Artificial Intelligence and Automation

IT IS NOT COMING FOR THE LADDER

Artificial intelligence, automation, and the half of this business nobody pays for

Shining Windows · Article Five of Six Written 15 September 2026 · Northampton

There is a kind of tiredness that has nothing to do with lifting anything.

It arrives at about half past eight in the evening. The van is on the drive, the poles are rinsed and racked, the tank is topped for the morning. And a laptop opens on a kitchen table, and the second half of the working day begins.

Quotes that were promised this week. An invoice that needs chasing for the third time, rewritten twice so it does not sound like an accusation. A message from somebody who wants a price for a conservatory roof from a photograph. Bank reconciliation. The diary for next week, which depends on weather nobody can see yet. Somewhere in there, a set of records that now has to be kept in software because the law says so.

None of it is paid. All of it is the business.

If you read what is written about artificial intelligence and small firms, you are told that a displacement is coming for manual work. What you are rarely told is that almost everybody writing that sentence has a job in which the physical work and the paperwork belong to two different people.

For a one-van business they are the same person. The man on the ladder at two in the afternoon is the man at the kitchen table at nine at night.

That is the whole subject of this article, and getting to it honestly meant throwing away most of what we started with.

ONE · THE OFFICE IS NOT OPTIONAL ANY MORE

Start with the part nobody argues about, because it is law.

The usual framing treats digital administration as a choice — a thing forward-looking businesses do to find efficiencies. For a VAT-registered trade in the United Kingdom that has not been true since 2019, and as of this year it is not true for a great many sole traders who are not VAT-registered either.

Making Tax Digital for VAT. The power sits in section 62 of the Finance (No. 2) Act 2017, headed Digital reporting and record-keeping for VAT, which amends Schedule 11 to the Value Added Tax Act 1994. The detail was delivered by the Value Added Tax (Amendment) Regulations 2018, SI 2018/261, in force 1 April 2019. That instrument inserted the phrase that governs the whole thing into regulation 24 of the VAT Regulations 1995:

"functional compatible software"

It means a program, or set of programs, that records and preserves records electronically, sends information to HMRC through its API platform, and receives information back. Regulation 32A(5) puts it plainly: the electronic account must be kept and maintained using it. The Value Added Tax (Amendment) Regulations 2021, SI 2021/986, in force 1 April 2022, removed the below-threshold exemption and brought every VAT-registered business in, whatever its turnover.

Exemptions survive, and they are narrow: insolvency, existing exemption from online filing, and on application — age, disability, remoteness of location, or religious belief. HMRC's published position is that finding the software difficult, or expensive, is not a ground.

Making Tax Digital for Income Tax. This is the one that matters to a sole trader who is not VAT-registered, and it started five months ago.

From 6 April 2026, anybody whose qualifying income was over £50,000 in the 2024–25 tax year is in. Qualifying income is gross self-employment plus property income, before expenses, taken from the return for the tax year two years before the start date. The threshold falls to £30,000 from 6 April 2027 and £20,000 from 6 April 2028. The statutory basis is The Income Tax (Digital Obligations) Regulations 2026, SI 2026/336.

Once you are in it: digital records in compatible software, four cumulative quarterly updates per business per year — due 7 August, 7 November, 7 February and 7 May — and then the return itself through the software by 31 January. Records kept five years after that deadline.

Read the threshold again, because it is widely misunderstood. £20,000 of gross income before expenses, by April 2028. A window cleaner turning over four hundred pounds a week, with a van, fuel, chemistry, insurance and a pole system to pay for out of it, is inside that.

There is a grace period on the quarterly updates for 2026–27, and then a points-based penalty system: four points and a two hundred pound charge, then two hundred pounds per further default.

Partnerships are not in scope at all, and gov.uk gives no start date for them.

So before any question about whether these tools are useful, there is a prior fact. The state now requires a sole trader to operate a piece of business software and file five times a year. The office is not an ambition or an overhead to be trimmed. It is a statutory obligation with a penalty regime attached, and it sits on the same person who carried the ladders.

And the burden is officially recognised, just not in the way we were told

There is a claim circulating that Parliament legislated to protect the smallest businesses from exactly this, in section 33 of the Small Business, Enterprise and Employment Act 2015, said to mandate a Small and Micro Business Assessment for every new regulation.

We checked it, because it would have been a good line.

Section 33 does nothing of the kind. Its heading is Definitions of small and micro business. It supplies the headcount tests — under fifty, under ten — for secondary legislation that chooses to use those terms. It imposes no duty on anybody to assess anything.

