Everyone is going all-in on AI, including us, but the cost of managing all of that AI is often bigger than the benefits

The subscription is not the cost of an AI tool. The person who integrates it is, and that person is already on someone's headcount under a title that has nothing to do with AI.

A company running 7 or 8 of them is not running a software problem. It is running an attention problem. The 7 or 8 figure is a vendor's number rather than survey data, so treat it as a place to start looking, not a threshold. What is hard to argue with is that these tools are bought one at a time, and the cost of owning them is not a sum of invoices.

Picture a $5M to $50M growing company. There is one person on the ops team who owns the integrations, sometimes one and a half, and they also own the payroll vendor, the CRM, the ERP, the bank portal, the 401k administrator, and the scheduling tool nobody will give up.

Now the ninth AI tool lands on that desk. It looks exactly like the last one did, which is the whole problem. It reads from the same systems the other eight read from. Somebody has to security review it, wire it in, and then be the person who answers the question about it every week after that.

None of that is on the invoice, because the vendor isn't charging for it. The vendor is charging for the seat.

All the cost shows up in the quarter you signed, and all the work of leaving shows up in the quarter you decided to quit, and that second quarter belongs to a different decision with a different person in the room.

Data doesn't move cleanly. A tool that read from the ERP won't necessarily export back into one. Prompts get built, and everything anyone worked out about how the tool is supposed to be used ends up inside those prompts, and nobody wrote any of it down. A vendor that does four things badly can be less trouble than two vendors that do two things well, which is why the tool count is only ever a stand-in for something nobody has measured. Somewhere inside the 7 or 8 there is one nobody has opened in 90 days, and it is still on the invoice.

The person who signs off on the ninth subscription never sits in the meeting where somebody decides to get rid of a tool, and the person in that meeting never sees the invoice for the ninth one.

The obvious objection is that this is ordinary software sprawl, that finance has been telling companies they buy too many tools for a decade, and that a chatbot does not change the arithmetic. I don't think that carries.

Ordinary software sits beside the other software. It reads a file, it writes a file, and the surface area is bounded by whoever designed the API. AI tools all reach into the same systems. The same ERP, the same CRM, the same shared drive, the same chains of prompts that now quietly encode how work gets done. So the cost of adding the ninth one compounds against the cost of removing the third, and the ordinary case runs the other way: as vendors standardize their interfaces, sprawl gets cheaper the longer you let it sit. This gets more expensive.

Airlines ran into this a long time ago, and the fix they landed on is the tell. Maintenance burden in a fleet is not a function of how many aircraft you own. It is a function of how many types and models you have, and how far each one drifts from the parts, the tooling and the manuals the rest of the fleet shares. Carriers spent decades consolidating fleets, and it was never about the paint. Every distinct type carries its own spares, its own training and its own manuals, and that bill arrives forever.

The transfer holds because the mechanism is identical. The count is a stand-in. Variety is the cost. A portfolio running 7 or 8 tools across three business units that each organize their data differently is a mixed fleet, and the coordination burden scales with variety rather than the tail number. Picture a second company on the same five tools with one chart of accounts, one geography and one sales motion. Five tools there isn't sprawl, it's tools. Same count, opposite problem.

I think the first company to publish an AI tool budget in the open will publish it as headcount rather than as software. A line called AI coordination sitting next to a salary is a strange thing to put in front of a board, and I expect to see one inside a couple of years, because it's the only one of the two that anyone in the room can actually cut.

It will be recognizable when it shows up. The person doing the coordination already has a ranked list of what breaks first, and it isn't ambiguous. They know. They've known for a while.