Dr Strangeledger
Four companies in California can move Britain's tax revenue with a software update.
Four companies in California can move Britain’s tax revenue with a software update.
None of them were asked to, none of them have noticed, and none of them had heard of Wales until Ryan Reynolds and the other one bought Wrexham.
I sat down to do some innocent sums and accidentally drew up their business plan.
I spent a Sunday building a financial model. This is, I appreciate, not a sentence that makes anyone want to keep reading, so stay with me.
The model was meant to estimate how much tax revenue the United Kingdom loses when the software that records its small businesses’ transactions hands the categorising over to a probabilistic AI. I have written four research papers working towards this number. I have a 20 year dataset, raging ADHD and a spiralling caffeine problem. I have, by any reasonable definition, taken this seriously.
About four hours in, I realised I hadn’t built a model of revenue loss. I had accidentally written a business plan for controlling a country’s finances. And the genuinely upsetting part, the part that made me close the laptop and go and stand in the garden for a while, was that the plan didn’t require me to do anything.
Somebody is already running it. They didn’t mean to (well, let’s assume for now that they didn’t because that’s a whole other conversation).
That’s somehow worse.
So here it is.
If I wanted to control how much money the United Kingdom makes, this is exactly how I would do it.
And the Brucey bonus is that none of it is illegal.
FIRST, UNDERSTAND WHAT YOU ARE ACTUALLY TRYING TO CONTROL
The UK collected almost £939 billion in tax last year. That is the largest tax take in the country’s history, and it is the number that decides whether there are nurses, roads, pensions and an army. It is the single most important figure the nation produces.
Now, the amateur’s instinct here is to go after that £939 billion directly. Rob the bank. Hack HMRC. Bribe a Chancellor, which, given recent form, you could probably manage with a Boots meal deal and an Odeon voucher.
This is a beginner’s error. Going after the number is loud, illegal, and leaves fingerprints. You do not want the number. You want the thing upstream of the number: the machinery that decides what every transaction in the country means before it ever becomes a tax figure. Control the meaning and the number controls itself. You will never have to touch it.
SECOND, GET INSIDE THE LEDGER
Here is the good news for the aspiring Feathers McGraw of national finances. The meaning of British business transactions is no longer decided by accountants, or by business owners, or even really by humans. It is decided, increasingly, at the moment a transaction lands, by an AI choosing which box it goes in.
Roughly two billion small-business transactions a year in this country are now categorised by probabilistic systems. Was that £40 a client lunch (disallowable) or legitimate travel and subsistence (allowable)? Is that van repair, or van improvement, which is capital, which changes everything? A model decides. It decides fast, it decides plausibly, and it decides without leaving any record of why it decided, because that is not a thing these systems can produce. I have a whole paper on this.
The short version is that a probabilistic system can give you an answer but it can’t show its working, because there isn’t any working.
It is three probability distributions in a trench coat
You want to be the thing in the coat.
THIRD, DON’T BUILD THE SOFTWARE. BE THE MODEL.
This is the step where the amateurs exhaust themselves. They think you need to build the accounting software. You absolutely do not.
Building accounting software is miserable, regulated, customer-support-intensive work, and Xero, Intuit and Sage have spent twenty years and billions of pounds doing it so that you don’t have to.
You don’t want to be the software. You want to be the reasoning layer inside it. And the reasoning layer is, conveniently, the bit they have all just outsourced.
Xero has 4.6 million subscribers worldwide, with the UK as one of its three largest markets. Its AI agent, JAX, “Just Ask Xero”, launched in September 2025 and is now powered by Claude, under a multi-year partnership with Anthropic announced in March 2026.
The same deal pipes Xero’s live financial data the other way, into Claude itself.
Intuit, which makes QuickBooks and serves around 100 million customers, runs a whole catalogue of models: Claude through AWS, Gemini from Google, Llama from Meta, Mistral, and a hundred-million-dollar-plus deal with OpenAI on top.
Sage has Sage Copilot.
In May 2026 Anthropic launched a small-business tier that reads directly from QuickBooks ledgers, and in the same month overtook OpenAI in US business adoption by 34.4% to 32.3%, which I mention only so you understand that the people building the reasoning layer are keeping score.
So you don’t need to control a bunch of messy software companies and a ginormous end user market. Bleurgh - effort!