The Small and Micro Business Assessment is real, but it is not statutory. It is a requirement of the Better Regulation Framework guidance published by the Department for Business and Trade, scrutinised by the Regulatory Policy Committee, and it applies only to domestic measures with a business impact above a ten million pound threshold. Its published starting point is worth quoting exactly, because it is more generous than most people in this trade would guess:

"The starting point for consideration should be exempting small and micro-businesses from the requirements of new Regulatory Provisions."

And the genuinely statutory better-regulation machinery that did exist has been dismantled. Sections 21 to 27 of the 2015 Act, which created the business impact target, were repealed on 29 August 2023 by section 18 of the Retained EU Law (Revocation and Reform) Act 2023 — a section whose own heading is Abolition of business impact target. What survives is section 28, a duty to review regulatory provisions in secondary legislation.

We are not making an argument out of that. We are recording it because we went to check a claim that suited us, found it was wrong, and would rather print the correction than the claim.

TWO · WHOSE JOB IS ACTUALLY AT RISK, AND WHY WE CANNOT TELL YOU PRECISELY

Here is where this article originally went wrong, and the correction is more interesting than the error.

The first version of this page said that skilled trades have fewer than ten per cent of their tasks automatable, and that bookkeepers, telephone salespeople and bank clerks face eighty-seven to ninety-four per cent risk. It built its central sentence on those two numbers.

Neither figure has a source. We went looking and could not find one.

What does exist is the Office for National Statistics study Which occupations are at highest risk of being automated?, published 25 March 2019. Its actual figures are these:

Occupation — Probability of automation

Bank and post office clerks — 58.9%

Telephone salespersons — 57.6%

Book-keepers, payroll managers and wages clerks — 56.4%

Plumbers and heating and ventilating engineers — 52.1%

Electricians and electrical fitters — 50.7%

Look at the spread. It is six points, not eighty. On the most-cited UK source in the field, the admin jobs and the trade jobs sit almost on top of each other — and ONS does not classify any of them as high risk, a category it reserves for probabilities above seventy per cent.

That study was published in March 2019, before generative AI existed in any form the public could use. It is a measure of automation in the older sense — machinery, software, process — and it was never designed to answer the question everybody is now asking it.

So we withdraw the numbers. What we can put in their place is sourced, and it survives contact with the evidence.

First, the scale of exposure. The Department for Science, Innovation and Technology, with the AI Security Institute, published Assessment of AI capabilities and the impact on the UK labour market on 28 January 2026. It reports, on International Monetary Fund estimates, that around seventy per cent of UK workers are in occupations containing tasks that AI could potentially perform or enhance — higher than the United States, at around sixty. It then adds two qualifications that almost never travel with the figure:

About half of those exposed are in roles of high complementarity, where the technology assists rather than replaces.

Such indices "have not been extensively validated against real-world outcomes, and should be interpreted as indicative rather than precise."

Second, the shape of the least exposed. The Department for Education's 2023 analysis of AI's impact on UK jobs and training names the occupations at the bottom of the exposure ranking. They include roofers, roof tilers and slaters, elementary construction occupations, plasterers, and steel erectors. The reason it gives is the useful part:

more manual work that is technically difficult, in unpredictable environments, and with lower wages — reducing the incentive to automate.

That last clause is the honest one and nobody enjoys it. Part of why this trade is safe is that it is difficult and unpredictable. Part of why this trade is safe is that it is not worth enough to anybody to bother.

Third — and this is the one that actually carries the argument — what businesses are using it for.

ONS asked. In June 2026, by technology: large language models eighteen per cent, visual content creation sixteen, data processing using machine learning twelve, image processing using machine learning six.

The dominant business use of artificial intelligence in the United Kingdom is language. Writing. Reading. Summarising. Corresponding.

And that is not the work. That is the office.

So the sentence this article rests on is no longer a claim about two percentages. It is a claim about two halves of one person's day, and it is better for the change:

Nothing in the evidence suggests a machine is coming for the ladder. Everything in the evidence says the tools are landing on the desk — and in a one-van business, the desk is the same man, later, unpaid.

THREE · HOW MANY BUSINESSES ACTUALLY USE IT

We went looking for a straight answer and found three.

The Department for Science, Innovation and Technology, surveying 3,500 UK businesses by telephone between 12 February and 2 May 2025 and publishing on 28 January 2026, reported around one in six — sixteen per cent — currently using at least one AI technology.

Firm-level data for the end of 2024 gave roughly sixteen to twenty per cent of businesses with ten or more employees.

The British Chambers of Commerce, with Atos, said fifty-four per cent in March 2026. Atos sells technology services, which should be stated whenever the figure is used.

Sixteen to fifty-four. We published all three and declined to choose, because we could not separate them honestly.

Then the Office for National Statistics published the answer.

