Gateway to Research; Innovate UK competition 2512 (opened 1 July 2026); HMRC CIRD81960; HMRC Guidelines for Compliance GfC3, Part 4; ITJobsWatch UK contract rates; Made Smarter.