$13 Trillion in Salaries Is the Real AI Market. Most of It Won't Install.

US employers pay around $13 trillion a year in wages and salaries. Global software spend is around $1.4 trillion. Jake Van Clief put those two numbers next to each other in August and drew the conclusion you would expect: every AI company is fighting over the small one.

He cites BEA for the first figure and Gartner for the second, and the ratio is roughly 9 to 1 either way you cut it. His argument from there is that the next wave of large companies will go after the salary line by letting domain experts package their own judgment into reusable systems built from folders and files rather than code.

I install exactly that for a living. I should be nodding along. Instead I published a calculator last month whose own numbers say most of that salary line isn’t for sale.

The 9 to 1 argument, and the part of it I’d defend

Start with what’s right, because it’s the more useful half.

The software budget is a category that exists because software used to be the only thing you could buy. It’s the line you go to when you want a tool. Judgment work has never had a budget line of its own: when a business needs judgment, it hires a person and the cost lands in payroll. So if you build something that genuinely does a slice of that work, comparing your price to a SaaS subscription is the wrong comparison. You’re not competing with a £40 a month tool. You’re competing with a salary.

That reframing is worth having, and it’s the strongest version of the pitch I’ve seen anyone make publicly, including me. My own £6,500 install gets compared to software prices constantly, and every time it does, the conversation goes badly for reasons that have nothing to do with whether the install works.

What my own numbers say

Here’s the problem. Last month I published a calculator that breaks the work nobody in your business is doing into 9 categories and scores each one for how much a named AI role can realistically own. I wrote those scores myself. They run from 10% to 85%.

The three lowest are someone good leaving (10%), a bad hire a tighter process would have caught (20%), and strategy drift (30%). On the default numbers, those three come to £47,500 of a £102,100 gap. So 47% of the money sits in the rows an AI role barely touches, and I scored them that low on purpose, knowing the tool sits on my own site and recommends my own product.

Both things can’t be true as stated. Either judgment packages, in which case my 10% is cowardice dressed up as honesty, or it doesn’t, in which case the salary line is not the addressable market Van Clief says it is.

Two things get called judgment, and only one of them installs

I sat with that for a while and I think the disagreement is a word problem. Two different things are wearing the same label.

The first is the method: the criteria an expert applies, what they look at first, what they weigh heavily, the questions they ask before they’ll commit, the things that make them say no. That’s information. It’s difficult to write down, and almost nobody has, but there’s nothing about it that resists being written down. Van Clief is right that this packages, and it’s most of what I actually install.

The second is accountability: being the person whose name is against the call, who has to be wrong about it in front of the board, who gets fired if it keeps going wrong. That doesn’t package. Not because the models aren’t good enough yet, and not for any reason that improves with the next release. It isn’t an information problem at all. Nobody has ever been able to buy it from a vendor, which is precisely why it lives in payroll.

Now look at the salary line again. When you pay someone £100,000, you’re buying both halves in one bundle, and you’ve never had to price them separately because they arrived attached to the same human. Packaging splits that bundle for the first time. You can take the method out. You can’t take the accountability out.

So the market is real and much smaller than 9 to 1 makes it look. It’s also, and this is the part that matters commercially, badly mispriced right now, because everyone is still quoting the method half against software prices.

What I got wrong in my own tool

My scores aren’t cowardice, but the label on them is wrong, and I’d change it.

When I scored “someone good leaving” at 10%, I was answering the question a founder deciding whether to hire actually has: can this be owned without a person? For that question, 10% is right. Retention is somebody’s job, and an agent cannot hold it.

But it reads as though I’m claiming an AI role can only do 10% of the thinking involved, and that’s not true. The rubric for who’s at risk and why, the evidence-gathering, the prompt to have the conversation this week rather than next quarter: that’s most of the intellectual work, it’s writable down, and I’d score it far above 10%. What can’t move is who owns the outcome.

The honest version is that my low rows measure ownership, not capability, and I conflated the two. I’d add a second column rather than change the numbers, because for the decision the tool exists to serve, ownership is still the right thing to score.

What to do about it

Van Clief cites Polanyi on tacit knowledge for why this has to be built by the expert rather than extracted from them, and that’s the bit I under-sold in my own writing. The reason your best person’s method isn’t documented is not that nobody got round to it. It’s that they genuinely can’t tell you all of it if you sit them down and ask. It comes out while they’re doing the work, in the form of corrections to a draft that’s already wrong.

Which is a slightly awkward thing for me to write, because it’s an argument against buying a done-for-you install and in favour of the expert doing it themselves with help. That’s true, and it’s why the work I do that lasts is the kind where the client is at the keyboard.

The practical test, if you’re deciding whether to hire for something: separate the two halves before you price either. Write down what the method actually is, in enough detail that a competent stranger could apply it. If you can’t get it out of your own head in an afternoon, that’s not a reason to skip the exercise, it’s the finding. It means the thing you were about to pay £100,000 a year for exists in one person’s head, is written down nowhere, and walks out of the building every evening.

Then hire for the half that needs a name against it, and install the half that doesn’t. The salary line is the right budget to be arguing with. Just don’t let anyone, me included, tell you all of it is available.

Run your own version of the split: the Cost of Not Hiring calculator.