Its article Artificial intelligence in UK businesses: 2023 to 2026, released 20 July 2026 and built on the Business Insights and Conditions Survey, gives this:

Finding — Figure

Self-reported AI use, businesses with 10 or more employees, late 2023 — around 12%

Same measure, 2026 — around 35%

Of those using at least one AI technology, share using it extensively — only 10%

The spread was never a disagreement. It was two things at once.

Time. The figure moved from twelve per cent to thirty-five between late 2023 and 2026. A survey fielded in early 2025 and one fielded in March 2026 would legitimately differ by a wide margin. Neither is wrong.

Depth. Thirty-five per cent use at least one AI technology. Only ten per cent of those use it extensively — roughly three and a half per cent of all businesses.

"Using AI" and "using AI extensively" differ by a factor of ten, and almost every headline conflates them.

So the honest answer is: about a third of UK businesses with ten or more employees touch it, and about one in thirty uses it seriously.

We are recording how we got there because the method is the point. Publishing all three figures and refusing to pick kept the question open until a neutral source arrived that explained why they differed. Choosing early would have produced a confident wrong answer.

Construction

ONS puts construction at thirteen per cent, against fifty-eight per cent in information and communication. An exterior trade sits near the bottom of the national adoption table.

An earlier version of this page said DSIT put the same sector at around twelve, and described the two as independently confirming each other. That was our arithmetic, not DSIT's finding. What DSIT actually publishes is the other side of the coin: eighty-eight per cent of construction businesses neither use AI nor have any plans to. Subtracting that from a hundred and calling the remainder an adoption rate is not the same thing, because the remainder includes firms that are planning to adopt and have not yet. The corrected figure is above; the inference has been removed.

FOUR · THE ONE ADVANTAGE A SMALL FIRM ACTUALLY HAS

There is a finding in the ONS data that runs the other way, and it is the best news in this article.

Adoption rises with size, as you would expect. Twenty-eight per cent of businesses with nought to nine employees use AI, against forty-nine per cent of those with two hundred and fifty or more.

Then depth of use inverts it.

Seventeen per cent of businesses with nought to nine employees report using AI extensively, compared with nine per cent of those with two hundred and fifty or more.

The smallest businesses in the country adopt least often, and when they do adopt they go nearly twice as deep as the largest.

The mechanism is obvious once stated. A large organisation needs a steering group, a procurement route, a budget line, a pilot, a compliance review and a training programme to change how one email template is written. A business of one or two people needs a person on a laptop on a Tuesday evening deciding to do it differently.

The corporation has scale. The sole trader has velocity.

And when the time comes back, it comes back whole. A large firm that halves an administrator's workload still pays that administrator for a full week, because its structure has no way to convert the saving into anything. A sole trader who saves five hours a week has five hours. That is the entire chain of command.

One caveat, from ONS itself. The nought-to-nine band carries disproportionately large survey weights, which is why some ONS estimates exclude it altogether. The seventeen per cent is published and citable. It should not be leaned on as though it were as robust as the headline series.

FIVE · THE FINDING THAT UNDERCUTS THE PROMISE

Every small business in the country is being told this will transform it.

Among UK businesses that have actually adopted it, seventy-seven per cent report no change in revenue yet. Just over one in ten — twelve per cent — report an increase. That is DSIT, from a base of seven hundred businesses currently using AI.

That is not an argument against using the tools. It is an argument against the promise being made about them, and it is precisely the kind of fact a page built on evidence should publish, because it cuts against the interest of everyone selling the tools, the training and the consultancy.

And then a second figure, apparently contradicting it

Research from the University of St Andrews reports productivity gains of between twenty-seven and one hundred and thirty-three per cent for small and medium businesses adopting AI.

Seventy-seven per cent see no revenue change. Twenty-seven to one hundred and thirty-three per cent productivity gain. Both cannot be describing the same thing.

We published both. Then we went and read what the study actually says.

The paper is "Can We See it in the Numbers Yet? Artificial Intelligence, Firm-Level Productivity and SMEs" by Ross Brown, Haoran Sun, Marc Cowling and David Dowell. Brown, Sun and Dowell are at St Andrews; Cowling is at Oxford Brookes. The university describes it as an ESRC-funded study, and it is built on the Longitudinal Small Business Survey — the Department for Business and Trade's own panel, interviewing just under ten thousand UK small and medium enterprises under a stratified random sampling design for representativeness by region, sector and firm size.

One thing has to be said before any of its numbers are used. It is a working paper. It was posted to SSRN in April 2025 and, seventeen months later, it has not been peer-reviewed and has not appeared in a journal. That does not make it wrong. It makes it provisional, and a page that demands parentage for other people's figures has to declare that about its own.

And it contains the reconciliation. The authors' own words:

"firms with lower unobserved productivity potential are more likely to adopt AI technologies"

That is the answer.

