The AI gain has a ceiling

Most of the AI value in a tech company lands in operations. Support, marketing production, the back office. It's a real gain and it arrives fast.

It also has a ceiling, and the ceiling is where people stop looking.

This came out of a conversation with a friend. I've been chewing on it since, so here it is properly.

Where people look first

What's your company doing with AI? You'll usually hear about the product first, something new in the interface with a model behind it.

There's no moat there. Whatever you put in front of the customer, someone ships the same thing two quarters later.

Adobe is the obvious objection. Photoshop is a beast, you can't clone it, people don't switch out of it casually, and Adobe can put AI in front of an enormous installed base tomorrow. All true. It also argues the opposite of what it looks like it argues.

The moat is Photoshop. Adobe's AI is defensible because Adobe is defensible, and not the other way round. Get that direction backwards and you go looking for the moat one layer too high.

So the feature is a defensive expense. You ship it so you don't bleed customers at the edges, which puts it on the maintenance line.

Look at what these features actually are. Your spreadsheet offers to take the mean of a column, or asks whether you wanted a pivot table off those three. Handy. Has anyone ever changed spreadsheets over it? And yet every spreadsheet has to have it now, because the one that doesn't looks broken.

That's table stakes, essentially. You pay it to stay at the table, and it gets filed under innovation.

The tell is timing. When four competitors ship the same capability inside a year, it was a deadline everybody met.

The hundred ads

Here's the part I believe in.

A marketing team that shipped three ad variants a week can ship a hundred. A support queue that needed forty people needs twelve. A back office full of people moving documents between systems stops needing most of them.

That's a real line coming out of a real budget, and it lands this year. Nobody has to believe anything about AGI for it to work.

Still, it caps.

It caps in size, because it's cost-side. Costs floor at zero. You can't save more than a hundred percent of a line item, and every year you take a bite, the base you're biting into is smaller. Revenue has no equivalent floor. So the operational gain is bounded by your own expense base, and the ceiling drops each time you approach it.

It also caps in time, which is the harder one to see. If AI lets you spin up a hundred ads, what stops your competitor spinning up a hundred ads (they have the same models you do)? Whoever gets there first eats well for a few quarters. Then everyone has the same cost structure, the savings get competed away into price, and you're left with a lower cost base and a much noisier ad auction than the one you started in.

Features mellow out into table stakes. Operations mellows out the same way, just slower, and the slowness is what makes it look permanent.

So capped upside undersells it. Capped in how much you get, capped in how long you keep it.

The bank

A bank that automates a chunk of its back office is more profitable. That's certain, and it's a gain to the firm.

Is it a gain to the economy? That depends on what happens next to what the bank freed. If the money and the people go into building something that didn't exist before, output rises and the economy is genuinely larger. If they don't, nothing was created. The wage bill moved into the margin, and a transfer from one set of people to another gets counted as growth.

This is an accounting question before it's anything else. Cost savings hit the firm's numbers immediately and unconditionally. They hit the economy's numbers later, partially and only if something absorbs what got let go.

Most commentary collapses the two numbers into one, and the collapse flatters the result.

When AI is the product

There's one place the ceiling doesn't apply.

When the model is the product, or the thing is built so tightly around the model that the two aren't separable, you're selling an input other things get built on top of. It makes processes possible that nobody was running.

However you feel about the valuations, that's the difference in kind, and it's why the value concentrates at that layer while everyone adopting it shares out much less. It's also why those companies keep pushing toward robotics (same move, new substrate). Become the input a new set of workflows depends on.

Everything else is downstream of that. Useful, worth doing and downstream.

Where I might be wrong

Is the feature the thing? Maybe not. A feature that's been sitting on your specific data for two years knows things a fresh clone doesn't. Anyone can copy the behaviour. Nobody can copy the two years. So if the moat turns out to be what a feature remembers, my no-moat claim is too strong.

And I'm treating the operational gain as a step change. You automate once, you bank it, you're done. Maybe those systems compound instead. If the operation runs measurably better each year (not just cheaper once), then it's a rate, and a rate doesn't hit the ceiling I've described.

I think both are open. I'd want a couple more years of numbers before betting hard either way.

Meanwhile the question I'd put to anyone selling me on AI transformation is where the gain actually lands, and what the ceiling is once it lands there.

I want to hear about the compounding case especially. If you use this stuff at scale and the gains are still climbing rather than flattening, let me know, because that's the version where I'm wrong.