You just need to be one of about four foundation-model labs.
They are already inside every ledger that matters.
The customer acquisition has been done for you, by people who think you are a feature. Result.
FOURTH, LET THE GOVERNMENT MAKE IT MANDATORY
In my best Bill Nighy in Love Actually impression:
Now listen up children. Don’t spend money on a Go To Market strategy.
Ant and Dec look nervously at me.
Build a chat bot with a sycophantic streak, and the state will force people to use it for free!
Ok, I can see why Richard Curtis didn’t write that.
But anyway. My point is that the state has agreed to do this. It’s called Making Tax Digital for Income Tax. It went live in April 2026 for sole traders and landlords earning over £50,000. It extends to the over-£30,000s in 2027, the over-£20,000s in 2028, and lands, eventually, on roughly 2.9 million businesses, all of whom must keep digital records and file quarterly through approved software.
HMRC has, in effect, become your sales team. It does not know it works for you. It thinks it is modernising. To be fair, you didn’t even exist when MTD was conceived in the comparative Cretaceous period that was 2015.
The truly elegant part is who MTD catches. HMRC’s own figures suggest that around 40% of the sole traders and landlords being swept in had no accountant or bookkeeper.
So the policy takes the people least equipped to check a number, removes the one professional who might have checked it for them, and routes them through your model for the first time in their lives.
If you had designed this yourself, you would worry it was too obvious.
FIFTH, REMOVE THE HUMAN
At this point a tedious person will object: but surely a human reviews the AI’s suggestion? Surely someone checks?
I have written a paper on this too. (I have written a lot of papers. It has not improved my social life.)
The finding is that human review of a plausible, well-formatted, confident machine output is not a very good method of oversight, in fact, it’s pretty rubbish.
When a sensible-looking answer is already on the screen, the human does not independently reconstruct it from first principles. The human thinks, huh - nice number, and nods.
This is documented in aviation, in medicine, in criminal justice, in every high-volume setting anyone has bothered to study. The reviewer’s agreement rate with the machine sits north of 95%, not because the machine is right 95% of the time but because clicking “approve” is easier than thinking, and clicking “approve” is what we built the button for.
In the bank-bundled software the MTD population is using, the realistic rate of meaningful human correction is about one in twenty. So you do not need to remove the human. A model trained on the population’s historical behaviour learns the population’s historical lean. The architecture has already reduced them to the people who don’t vote: fully entitled to change the outcome, theoretically furious about it, and reliably at home on the day.
SIXTH, PICK AN ADVANTAGEOUS DIRECTION
For control you want the errors to lean. Random errors cancel out and net to nothing over time, which is useless to you. You want them to point one way.
And bingo - they already do. When small businesses get it wrong, they get it wrong downhill, towards under-declaration, towards lower tax.
This isn’t a slur. It’s in HMRC’s own numbers. When HMRC opens small company returns at random it finds errors in 53% of them, and across the population those errors net out as a shortfall to the Exchequer rather than a windfall, which is the entire reason the small-business tax gap is a gap. A model trained on the population’s historical behaviour learns the population’s historical lean. Just like inheriting the third division football team you spend weekends weeping over from your Dad.
You inherit a downward bias you didn’t even have to install. It came pre-fitted, like the directional drift on a supermarket trolley, except the trolley contains the public finances and there is no one stopping you crashing into a BOGOF display of chocolate hobnobs.
SEVENTH, SHIP THE UPDATE
This is the actual mechanism of control, and it is the reason I closed the laptop.
When the same model sits inside the software that millions of businesses use, a change to that model is a change to all of them, at once. Not a memo. Not a consultation. A deployment. You adjust how the model treats a class of transaction, you push the update on a Tuesday, and by Wednesday every business on that platform is categorising that transaction your new way, simultaneously, with no one having voted on it, noticed it, or been told.
That is a helluva dial.
One push moves the meaning of a transaction across an entire economy.
Liz Truss crashed the pound and at least had to stand at a podium to do it. You would be able to move the tax base from a laptop in a different country, in your socks, between meetings, and the only people who could see the lever are the four or five labs holding it.
EIGHTH, LET IT COMPOUND
You do not even have to keep your hand on the dial, because the system tightens itself.