A gain of twenty-seven to one hundred and thirty-three per cent measured from a low base is entirely compatible with no change in revenue at all. A business operating at a quarter of its sector's productivity which moves to half of it has more than doubled — and may be earning exactly what it earned last year.

The two figures were never in conflict. They were measuring different firms at different points on the same curve. Both are true. Neither body got it wrong.

We only know that because we published both and then went and read the second one, instead of choosing between them at the start.

Two caveats that have to travel with it

The sector finding. The benefits are sector specific, and the university names catering and hospitality as prominent among the beneficiaries. Construction is not among the sectors highlighted. So nothing here entitles an exterior cleaning business to claim a productivity dividend.

And a wording correction. The university's press release says the paper found it is firms with "lowest productivity levels" that are most likely to adopt. The paper's own abstract says "lower unobserved productivity potential" — an econometric residual, not the same statement. We have quoted the authors rather than the press office, and where those two disagree, that is the rule.

What it means at a kitchen table

Strip the econometrics out and the finding is this.

If quotes, invoices, scheduling and correspondence take fifteen hours a week and a tool reduces that to five, administrative productivity has tripled. If the diary was already full, and the day still ends when the light does, revenue moves by nothing at all.

The constraint on a window cleaner's income is not how fast he writes a quote. It is how many houses he can physically reach before dark, and how many of those will pay.

So the technology does not buy the ladder. It buys back the evening. That is a smaller claim than the one being sold, and it happens to be true, and on the evidence above it is the only claim this trade is entitled to make.

SIX · THE GROWTH THAT IS NOT YOURS

On 11 September 2026 the ONS reported that monthly GDP grew by 0.4 per cent in July. Much of the coverage attributed it to an artificial intelligence boom.

The ONS did not say that. What it said, in full, was this:

"many of the businesses reporting the largest turnover in July 2026 are involved in activities related to artificial intelligence and cloud computing. However, because of the nature of our data collection, it is difficult for us to quantify the exact impact of these types of activities on turnover."

That is a careful sentence and it was widely rewritten into a less careful one. We are quoting it because the difference matters and because we have twice now been caught tidying a hedged official statement into a confident one.

Read it with the seventy-seven per cent alongside it and a shape appears. The turnover is being recorded among the businesses selling the technology. The businesses buying it mostly report no revenue change yet.

The economy can grow because small businesses are buying tools without those tools yet making small businesses richer. Both things can be true at once, and for the moment they are.

That is the same shape this series keeps meeting. The International Monetary Fund describes inflation as quietly reallocating real income away from the households least able to absorb the loss. The adoption data describes a technology whose gains accrue first to the organisations most able to capture them.

Same shape. Different mechanism. The advantage moves upward in both.

That is worth knowing about the one subject everybody assumes is a leveller.

SEVEN · THE MONEY AIMED AT THIS SECTOR, AND WHY NONE OF IT ARRIVED

Two public programmes were pointed at exactly this problem. The story of both is worth telling.

Innovate UK BridgeAI

A one hundred million pound programme, launched publicly in April 2023, with the Alan Turing Institute — the United Kingdom's national institute for data science and artificial intelligence — among its delivery partners, alongside BSI, the STFC Hartree Centre, Digital Catapult and Innovate UK Business Connect.

Its purpose was to drive AI adoption in sectors with high growth potential and low take-up. It named four priority industries.

Transport. Construction. Agriculture and food processing. The creative industries.

Construction — thirteen per cent adoption, near the bottom of the national table — was one of four named priorities for a nine-figure national programme.

An earlier version of this page said BridgeAI closed in June 2026. That was wrong. At the Autumn Budget 2025 the government committed to expanding it from four Industrial Strategy sectors to eight, and gov.uk was still writing about "Innovate UK's expanded BridgeAI programme" in the present tense in January 2026. We had it backwards and we are correcting it here.

What is live and free today, for anybody who wants it:

•  The AI Skills for Business Competency Framework, now at version three, published openly and free to download. It tells a business what skills it actually needs before it buys anything. Registration is open indefinitely.

•  Capsules of AI Knowledge — short videos, around three minutes each, aimed at common adoption problems.

•  Bespoke advice for small and medium businesses from the Turing Institute's scientific advisers. The stated offer has been up to sixteen hours of free one-to-one guidance, and the programme reports having supported over a hundred and fifty organisations. The application page currently shows a closing date that has passed while other pages describe it as continuing, so check before relying on it.

Independent. Non-commercial. Publicly funded. Free at the point of use. It is the exact opposite of the material that fills the space instead, and almost nobody in this trade knows it exists.

The AI Upskilling Fund, and a lesson in who these things are for

The second programme is a harder story and we think it should be told plainly.

