On Thursday, Anthropic is due to start trading on the Nasdaq, in a share sale of more than $60 billion that only SpaceX’s has beaten. The sale would value the company at about $1.2 trillion, and a valuation like that assumes its sales will grow enormously. Azeem Azhar argued in the New York Times this week that software budgets are too small to pay for that growth, and that the sums only work if AI is paid for out of wages. He puts American white-collar pay at $8 trillion to $10 trillion a year.

That leaves a question for those of us who run something smaller. Where does the AI bill sit in our own accounts? For most of us it’s under software, a line or two below Xero. I think that’s the wrong place to judge it from, or soon will be. Once AI does a job from start to finish, its cost belongs beside the wage bill. I don’t mean in the ledger, where the accountant will still code it as software. I mean in the budget we hold it against when we decide how much to spend, who looks after it and how we judge it.

A tool and a worker

Xero is a tool. It helps a person do the books, and that person still does them. Nobody asks what Xero produced this month. We ask whether the bookkeeper got the VAT return in on time.

AI started out the same way. A chat window helps someone write the email or tidy the spreadsheet, and they’re still the one doing the work. That has changed. AI can now take a customer enquiry, look up the order, write the reply and send it. It can match a month of supplier invoices against the bank feed and list the three that don’t agree. When it does a whole task like that, it’s doing work we’d otherwise pay a person to do.

Large companies have stopped treating it as an IT purchase. BCG surveyed 1,330 senior executives this year and found that AI spending had doubled in twelve months to 3.3% of revenue. More than 80% of it now sits outside the IT budget. The same companies expect their workforces to be 10% to 15% smaller by 2030. BCG sells AI advice, so its survey isn’t neutral, and it stops short of saying the money is coming out of wages. It shows AI being paid for by the departments that do the work. The comparison with the wage bill is mine.

What changes when we move it

I argued in July that the AI bill had become a meter and needed a variable line of its own. This is the next step, which is deciding what that line gets compared with. Three things change once the answer is wages.

  1. The budget. Software has a natural ceiling. Twenty pounds or so a month for each person feels about right, and anything much above it gets questioned. The wage bill works on a different scale. A full-time job on the National Living Wage costs an employer more than £2,000 a month in pay alone. If an AI system reliably does a fifth of that job, £200 a month can be good value, and a rule that says “nothing over twenty pounds” will stop us buying it. Count the time someone spends checking its work as part of the cost, and be clear what the freed hours will be used for.
  2. The owner. Software belongs to whoever looks after the laptops and the logins. Work belongs to whoever manages it. If AI is answering customer enquiries, the person who runs customer service should hold that budget and be responsible for the replies. They’re the one who knows what a good reply looks like. Whoever looks after security and data protection still gets a say in which tool is used.
  3. The measure. We judge software by whether people use it. We judge a new hire by the work they take on and how much of it has to be redone. AI that does a task should face the second test. Which task has it taken over? Who checks it, and how often are they correcting it?

The job that never gets advertised

Wage bill is an uncomfortable phrase, because it sounds like redundancies. In a firm of twenty or two hundred people, I don’t expect it to mean that. What I’d expect to see is the hire not made. The accounts assistant leaves and the advert never goes out. The second person in customer service, pencilled in for next spring, drops off the plan because the queue is under control.

Nobody loses a job in that version, which is why it’s easy to drift into. A role disappears without anyone deciding it should. I’d rather we decided. If AI is doing the work of a post we haven’t filled, we should say so and choose what happens to the money. Some of it pays for the AI. The rest might pay for a senior person two days a month, the kind of fractional role a small firm can now build around. It might pay for training the people we already have to check the machine’s work properly.

Where the line is

The wage comparison doesn’t apply to most of what we use AI for today. If our AI costs amount to a few chat subscriptions that help people write faster, they’re software, and filing them next to Xero is right. The line is crossed when AI does a task from start to finish and a person only checks the result. Below the line it’s a tool. Above it, it’s work.

The comparison can also be abused. “It’s cheaper than a person” will justify almost any invoice if we let it. The test is work that has been done and checked. A tool that everyone is excited about, but that hasn’t taken over a single task, fails it.

The payoff isn’t automatic either. BCG scores almost half of the companies in its index as generating value from AI. The rest aren’t there yet. I don’t know how quickly they’ll catch up, or whether a small firm will find it easier or harder than a large one. We can see our own work more clearly than they can. We also have nobody spare to do the checking.

Find the AI spend in last month’s accounts. Then write down the one task it now does from start to finish, and what you’d have paid a person to do it. If there’s no such task yet, leave the line where it is. If there is, the cost belongs with your people, and the next hiring decision is one you can now make on purpose.