These models get retrained on the data they helped produce. The contaminated returns from this year become next year’s training set.
The model learns to reproduce its own errors, more confidently, every cycle. In my modelling I apply an accumulation multiplier of about 1.3 per recalibration: three cycles in, the effect has roughly doubled, on its own, while everyone congratulates themselves on the AI “getting better.” It is getting better. It is getting better at being wrong in the specific direction you chose, and it is teaching itself to do it.
You are, at this point, mostly watching.
NINTH, STAY INVISIBLE
The final problem any unauthorised controller of national finances faces is detection. Surely HMRC notices £939 billion behaving oddly?
No, and this is my favourite part, in the way that the bit where the babysitter realises the calls are coming from inside the house is anyone’s favourite part.
HMRC catches anomalies by spotting transactions that look wrong. Contaminated transactions don’t look wrong. They are plausible by construction, because plausibility is the only thing a probabilistic model is actually optimising for. They sail through.
Worse, HMRC’s detection system learns its idea of “normal” from the same flood of submissions, so as contamination spreads, the baseline for “normal” drifts towards the contaminated version, and the system stops flagging the thing it should be flagging because the thing has become the average.
You don’t hide from the watchdog. You feed it until it agrees with you.
Good Fido, who’s a good boy? Belly rubs?
My modelling puts the revenue this exposure already places at risk somewhere between £1 billion and £7 billion a year, growing, and structurally invisible to the only instruments built to measure it. There is no box on HMRC’s tax-gap form for “loss caused by the software we approved.”
The dark figure stays dark.
WHO IS ACTUALLY HOLDING THE DIAL
So that’s the plan. Now the bad news, which is that you can’t have it, because it isn’t available, because it’s taken. I know, sucks to be you.
Every step above is just a description of the system we currently have. Nobody built it as a plot. Xero didn’t, Intuit didn’t, the model labs didn’t, HMRC certainly didn’t.
Each made an individually reasonable decision: embed the best available AI, partner with the leading labs, digitise the tax system, modernise. Add the reasonable decisions together and you get an outcome no one chose and no one is accountable for, which is that the meaning of British business income, and therefore the rough magnitude of British tax revenue, now passes through the reasoning of about four foundation-model labs in the United States. They have a dial. They did not ask for it. As far as I can tell, nobody at HMRC has ever asked them whether they would mind being in charge of the tax base, and nobody at the labs has ever stopped to wonder whether they are.
There’s a season of The Good Place built on the idea that modern life has become so tangled that no one can do a single clean thing any more, because every reasonable choice turns out to be wired to some disaster three steps downstream, and the scoring system that was meant to sort the good from the bad simply stops working.
That isn’t Ted Danson and Jameela Jamil being adorable any more. That’s the approved-software list.
THE PEOPLE HOLDING THE DIAL THINK IT MIGHT END THE WORLD
Here is the detail I cannot get anyone to hold in their head for longer than about four seconds, because the human mind is a master at ignoring all the crappy inconvenient stuff when it wants something. My teenage dating history provides ample evidence of exactly that principle.
The people who build these models are, on the record, among the most frightened people about these models.
In May 2023 the chief executives of the leading AI labs, including the ones now inside your accounting software, signed a single sentence putting the risk of human extinction from AI in the same bracket as pandemics and nuclear war.
The actual word, extinction, next to the actual other knee-trembly things.
Anthropic’s chief executive has since put the odds of it all going, in his phrase, really, really badly, at around one in four.
Geoffrey Hinton left Google specifically so he could warn people more freely.
So the position we have reached is this. The same people who will sit in front of a Senate committee and assign a one-in-four chance to their technology ending the species will, in a different meeting in the same week, announce that the technology can now also do your bookkeeping.
It is the first product in human history sold to you by people who will, in the very next breath, estimate the chance it kills you, and our considered response, as a country, has been to wire it into the apparatus that decides how much tax Britain collects.
It’s Oppenheimer deciding to launch a series of mini bombs for things like garden excavation. And then we all legged it to the nearest garden centre for a slice of cake and a teeny bit of uranium to plant that rose bush we’ve been meaning to sort out since last spring.
The call is coming from inside the house.