The Flexible AI Upskilling Fund was announced as a £7.4 million pilot, offering to match up to fifty per cent of the cost of AI training for small and medium businesses — capped at two and a half thousand pounds for a micro business.

It was open to businesses in the Professional and Business Services sector, defined by a closed list of SIC codes: divisions 69 to 74, 77, 78 and 82. The published guidance is unambiguous — "Other codes are not eligible."

Cleaning activities are SIC division 81. Window cleaning is 81221.

It also required a business to "employ between 1 to 249 employees in the UK", which on its face excludes a sole trader with no staff.

So a window cleaner was ineligible twice over: wrong sector, and probably no employees.

Applications originally closed on 31 May 2024 and were extended to 18 August 2024. DSIT's own evaluation, published 28 January 2026, records what happened. Of £7.4 million announced, £381,096 was actually awarded — approximately fifteen per cent of the original target. Five hundred and fifty-three applied, three hundred and twenty-seven succeeded, a hundred and eighty-one claimed. The planned randomised trial was abandoned as underpowered.

And the evaluation's own first recommendation names the cause: focusing on professional business services "limited the potential pool of eligible businesses."

We are not going to turn that into a grievance. It is more useful as a fact. Public money aimed at AI skills for small business was written around offices, and the department that wrote it has said so in print. Anybody in a trade reading advice to "apply for the government AI fund" should know it was never open to them and is now closed.

EIGHT · THE THING YOU CANNOT HAND TO A MACHINE

Article Four of this series set out section 50 of the Consumer Rights Act 2015, which is the most important provision most tradespeople have never heard of. It matters here for a specific reason.

Section 50 provides that every contract to supply a service is treated as including, as a term of the contract, anything said or written to the consumer by or on behalf of the trader about the trader or the service, where the consumer takes it into account in deciding to enter the contract, or in any decision about the service afterwards.

Two qualifications that must travel with it, because they usually do not. Under section 50(2)(a), what is said is subject to anything that qualified it and was said to the consumer on the same occasion. And under section 57(2), subject to that, a term excluding liability arising under section 50 is not binding on the consumer.

Now put a chatbot on the front of a cleaning business.

A customer asks whether the chemistry is safe around a dog. The machine, which does not know and cannot know, produces a confident, pleasant, entirely invented answer. The customer books on the strength of it.

The statutory hook is the phrase "by or on behalf of the trader". A language model is not an agent in law — it has no legal personality and forms no intention — so the question is one of attribution: the trader chose to put the thing on its own website, held it out as the way to get information about its services, and the output is therefore the trader's own communication. On that footing section 50 does the rest automatically, and section 57 stops the trader excluding the consequence in its terms.

We should be careful here, and we will be. There is no UK appellate decision on chatbot statements under section 50. The Canadian case people cite is a British Columbia tribunal decision and is not authority here. What we are describing is how the wording of the section would be likely to apply, not settled law, and nothing on this page is legal advice.

But the practical point does not depend on resolving any of that.

A large company whose chatbot invents a promise has a legal department. A sole trader whose chatbot invents a promise has a problem that can end the business. The asymmetry is not in the law. It is in who can afford to be on the wrong side of it.

Which produces a rule we can state without hedging, because it is our rule rather than a legal opinion:

Use it for the diary. Never for the science.

Nothing that makes a promise to a customer about a chemical, a surface, a warranty, an outcome or a price leaves this business without a person having written it or checked it. Not because a machine could not produce a plausible sentence. Because a plausible sentence is exactly what section 50 turns into a contractual term.

NINE · THE PART OF ARTICLE THREE THAT BELONGS TO THIS SUBJECT

The third article in this series sets out the enforcement volumes the review platforms publish, what the regulator has done about fake reviews, and who is currently under investigation. Those numbers live there and we are not going to reprint them here.

One thing about them belongs in this article rather than that one, and only one.

Both sides of that fight are now automated.

Tripadvisor states in its own transparency reporting that generative AI is having a noticeable impact on what gets submitted to the platform, and reports AI-generated reviews as a category it now removes separately. Google, describing how it handles the volume, credits its own machine learning with doing the removing.

So the position in 2026 is a machine writing the accusation and a machine deciding whether to take it down, with the business it concerns present at neither end.

What changed is not the existence of fake reviews. It is the price of one. Producing a plausible, specific, human-sounding account of a job that never happened used to take a person twenty minutes. It now takes nobody any time at all. Anything whose marginal cost falls to zero gets produced in volumes that have nothing to do with how often the underlying event occurs.

That is worth stating precisely because of what it does to the meaning of the number on the page. A star rating was always a weak instrument — Article Three sets out exactly how weak, and why. What automation adds is that the rating now also reflects who was willing to generate text, which is not a fact about the business at all.

