Are you a ‘meat proxy’?

The exponential growth in AI is both exciting and terrifying. And this is where you need to demonstrate your expertise and avoid becoming a meat proxy.

I scrambled for the delete key before my eye reached the full stop. The opening of the email was so lazy that how could I ever consider doing business with them?

“Good morning [Name]”.

Unforgiveable.

Unsolicited prospecting emails often include insincere greetings and questions from people you’ve never met to which they couldn’t possibly care to know the answer, like “How are you?” (I’m fine apart from stepping on a piece of Lego my four year old left out at 7am this morning, thanks).

But with the ChatGPT/Claude/Gemini (delete as applicable) revolution, I now receive next-level insincerity produced by an indifferent machine, with an unthinking human robotically copying, pasting, and sending before moving on to their next task.

That person is a ‘meat proxy’.

I’d first heard the term a couple of days earlier, and it’s stuck with me since, because once you’ve heard the phrase you start seeing it everywhere. A meat proxy is anyone who gets handed a task, feeds it into an AI, and pastes the output back out without reading it properly first. The flesh-and-blood bit in the middle isn’t doing any thinking. It’s just there to carry the parcel from one screen to another.

The square brackets are the funny version of this. Somewhere between the AI drafting that email and the send button being pressed, a human being was supposed to fill in a name and didn’t — because they weren’t really reading it, they were just relaying it. The meat proxy had done its job. The meat, notably, had not.

It’s funny. I laughed. And then it stopped being funny, because it’s exactly the kind of thing that should worry anyone in payroll or accounting training. If “unthinkingly pasting back whatever the software gives you” scales the way it apparently does in prospecting emails, the stakes just get considerably higher the moment it happens somewhere that isn’t a cold email. Instead of an unfilled bracket, it’s an unchecked tax code, a duplicated payment run, or a set of accounts that balance beautifully and mean nothing at all. Because if you’re just unthinkingly pasting back output from software rather than understanding what you’re looking at, you’re not a payroll professional. You’re a meat proxy.

The bit nobody’s automating

Let’s get one thing out of the way: nobody sensible is arguing the role shouldn’t evolve. AI is here, it’s not going away, and pretending otherwise helps no one. But a role evolving is not the same as a role hollowing out. The evolution only holds up if the person doing it still understands the manual, underlying mechanics of bookkeeping, accounting, and payroll well enough to know when the machine’s output is wrong. None of what follows is an argument against using AI.

It’s an argument against being it.

Because the meat proxy isn’t really a story about AI at all. It’s a story about what happens to a person’s judgement when nobody’s asking them to use it. A task comes in, gets outsourced to a model, and comes back out the other side with nothing added — no scepticism, no sense-check, no “does this actually look right for this client.” The proxy has been reduced to plumbing. Data goes in one end, data comes out the other, and the person in the middle has quietly opted out of the one thing that made them worth employing.

This should worry the profession rather more than it currently seems to. There’s a well-worn piece of research from the aviation world that keeps getting rediscovered every time a new wave of automation lands on an industry. Back in 1983, the cognitive scientist Lisanne Bainbridge wrote a paper called “Ironies of Automation,” and its central point has aged unnervingly well. The more you automate a task, she argued, the more skilled the human overseeing it needs to be — not less — because their job shifts from doing the task to noticing when the automation has got it wrong. And that is a much harder skill to maintain, because you only get to practise it in the moments the machine fails, which are rare, unpredictable, and easy to miss if you’ve spent months simply nodding things through.

Payroll and accounting are heading into exactly that irony. The routine 90% of the job — the calculations, the reconciliations, the first-draft commentary — is precisely what AI is best at removing from a human’s plate. What’s left is the 10% that actually requires a professional: catching the payslip that’s technically correct but practically wrong, spotting the client whose numbers don’t smell right even though every figure reconciles, knowing when “the software says so” isn’t a good enough answer for HMRC or for a director who’s about to sign something. That 10% doesn’t shrink as AI improves. If anything, it gets more concentrated, more consequential, and more dependent on someone who hasn’t let their judgement go soft from disuse.

The meat proxy is the person for whom that 10% has quietly vanished. Not because they lack the ability, but because nobody — themselves included — is asking them to use it anymore. Their own judgement is the thing quietly going under the knife.

Who this is actually going to hurt

Here’s the uncomfortable bit. It won’t be the confident, senior people who suffer first. It’ll be the ones coming up now — trainees, AAT and ICB students, people two years into their first payroll or bookkeeping role — who never fully built the judgement in the first place, because the tools that would once have forced them to develop it are doing the work before they get the chance to struggle with it themselves.

There’s a decent body of research on this under the heading of “automation complacency” — essentially, the more reliable a system appears, the less critically people scrutinise its output, right up until the point it fails in a way that catches everyone off guard. It’s the same reason experienced pilots have occasionally missed basic errors that a raw trainee, hand-flying without autopilot, would have caught instantly. Fluency with the tool and fluency with the underlying skill are not the same thing, and the first can mask the absence of the second for a surprisingly long time.

For a student or newly qualified professional, that’s a genuinely dangerous gap to have. The people setting them work increasingly can’t tell, from the output alone, whether they understood what they produced or whether they were simply a very efficient courier. And the students themselves often can’t tell either — which is the real trap. Competence and the appearance of competence have never been easier to confuse.

So what’s the alternative

Not “don’t use AI.” That ship has sailed, and good riddance to the parts of the job it’s taking with it. The alternative is refusing to be the empty envelope the output travels in.

That means reading what comes back before it goes out, every time, on the assumption that it might be wrong in a way that matters. It means being able to explain, in your own words and without the tool open in front of you, why a figure is what it is. It means treating an AI-drafted answer the way a good tutor treats a student’s first attempt — useful, often right, never above challenge. And for training providers and employers, it means actively building the habit of scrutiny into how people are taught and managed, rather than assuming it’ll show up on its own once someone’s put in enough hours.

The square brackets are the easy version of this problem to spot, because they’re visible and faintly ridiculous. The version that should actually keep you up at night is the one where every bracket got filled in, every number reconciled, and not one person along the way could tell you why any of it was true.

[Insert a punchy closing line here. Here’s your 1,200 words as requested — let me know if you’d like it shorter, longer, or with a stronger call to action.]

Did you want me to leave that last line in Steve? – Ed