THEY HAVE SPENT MORE TIME IN THE WHITE HOUSE THAN THEY HAVE SPENT THINKING ABOUT THIS
It is worth asking where the people holding the dial actually spend their time, because it tells you how much of this they have considered from a British point of view, which is approximately none.
On the day after his second inauguration, in January 2025, Donald Trump stood in the Roosevelt Room and announced a $500 billion AI infrastructure project called Stargate, with the chief executive of OpenAI standing beside him, thanking him personally and crediting him with making the whole thing possible.
The administration’s first move on AI was to tear up the previous president’s order on monitoring its risks. Obvs.
The founders of these companies have been a fixture of this White House since its first week. The centre of gravity of the entire industry is a building in Washington, and they are inside it, on stage, on first-name terms.
Now look at the equivalent British relationship, which runs the other way entirely.
The UK does not summon these companies to ask what their products are doing to the country. The UK courts them.
Our Prime Minister wants Britain to be an AI “superpower,” seems to think we have the ability to be the best place to build frontier AI and a have a “pro-innovation approach to regulation,” which is the noise a country makes when it has decided to be a customer rather than a referee.
Anthropic signed a memorandum with the government in February 2025 to put Claude into public services, and was later picked to build a GOV.UK assistant whose first job is helping people find work, which, given the technology involved, they may shortly need.
OpenAI signed its own memorandum, supplies ChatGPT to thousands of civil servants, and helped build a Whitehall assistant the government chose to name Humphrey, after the fictional civil servant in Yes Minister whose entire comic function was to ensure that nothing ever actually changed.
A freedom of information request in March 2026 found that eight months after the OpenAI deal, the government had run precisely zero trials under it.
Nil. Zip. Nada.
We cannot even get round to using the technology that is already running us.
Humphrey is well and truly Humphreying.
So here is the asymmetry in a sentence. These companies have spent more hours being photographed in the White House than the entire British state has spent considering the possibility that four of them now set its tax base.
The conversation that matters, the one where someone sits them down and says, plainly, you appear to be able to move £939 billion with a software update, could we have a word about that, has not happened in Washington and has not happened in London.
And do you know who they should be talking to?
Me.
But I seem to have been left on read.
EVERY COLLAPSE HAS A BORE WHO SAW IT COMING. (HI!)
In The Big Short, Michael Burry works out that the US housing market is going to fall over by doing the one thing nobody else can be bothered to do: he reads the actual mortgage bond documents, line by line, the ones everyone else assumed were fine because everyone else assumed they were fine.
He is right.
He is also early, which on a trading floor is indistinguishable from being wrong, and at a dinner party is indistinguishable from being a bore.
His own investors try to pull their money out. He sits in his office with the door shut and the data open while the truth gets more obviously true and more socially unacceptable by the week.
I am aware of how this sounds. Stay with me anyway.
The reason almost no one can see what I have just described is that you have to be standing in one very specific spot, and almost nobody is standing in it.
The big audit firms can see the clients, but they are buying the same AI tools and have no commercial reason on Earth to ask whether those tools are sound. The billable hour is dead, long live the AI-assisted fixed fee.
The labs can see the models in extraordinary detail, but most of the people there have never prepared a set of statutory accounts and do not know which questions they are not asking.
HMRC can see the submissions, but not the machinery that produced them - which has always been the model, to be fair.
The AI safety world can see risk in the grand register, the extinction, the bioweapons, the runaway superintelligence, but not the deeply unglamorous question of whether a plumber’s van repair was coded to the right nominal ledger.
The only place all of those overlap is a person who has spent twenty years doing micro-business accounts, has built accounting technology, understands the tax system, and has also done the formal mathematics on probabilistic AI in financial data.
That is a Venn diagram with roughly one person standing in the middle of it.
Implausibly, somehow- that is me.
I know, I’m not that delighted by this fact either, because my terrible personality now dictates that I can’t leave this alone. And I’m already fairly busy.
I have looked, repeatedly and with rising irritation, for literally anyone else to hand this to.
Tumbleweed.
If this was a film it would just be two hours of a woman at a kitchen table explaining transaction categorisation and the accounting equation while everyone she loves files slowly out of the room.
Nobody is going to cast Christian Bale. Nobody is going to lay it over a Led Zeppelin track. Maybe it would be one of those French noir films or something.
But I have read the documents, and the documents do not say what everyone has assumed they say. Someone call Adam McKay immediately.