One note of caution about a figure that circulates in this area, including on a government website. A widely repeated claim that up to half of online reviews are fake traces to a commercial detection vendor citing its own white paper. We are not repeating it, and the rule does not stop applying because the number turned up somewhere respectable.

And this is where the Forensic Case Log stops being a nice idea and becomes the answer. A language model can write a review of a job that never happened. It cannot produce a dated record of what was used, at what dilution, onto what substrate, in what conditions, with photographs taken at the property.

The defence against automated fiction is not a better argument. It is a contemporaneous record.

TEN · WHAT IT COSTS TO RUN

Artificial intelligence runs inside buildings, and those buildings drink.

We found the figures while researching this article and they turned out to belong somewhere else, so that is where we put them. The first article in this series sets out what UK data centres consume in electricity, what a single hyperscale facility consumes in water, and the fact that nobody is required to report the water at all.

We have re-checked that last point for this article and it holds. As at September 2026 there is no UK statutory or regulatory provision requiring data centres to report water consumption. Everything that exists is proposal: the Defra-convened report recommends mandatory reporting above about one megawatt — which is itself proof that none exists — and the Environment Agency's own national framework records that it is having difficulty obtaining the information.

Set against that, the offence a man with a tank on a van can commit is section 118(5) of the Water Industry Act 1991, discharging trade effluent into a public sewer without consent. One correction to what is usually printed about it, including by us: the section text still says "not exceeding the statutory maximum", but that £5,000 cap was removed by section 85 of the Legal Aid, Sentencing and Punishment of Offenders Act 2012 for offences committed on or after 12 March 2015. The fine is unlimited. The section has never been amended to say so.

It sits in Article One rather than here because the comparison that makes it matter is with water and custody rather than automation. If you read only one section of this series on the physical cost of this technology, read that one.

ELEVEN · WHAT THIS BUSINESS ACTUALLY DOES

We are not going to pretend to neutrality. We use these tools, and we should say how.

They are used for the office. Research, drafting, checking, structuring, correspondence, and reading long documents we would otherwise not have time to read. This article is an example: three government publications, an academic working paper, a statutory instrument and two pieces of primary legislation were read to write it, and the reading was made affordable by the tools.

They are used to argue with us. Every figure in this article was checked against its primary source, and the checking is what removed most of what we started with. Two of the corrections in section twelve below were found that way.

They are not used to decide anything about a job. No measurement, no assessment, no method and no outcome in the Forensic Case Log is generated by a machine. Those come from a person standing in front of a building.

And nothing in the case log is written by one. The record of what was done at a specific property on a specific date is written by the people who did it, because the whole point of a contemporaneous record is that somebody was there.

If that distinction ever stops being true, we will say so on this page.

TWELVE · WHERE WE WERE WRONG

The first version of this article was published on 14 September 2026. Checking it against primary sources for this version produced the following. All of it was our error.

The automation percentages. We printed "fewer than ten per cent" for skilled trades and "eighty-seven to ninety-four per cent" for bookkeepers, telephone salespeople and bank clerks. Neither figure is traceable to any published study. The ONS 2019 figures are 50.7 to 52.1 per cent for electricians and plumbers and 56.4 to 58.9 per cent for the clerical roles — a six-point gap, not eighty. Both numbers are withdrawn and section two is rebuilt on sourced material.

The displacement estimates. We referred to "seven-to-eleven per cent displacement estimates". We can no longer find a source for them. The nearest real figures are PwC's 2021 work for BEIS — around seven per cent of existing UK jobs at high probability of automation over five years — and IPPR's separate finding that about eleven per cent of tasks are exposed. Those are different concepts over different horizons and should never have been welded together. Withdrawn.

BridgeAI. We said it closed in June 2026. It was expanded, from four sectors to eight, at the Autumn Budget 2025. Corrected in section seven.

The construction comparison. We said ONS and DSIT independently gave construction thirteen and twelve per cent. The twelve was our own subtraction from DSIT's published eighty-eight per cent "neither use nor plan to use", and it is not the same measure. Corrected in section three.

The micro-business staff figure. We said thirty-eight per cent of staff use AI in adopting micro businesses. DSIT's figure is for micro businesses with five to nine employees specifically. Since it does not describe a one or two person firm at all, it has been replaced in section four with the ONS extensive-use finding, which does.

The productivity paper. We described it as academic research without saying that it is a working paper which has not been peer-reviewed seventeen months after posting. That should have been said the first time. It is said now.

The Water Industry Act penalty. Where this series has described the trade effluent offence, the statutory maximum is out of date. The fine is unlimited.

Seven corrections in one article is a lot. We would rather print them than have them found.

THIRTEEN · WHAT WE REFUSED TO PRINT

A quantity of material was offered for this article and did not survive checking. It is listed because the list is more useful than the material.

