The Government did not actually measure automation. I know, because I read the methodology after my profession came first. Unfortunately the correction did not make me feel much better.
The Department for Education ranked 365 occupations by their exposure to artificial intelligence.
I came first.
Not “consultants featured prominently”. First. Rank one of three hundred and sixty-five. The occupation is “management consultants and business analysts”, which is me, and also most of the people selling AI readiness workshops.
A note for anyone American reading this, before you file it under foreign government trivia. The scoring model is American, built on US occupations and then mapped onto British job codes. This is your list too, and nobody asked you either. Accountants, solicitors and analysts are all close behind.
I should correct my own headline before somebody from the Department does. The report, from November 2023, did not measure which jobs were most automatable. It measured exposure, and it says so twice. Which matters rather a lot when you have just come first.
My immediate instinct was to read the methodology, build a framework and write a couple of thousand words about it. The Department may have a point.
On this page
- Marking my own homework
- The ranking is wobblier than the rank suggests
- The comfortable answer does not survive
- The awkward fact I have to admit
- What is left, and it is less than I would like
- How I actually work, since you are entitled to ask
- This was never about consultants
Marking my own homework
Before I take the ranking apart, which I intend to enjoy, here is what I get paid to do, scored honestly.
Desk research and market scans. Gone. Genuinely gone.
Summarising two hundred pages into six. Gone, and better than I did it.
Competitor benchmarking. Going.
Producing forty options before lunch. It does not need lunch.
Turning that into a chart with an arrow pointing north-east. I may have a problem.
Writing “unlocking strategic value” underneath the arrow. I am finished.
Six things I would have called skilled work five years ago. A fair number of my clients have tried the first three themselves at two in the morning, and they know how good it has got.
AI has not made expertise worthless. It has made the proof of expertise worthless.
There is probably an unnecessarily grand name for this. I am going with the proof-of-work problem, mostly because inventing a framework is the only part of my old job description I am not yet willing to surrender.
For years the output carried information about the process behind it. Two hundred credible pages implied that somebody capable had spent two months producing them, and that implication is what people were buying. AI has severed the link. The report still exists. The weeks do not. There is nothing on the cover to tell you which one you are holding.
The ranking is wobblier than the rank suggests
The list runs to 365. Rank 364 is roofers. Rank 365, the least exposed occupation in Britain, is “sports players”. The safest job in the country is being quite good at sport, and the least safe is advising the sports club on its operating model.
Then you read the appendix. The top five occupations are separated by 0.043 of a standard deviation, so rank one and rank five are a rounding error apart. The Department is more confident in its conclusion than in its leaderboard, which is not how the leaderboard got reported.
Underneath the American model sit Mechanical Turk workers, rating how closely ten pre-ChatGPT AI applications relate to a list of human abilities. The ten include abstract strategy games and recognising instrumental music tracks. This is not quite the robot uprising I had pictured.
Run the same model on the original American data and my occupation comes nineteenth out of 774, because Britain has fewer and broader job categories, and the big one marked “people who make slides about other people’s businesses” won. My greatest professional error appears to have been practising in a country with insufficient occupational categories.
The same report has a second column for large language models, the technology that actually turned up. On that one I come tenth, three places ahead of the clergy. Tenth is the sort of result I would normally put in an appendix.
And now the part where I stop enjoying it. Nineteenth of 774 is the top three per cent. So is tenth of 365. So is first. I have spent this entire section moving myself from the ninety-ninth percentile to the ninety-seventh, and none of it changes the finding. That is the most consultant thing in this article, and I am leaving it in so you can see me do it.
The comfortable answer does not survive
The reassuring line is that human judgement is irreplaceable. I went looking for the evidence and it is not there.
In 2023, academics from Harvard, Wharton, MIT and Warwick, with BCG’s own research arm, ran an experiment on 758 Boston Consulting Group consultants. Inside the model’s capability, those using GPT-4 were 25% faster and scored around 34% higher.
Then the researchers gave them a problem built to sit just outside it, where the spreadsheet pointed one way and the interview notes pointed the other. The consultants using AI got it right 19 percentage points less often.
The group who had been taught how to prompt did worse than the group left to work it out, though that particular gap is not statistically firm. As a consultant, I naturally conclude they needed a better training programme.
And on that same task, where they were more likely to be wrong, the graded quality of their recommendations went up. They were more persuasive while being incorrect. I would make a joke about that, but we already are one.
Note what did not happen. They were not short of information. The interview notes were in the pack. The confident thing on the screen beat the awkward thing in the transcript. That is not a failure of access but of weighting, and it is the one I worry about, because it is invisible from the inside.
A meta-analysis of 106 experiments, in Nature Human Behaviour in October 2024, found the same shape at scale. Human and machine together beat the human alone, but not the better of the two working alone: where the machine was stronger, the pairing came out worse than the machine by itself. That is hard to square with “human in the loop” as a general defence. The loop is only worth having when you can say which party is better at the specific thing.
Now ask the machine to do my job. In March this year, Harvard Business Review put seven strategic dilemmas to six frontier models. On six of the seven they picked the fashionable side: differentiate, augment, play the long game. Not because the cases pointed that way, but because the internet does. They called it trendslop, and likened the output to “a freshly minted MBA or junior consultant, parroting what’s popular rather than what’s right for a particular situation”.