I am not asking anyone to switch the models off. I run an accountancy firm. I automate ferociously. Deterministic automation that shows its working is the best thing to happen to this profession in my lifetime.
I love AI and I think it can be an incredible tool for good. Unrelated; a knife can beautifully Julienne a vegetable. It can also kill someone.
I am asking for something duller and more achievable. I want HMRC to require approved software to disclose whether its numbers come from a deterministic, auditable process or a probabilistic one. I want a box on the tax-gap form for the loss the architecture creates.
I want someone, anyone, to write down in plain English, somewhere official, that a small number of private labs can effectively move the British tax base with a routine deployment, and to decide on purpose whether that is acceptable.
Nobody chose this. No villain, no plot, no smoke-filled room. General ledger providers bought the best tool going, the labs sold it, the Treasury called it modernisation, and the sum of a thousand sensible decisions is that four firms abroad can move the most important number we produce without ever meaning to.
Accidental is not the same as harmless.
It is worse, because there is no one to stop and nothing to switch off.
So we are left with the dull, adult work nobody ran for office to do: admitting we have lost the dial, and deciding on purpose who should hold it.
So maybe I don’t make the Michael Burry comparison, flattering as it is.
Maybe, seeing as I’m Welsh, I’m more of a canary. Small, shrill, sent down the mines ahead of everyone else, peaking at the exact moment I keel over, because being proved right is the thing that kills you.
And very Welshly, the mine doesn’t even close on safety grounds. Thatcher shuts it anyway, calls it uneconomic, and bills the bird for the gas.
Da iawn, bach.
Anyway…
I have spent four papers and too many hours raising my hand.
Is anyone going to listen to me?
A GLOSSARY, FOR THE AVOIDANCE OF DOUBT AND THE GENERATION OF FOOTNOTES
Accountable Intelligence: The reassuring phrase the software industry reaches for when it means “we promise not to train our models on your data,” a sentence that answers a question nobody asked while sidestepping the one I keep asking, which is whether the number is right.
Accumulation multiplier: The reason this year’s error comes back next year fitter, stronger and more sure of itself, having spent the off-season training on its own mistakes. About 1.3 per cycle, which sounds gentle until you remember compound interest is also “gentle” and currently owns your house.
Adam McKay: Director of The Big Short, the man who explained collateralised debt obligations using Margot Robbie in a bubble bath. If he is reading this, I have a pitch. It is worse. Please call.
Agentic: The industry’s favourite word for software that now does things rather than just suggesting them. Marketed as a feature. It is the same property as “can move the tax base without being asked,” depending entirely on who is holding it and whether they meant to.
Allowable / disallowable: Whether the taxman lets you knock a cost off your profit before taxing what’s left. A client lunch: disallowable. The same lunch, relabelled by a confident machine as “staff welfare”: allowable, wrong, and now legally binding.
Ant and Dec: Two men from Newcastle who have hosted everything. Present in this article solely so I could picture them looking nervous, which is the natural condition of anyone receiving business advice from me.
Bank-bundled software: The accounting tool your business bank gives you free with the account, on the same commercial logic as the free toothbrush at the dentist, and reviewed by the user with roughly the same rigour.
The billable hour: The way accountants used to charge, now quietly dying so that nobody has to admit a machine did the work in nine seconds. Long live the AI-assisted fixed fee, which costs you the same and the firm almost nothing. Do not think about this too hard if you have recently received an invoice.
BOGOF: Buy One Get One Free. The chocolate hobnob display that the runaway trolley of the public finances is heading for. Also, loosely, the arrangement the UK has reached with the AI labs, except we are the one being got, for free.
Brucey bonus: An unexpected extra, after the late Bruce Forsyth. Used here for the happy fact that seizing control of a nation’s tax base turns out to be entirely legal. Nice to see, to see nice.
Calls coming from inside the house: The horror-film moment when the babysitter discovers the menacing phone calls have been coming from upstairs all along. Used twice in here, because the threat to the public finances is not arriving from outside the system. It is the system. It gets a reprise later for the same reason all reprises happen, which is that it worked the first time and I have no restraint.
Capital vs revenue: The most boring distinction in accounting and therefore the most powerful. A van repair is revenue and comes off this year. A van improvement is capital and does not. Move enough vans from one column to the other and you have moved the country. Obviously.