A claim that Making Tax Digital rests on the Finance Act 2019 and the Value Added Tax (Amendment) Regulations 2022, SI 2022/222. The primary power is section 62 of the Finance (No. 2) Act 2017. The instrument is SI 2018/261, extended by SI 2021/986. SI 2022/222 is the M48 Motorway (Severn Bridge Weight) (Temporary Restriction of Traffic) Order 2022, a Department for Transport order about lorries on the Severn Bridge. It has nothing whatever to do with VAT.

A claim that section 33 of the Small Business, Enterprise and Employment Act 2015 mandates the Small and Micro Business Assessment. Section 33 defines "small business" and "micro business" and does nothing else. SaMBA is guidance, not statute. Set out in section one.

A claim that Pimlico Plumbers Ltd v Smith [2018] UKSC 29 establishes that an independent trader is defined by command over invoices, tools and administration. The case is real — Supreme Court, 13 June 2018, Lord Wilson giving judgment. It decides the opposite of what was claimed. Pimlico's appeal was dismissed and Gary Smith was held to be a worker under section 230(3)(b) of the Employment Rights Act 1996, not an independent contractor. He presented as self-employed for income tax and VAT, turned over around £131,000, and supplied his own tools — and none of that prevented worker status. It is an employment-rights case. It imposes no record-keeping obligation on anybody. It is not cited in this article.

A claim that the ONS attributed July's GDP growth to an AI boom. It did not, and its actual wording is quoted in section six.

A government productivity figure given as 1.5 per cent annually, worth up to £47 billion over a decade. The published figure is up to an average £47 billion each year over a decade — the version offered to us understated it by roughly a factor of ten. It is also not properly a government estimate: the notes to editors state it is based on IMF figures but derived from separate calculations and is not an official IMF estimate. We have not used it, because a projection we cannot interrogate is not evidence, however large.

A source given as "HeyBRB Verified UK AI Datasets". No dataset of that name exists. HeyBRB is a commercial AI consultancy. Not a citable source, and the fact that it was offered as one is the clearest possible argument for checking.

A reading of the St Andrews paper claiming the 27 to 133 per cent range applies specifically to bottom-quartile firms while top-decile firms saw no gain. That does not appear in the abstract or in either university release, and the abstract attaches the range to adopters against non-adopters. It may be somewhere in the full paper. Until somebody produces the page, we are not printing it.

A claim that up to half of online reviews are fake. Set out in section nine. Vendor-sourced.

And a general note. Several widely quoted AI commentators were considered and left out. They are prolific and often interesting. They are also, in most cases, selling something into the market they are describing.

FOURTEEN · WHAT WE THINK

The honest position for an exterior trade is not AI is coming for us and not AI is irrelevant to us.

It is that the half of this business that artificial intelligence is transforming is the half nobody pays for.

The work is bounded by ladders, daylight, weather and the length of a road. No tool moves those. A model cannot judge whether a lime mortar joint will survive being cleaned, cannot feel a pole loading in a gust, cannot decide that a wall should be left alone. The physical world stays stubborn.

What the tools move is the evening. The books, the correspondence, the scheduling, the quoting, the reading — the second shift that the state has now made compulsory and that nobody has ever paid for.

That is a smaller promise than the one being made in public. It is also, on the evidence, the only one this trade is entitled to make. And for a person who has been doing the books at nine at night since 2001, it is not a small thing at all.

The most useful automation is the kind nobody photographs. It is not a machine on a scaffold. It is a set of records that reconciles itself while the person who earned the money is asleep.

SOURCES

Statute and statutory instruments

•  Finance (No. 2) Act 2017, section 62 — Digital reporting and record-keeping for VAT — legislation.gov.uk/ukpga/2017/32/section/62

•  Value Added Tax Act 1994, Schedule 11

•  The Value Added Tax (Amendment) Regulations 2018, SI 2018/261 — legislation.gov.uk/uksi/2018/261/made

•  The Value Added Tax (Amendment) Regulations 2021, SI 2021/986 — legislation.gov.uk/uksi/2021/986/made

•  The Income Tax (Digital Obligations) Regulations 2026, SI 2026/336 — legislation.gov.uk/uksi/2026/336/contents/made

•  Small Business, Enterprise and Employment Act 2015, sections 28 and 33 — legislation.gov.uk/ukpga/2015/26/section/33

•  Retained EU Law (Revocation and Reform) Act 2023, section 18 — Abolition of business impact target

•  Consumer Rights Act 2015, sections 50 and 57 — legislation.gov.uk/ukpga/2015/15/section/50

•  Water Industry Act 1991, section 118(5)

•  Legal Aid, Sentencing and Punishment of Offenders Act 2012, section 85

Official statistics and government publications

•  Office for National Statistics, Artificial intelligence in UK businesses: 2023 to 2026, released 20 July 2026 — ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026