The machine has not learned to replace consultants. It has learned to impersonate one, reproducing the house style perfectly: fluent, confident, context-free, reaching for the word “unlocking”, and wrong in a font I recognise.
The awkward fact I have to admit
There is an obvious problem with all this: clients keep hiring consultants. The trade body expects the UK market to grow six per cent this year and eight per cent next. Nearly four years after the tool became free, buyers with ChatGPT open on the next monitor keep signing the engagement letter. That is the market voting with money on exactly my question, and voting against me.
The underlying numbers are shakier than the forecast. Employment in the industry hit its record in 2023 and fell the year after, every Big Four consulting arm in the UK shrank last year, and Deloitte’s UK revenue fell for the first time since 2010. None of which proves AI did it, and this is exactly the point at which a consultant finds the series that agrees with him.
So here is the one piece of direct evidence, and it is worth more than all of the above. In February, KPMG asked its own auditor for a discount on the grounds that AI had made the work cheaper. It got fourteen per cent off. That is the actual threat, and it is quieter than redundancy.
AI can take the revenue without taking the jobs.
Four years is enough to know AI has not killed consultancy. It is nowhere near enough to know what it has done to the economics. If you buy this kind of work, I would like to know whether your own invoices have moved, because I can see the market data and I cannot see your invoices.
What is left, and it is less than I would like
If the proof is free, the scarce thing is no longer an answer. It is a reason to trust one. Verification is the bottleneck now, and verification is the one job that gets harder as the output gets better.
So what can an outsider do that a good employee and a good model cannot? Three things, all narrower than my profession likes to claim, and none of them is being cleverer.
The first is independence, and it is about weighting rather than knowing. Your people almost certainly know the awkward thing already: the information was in the pack. What I have is no career reason to let the confident thing outrank it. Being the person who keeps saying the forecast is soft is not an obvious route to promotion.
The second is permission. Everybody already knows the personnel change the whole plan quietly assumes is not going to happen. Almost nobody who works there can put that in a document. The obvious retort is that I am on next year’s renewal panel too, and that is fair. But the worst thing that can happen to me is that I lose a client, and that is not the worst thing that can happen to them.
The third is pattern recognition with consequences attached. The model has read vastly more than I ever will, and that is its advantage and occasionally its problem. Ask it a generic strategy question and it has an extraordinary amount of evidence about what businesses in general say they ought to do. The fashionable answer is the population mean with no company attached to it. Having watched the same failure in person, a few dozen times, in businesses that each believed they were the exception, is a smaller and more biased sample than the internet. It is also the only one carrying the detail of how it actually went wrong, which is what lets you rank the awkward thing in the transcript above the confident thing on the screen.
Accountability sits underneath all three, but not for the reason I first thought. Insurance can absorb a financial claim. It cannot make a repeated record of bad judgement go away. I have been doing this long enough to have checked. My name on this is not the product. It is the evidence that the judgement is real.
How I actually work, since you are entitled to ask
I use AI heavily. Research, first drafts, structuring arguments, pulling documents apart, arguing with my own conclusions. It usually wins. Occasionally it hands me something fluent, confident and wrong in a font I recognise. So do I.
I am not going to tell you a human checks everything and leave it there, because the evidence above says that is a weaker safeguard than it sounds. I use it where I can show it beats me, I own the parts where I beat it, and I go and find what it has no way of knowing. Behind anything that leaves here, a real person has read it and a real person answers you. That is exactly what someone would say either way, which is why I would rather be judged on whether the answers turn out to be any good.
On price, the first three things on that list are no longer line items on mine, because I cannot defend billing for something that now takes an afternoon. Work that took three weeks and takes one gets billed as one. That is a thinner concession than it sounds, since the days I removed were the cheap ones. I would like some credit for the rest of it. It was not entirely voluntary.
What I can defend is finding out which problem is real, getting at what never made it into the brief, and working out which sensible-looking answer will not survive this particular business. And putting my name behind it. That is a smaller claim than my profession used to make.
One last disclosure. AI helped me research this article, argue against it, restructure it, and turn it into the various formats in which you may be encountering it. Which is either responsible use of the technology or an unusually elaborate demonstration of my own point.
This was never about consultants
You are on the same list, and you did not have to read a government spreadsheet to suspect it.
That includes accountants, solicitors, analysts, agencies and researchers: anyone whose expensive output used to carry an implicit message underneath it, which was that somebody who knew what they were doing had spent a long time on this.
Which is the proof-of-work problem, and it is the whole thing really. The polished deliverable used to be evidence. It is now just a deliverable, and on the occasions it is confidently wrong it will look better than the version that was right.
The question is not whether AI can make your deliverable. It probably can. The question is what you were charging for underneath it.
AI is not about to put me out of work. What it has done is make it much harder to pretend that producing the work is the same thing as doing the job.
I have spent rather too long arguing with a government spreadsheet about whether I should have come first, nineteenth or tenth.
I still think the list is badly built. I no longer think that is the interesting part.
Which leaves the practical question of what you do about it. Some of it is admin: every UK director now has to prove who they are to Companies House, which is proof of identity arriving at almost exactly the moment proof of work stopped meaning anything. Some of it is judgement: the habit of judging the next big technology on evidence rather than press releases is the same one this piece is asking for. And if you assume the people starting out now are the ones in trouble, the figures on young founders say something more interesting than that.