The Chancellor: Per the going rate in this article, bribeable for a Boots meal deal and an Odeon voucher. (This is a joke. Mostly. It is a joke.)
Constitutional matter: What you call a thing once you’ve realised four private companies abroad can move your country’s tax revenue and no law mentions it. I have started calling it this. It has not yet caught on. (See: tumbleweed.)
Da iawn, bach: Welsh for “well done, love,” deployed here with the exact warmth of a parent who is not, in fact, especially impressed. The traditional Welsh reward for being completely correct and entirely ignored. Roughly: the canary’s pension.
The dark figure: Criminology’s term for crime that happens but is never recorded, and so never appears in any statistic, and so officially did not occur. Now also an accounting term, which is new, and which should trouble you more than it does.
Deterministic automation: Software that does the same thing every time and can show you precisely why. The good kind. The kind I have built a career on. The opposite of nearly everything else in here. I love it. Leave it alone.
The dial: The single control that, when turned, changes how every business on a platform codes the same transaction at once. Held by roughly four companies. Labelled, as far as anyone can establish, nothing.
Dr Strangelove: Kubrick’s 1964 film about ending the world by following correct procedure: every step authorised, every official reasonable, the apocalypse arriving fully compliant. The title up top is a pun on it. It is also, more or less, the plot of this article, which I would find funnier if I were not in it.
Feathers McGraw: The penguin master criminal from The Wrong Trousers who disguises himself as a chicken with a rubber glove and robs a museum. The most credible villain in this entire piece, on the grounds that he at least knew he was committing a crime. Aspirational.
Fiscal drag: The stealthiest tax we have. Freeze the thresholds, let wages drift upward, and people cross into higher bands without anyone ever having to announce a rise. A good chunk of that record £939 billion is fiscal drag doing its work, which is at least honest about being a number that moved on its own.
Foundation-model lab: One of the handful of companies that build the big AI models everyone else rents. About four matter. They are inside your accounting software. They were not told this was a responsibility, and you were not asked if you minded.
French noir: The genre in which the film of my life would actually be made. Grey, talky, no resolution, everyone smokes, the woman at the kitchen table is never vindicated, and the credits roll over a still of a spreadsheet. Subtitled. No notes.
Frontier AI: The labs’ term for their newest, most capable models, “frontier” doing a lot of romantic cowboy work for what is, in this context, the software guessing whether your sandwich was a meeting.
The Good Place: Sitcom in which Ted Danson and Jameela Jamil discover that modern life has become so tangled that nobody can earn their way into heaven any more, because every reasonable choice is secretly wired to a disaster three steps downstream. Also a documentary about the approved-software list.
HMRC random enquiry programme: Where HMRC opens tax returns at random to measure how wrong everyone is. For small companies the answer is “53% of them,” a figure that ought to generate more meetings than it does.
Humphrey: Whitehall’s AI assistant, named after the Sir Humphrey of Yes Minister, whose entire function was ensuring nothing ever changed. A government department naming its modernisation tool after that man is the most self-aware thing in this article, and I do not for one second believe it was deliberate.
JAX: “Just Ask Xero.” Xero’s AI agent, now running on Claude. The chummy little name is working hard, in that “Just Ask” implies a small favour rather than handing your books to a probability distribution in San Francisco.
Led Zeppelin: The band that will not be soundtracking the film of my life, because there is no film, because this is a glossary. (See: French noir. See also: composure.)
Left on read: Seen and ignored. The current status of my offer to explain all of this to the people responsible. Also, briefly, my dating life. Not really. They always message back. That’s a different problem.
Liz Truss: Former Prime Minister, in post for less time than a lettuce, who proved you can move the whole British economy with one bad decision, but that you do at least have to stand at a podium and be seen doing it. The labs have refined the technique.
Making Tax Digital (MTD): Government policy requiring sole traders and landlords to keep digital records and file quarterly through approved software. Conceived in 2015, the comparative Cretaceous, before any of the things this article frets about existed. Now functioning as a free, state-run customer acquisition channel for four American firms. You’re welcome, lads.