•  Office for National Statistics, Which occupations are at highest risk of being automated?, 25 March 2019 — ons.gov.uk/employmentandlabourmarket/peopleinwork/employmentandemployeetypes/articles/whichoccupationsareathighestriskofbeingautomated/2019-03-25

•  Office for National Statistics, GDP monthly estimate, UK: July 2026, released 11 September 2026 — ons.gov.uk/economy/grossdomesticproductgdp/bulletins/gdpmonthlyestimateuk/july2026

•  Department for Science, Innovation and Technology, AI Adoption Research, published 28 January 2026, updated 13 February 2026 — gov.uk/government/publications/ai-adoption-research/ai-adoption-research

•  Department for Science, Innovation and Technology and AI Security Institute, Assessment of AI capabilities and the impact on the UK labour market, 28 January 2026 — gov.uk/government/publications/assessment-of-ai-capabilities-and-the-impact-on-the-uk-labour-market

•  Department for Education, The impact of AI on UK jobs and training, November 2023

•  Department for Science, Innovation and Technology, AI Upskilling fund: application guide, 12 March 2024, updated 22 August 2024 — gov.uk/government/publications/flexible-ai-upskilling-fund/ai-upskilling-fund-application-guide

•  Department for Science, Innovation and Technology and Ipsos, Evaluation of the Flexible AI Upskilling Fund, 28 January 2026 — gov.uk/government/publications/evaluation-of-the-flexible-ai-upskilling-fund/evaluation-of-the-flexible-ai-upskilling-fund

•  HM Revenue and Customs, VAT Notice 700/22: Making Tax Digital for VAT, paragraph 3.2

•  HM Revenue and Customs, Use Making Tax Digital for Income Tax, updated 7 September 2026 — gov.uk/guidance/use-making-tax-digital-for-income-tax

•  Department for Business and Trade, Better Regulation Framework and Small and Micro Business Assessment (SaMBA) guidance — gov.uk/government/publications/better-regulation-framework

•  GOV.UK, Water use in data centre and AI report

•  House of Commons Library, Data centres: planning policy, sustainability, and resilience, CBP-10315, 27 May 2026, by Adam Clark — commonslibrary.parliament.uk/research-briefings/cbp-10315/

Academic

•  Brown, R., Sun, H., Cowling, M. and Dowell, D., Can We See it in the Numbers Yet? Artificial Intelligence, Firm-Level Productivity and SMEs, SSRN abstract 5210948, posted 9 April 2025. Working paper; not peer-reviewed as at September 2026. ESRC-funded, with Oxford Brookes University, on the Department for Business and Trade's Longitudinal Small Business Survey

•  International Monetary Fund, Gen-AI: Artificial Intelligence and the Future of Work, January 2024

•  IPPR, Transformed by AI: How generative artificial intelligence could affect work in the UK, Jung and Srinivasa Desikan, 27 March 2024 — ippr.org/articles/transformed-by-ai

Alan Turing Institute and Innovate UK

•  Innovate UK BridgeAI — turing.ac.uk/partnering-turing/current-partnerships-and-collaborations/innovateukbridgeai

•  AI Skills for Business Competency Framework, version 3 — iuk-business-connect.org.uk/opportunities/ai-skills-for-business-competency-framework-and-resources/

•  Capsules of AI Knowledge — iuk-business-connect.org.uk/news/capsules-of-ai-knowledge/

Platforms

•  Tripadvisor, 2025 Transparency Report, 18 March 2025

•  Google, New ways we're protecting businesses on Maps, 16 April 2026

•  Google, How we're using AI to fight fake reviews, 7 April 2025

Case law, named to explain why it is not relied on

•  Pimlico Plumbers Ltd and another v Smith [2018] UKSC 29, 13 June 2018 — supremecourt.uk/cases/uksc-2017-0053

Case law note. Pimlico Plumbers Ltd v Smith appears in these sources because section thirteen explains why we are not using it, not because it supports anything in this article.

Sources and licensing. Contains public sector information licensed under the Open Government Licence v3.0. Parliamentary Library material, where quoted, is licensed under the Open Parliament Licence. Quotations from copyright works are used under the quotation exception, section 30, Copyright, Designs and Patents Act 1988, and each is attributed in the text. Figures are reported as published by the named body. Where credible sources disagree, all are given. Where we could not verify a figure, it has been removed and the removal recorded.

Nothing in this article is legal advice. Anybody uncertain about their obligations under Making Tax Digital, the Consumer Rights Act 2015 or the Water Industry Act 1991 should speak to HMRC, Trading Standards or their water undertaker directly.

Matthew Kenneth McDaid, trading as Shining Windows 15 September 2026