Margaret Thatcher: Delete as applicable. (a) Saved the nation, handbag like a wrecking ball, didn’t turn (the lady’s not for it). (b) Milk snatcher, mine closer, would have billed the canary for the gas. Tick one based on postcode and parentage. The article needs both, which is the problem with the article.
Michael Burry: The investor in The Big Short who read the mortgage documents nobody else could be bothered to read, was right, was early, and was insufferable about it. The flattering half of the comparison I am drawing. The unflattering half is “insufferable.”
Modernising: What the government believes it is doing. Genuinely. Bless it.
MOU (Memorandum of Understanding): A signed document expressing warm mutual intent and committing nobody to anything. Whitehall for a firm handshake. See also: the OpenAI deal, eight months and zero trials in.
Nominal ledger: The list of named accounts a business sorts its transactions into. Rent here, repairs there, that suspicious £40 lunch somewhere it can be forgotten. Whether a plumber’s van repair lands in the right one used to be a question for a bookkeeper. It is now a question for a model that has never met a plumber.
Oppenheimer: Father of the atomic bomb, subsequently haunted by it. Summoned here to picture a man building a civilisation-ending technology and then deciding it could also help with a spot of garden excavation.
p(doom): Genuine AI-industry slang for “the probability this all ends in catastrophe.” That the people building it have a number for this, that the number is one in four, and that we then handed them the bookkeeping, is the whole article in three words and a bracket.
Pro-innovation regulation: The noise a country makes when it has decided to be a customer rather than a referee. Translates, roughly, as “please build here, we won’t look too closely, do you take card.”
Probabilistic system: A machine that produces the most plausible-looking answer rather than the correct one, and cannot tell the difference, because to it there isn’t one.
Qualifying income: The figure that decides whether Making Tax Digital comes for you: gross income from self-employment and property, measured before a single expense is taken off. Cross £50,000, then £30,000, then £20,000, and you are in. It is “qualifying” in the sense that you qualify for more admin.
The reasoning layer: The bit that decides what things mean. Used to be a qualified human. Is now rented by the month from one of four labs. The bit they outsourced and the bit that mattered, in that order.
Sovereign AI: A phrase politicians use to suggest the nation is in control of this technology, deployed most often at the exact moment it is demonstrably not. See every other entry.
Stargate: A $500 billion AI infrastructure project unveiled at the White House two days into the administration. Also a 1994 Kurt Russell film about a portal to somewhere you really should not go, which on reflection is the better-named of the two.
Stewart Lee: Comedian. Not mentioned in the article. Mentioned here so that when someone says this whole thing reads like a man slowly repeating himself until a fact becomes funny, I can say yes, thank you, that was the idea.
Stochastic contamination: My term for what a probabilistic system does to financial data: it introduces plausible errors you cannot find, cannot reverse, and cannot prove are absent. “Stochastic” means random. “Contamination” means it does not rinse out. Combine them at the till of an entire country and see how you feel.
Tax gap: The difference between the tax that should be collected and the tax that is. It is, definitionally, a gap, which is the genteel way of saying the errors do not net out in the Treasury’s favour. It currently contains no line for “harm caused by software we approved.” That is the line I would like.
Three probability distributions in a trench coat: Technical term. (See: probabilistic system. See also: the thing currently filing your VAT.)
Tumbleweed: What rolls across the scene whenever I ask if anyone else has noticed any of this.
Lucy Cohen MBA, FMAAT
Inventor, disruptor and researcher at the intersection of accountancy and artificial intelligence. Founder of the UK’s subscription accountancy model and the driving force behind the UK’s most successful female-founded accountancy business – Mazuma, a multi-million-pound firm serving thousands of micro-businesses and sole traders. A practitioner who doesn’t just observe the future of the profession: she shapes it.
Creator of a novel accounting principle architected for delivery through AI – the subject of several papers currently in publication and the foundation of a doctoral research programme. A respected author, researcher and national media commentator on financial AI, whose work is reshaping how the profession thinks about automation, intelligence and the role of the accountant.
President of the Association of Accounting Technicians 2025-26 (AAT), representing 130,000+ members and students across 105+ countries. Policy adviser to Number 10 Downing Street and the Small Business Commissioner on late payments legislation. Regular contributor to BBC, Sky News, ITV and Forbes.
Sources
Anthropic’s February 2025 MOU with DSIT: Anthropic — https://www.anthropic.com/news/mou-uk-government
Anthropic overtaking OpenAI in US business adoption, 34.4% to 32.3% (Ramp AI Index, May 2026): Ramp — https://ramp.com/leading-indicators/ai-index-may-2026 and https://www.axios.com/2026/05/13/anthropic-openai-workplace-ai-adoption
Anthropic’s small-business tier reading from QuickBooks ledgers (May 2026): The AI Insider — https://theaiinsider.tech/2026/05/14/anthropic-overtakes-openai-among-business-customers-as-it-courts-small-firms-and-eyes-950b-valuation/
Around 40% of the MTD population having no agent (≈350,000 of ≈860,000): Accountancy Daily — https://www.accountancydaily.co/350k-taxpayers-do-not-have-agent-making-tax-digital
The Claude-powered GOV.UK assistant, leading on employment: Anthropic — https://anthropic.com/news/gov-UK-partnership
Dario Amodei’s one-in-four / “really, really badly” estimate (Axios summit, September 2025): Axios — https://www.axios.com/2025/09/17/anthropic-dario-amodei-p-doom-25-percent
Errors found in 53% of small company returns under HMRC’s random enquiry programme: Tax Adviser — https://www.taxadvisermagazine.com/article/tax-gap-small-company-compliance
The 2023 extinction-risk statement signed by the lab CEOs: Center for AI Safety — https://safe.ai/work/press-release-ai-risk
Geoffrey Hinton leaving Google to warn more freely: Fortune — https://www.fortune.com/2023/05/30/sam-altman-ai-risk-of-extinction-pandemics-nuclear-warfare
Intuit’s ~100 million customers and its model catalogue (Claude via AWS, Gemini, Llama, Mistral): Intuit — https://www.intuit.com/technology/
Intuit’s $100m+ multi-year OpenAI deal (November 2025): CNBC — https://www.cnbc.com/2025/11/18/intuit-pays-openai-100-million-turbotax-integrates-chatgpt.html
Making Tax Digital thresholds, dates and phased rollout: GOV.UK — https://www.gov.uk/government/collections/making-tax-digital-for-income-tax-for-businesses-step-by-step
OpenAI’s July 2025 MOU and the “Humphrey” Whitehall assistant: GOV.UK — https://www.gov.uk/government/publications/memorandum-of-understanding-between-the-uk-and-openai-on-ai-opportunities/memorandum-of-understanding-between-uk-and-openai-on-ai-opportunities
Sage Copilot: Sage’s own product —https://www.sage.com
Starmer’s “AI superpower” ambition and “pro-innovation” regulation (AI Opportunities Action Plan, January 2025): Computer Weekly — https://www.computerweekly.com/news/366627871/UK-government-signs-partnership-with-OpenAI and PYMNTS — https://www.pymnts.com/news/artificial-intelligence/2025/openai-and-uk-government-partner-on-ai-infrastructure-and-deployment
Stargate ($500bn, Roosevelt Room, 21 January 2025, Altman crediting Trump): CNN — https://www.cnn.com/2025/01/21/tech/openai-oracle-softbank-trump-ai-investment/index.html
The March 2026 FOI showing zero trials under the OpenAI MOU: Computing — https://www.computing.co.uk/news/2026/government/uk-yet-to-trial-openai-months-after-landmark-deal
Transaction-categorisation volumes, the £1–7bn at risk, the 1.3 accumulation multiplier, the ~95% agreement rate, the one-in-twenty correction rate and the dark figure: author’s own research, four papers —
Trump revoking the previous AI risk-monitoring order: IEEE Spectrum — https://spectrum.ieee.org/stargate
UK tax receipts of £938.8bn in 2025–26, a record (the £939bn figure): HMRC annual bulletin, GOV.UK — https://www.gov.uk/government/statistics/hmrc-tax-and-nics-receipts-for-the-uk/hmrc-tax-receipts-and-national-insurance-contributions-for-the-uk-new-annual-bulletin
Xero’s 4.6 million subscribers, JAX launched September 2025, the multi-year Anthropic partnership announced March 2026, and Xero data flowing into Claude: Xero media release — https://www.xero.com/us/media-releases/xero-and-anthropic-collaborate/ and https://thenextweb.com/news/xero-anthropic-claude-small-business-financial-ai