More AI apps, the same re-explaining: notes on tool sprawl

Picture a thirty-person firm. Marketing has a writing assistant. The accounts person uses a chatbot to tidy spreadsheet formulas. Sales has an add-on in the email client that drafts replies. Someone installed a meeting recorder. The operations lead pays for a different chatbot on a personal card because she prefers it.

Ask the owner of that firm how many AI tools the business uses and the answer is two, maybe three.

How common is this?

Zapier surveyed more than 500 US enterprise leaders in late 2025. Twenty-eight percent said their organisation used more than ten AI apps. Seventy percent had not moved beyond basic integration between them. Thirty-one percent said they found unapproved "rogue" AI tools in use every month.

Those are large companies with IT departments. A firm of thirty people has fewer apps and no one whose job is to count them, so the share of tools nobody knows about is probably higher, not lower. We have not measured that, so treat it as a guess.

In a small firm, "basic integration" usually means a person copying text from one tool into another. The sales add-on does not know what the proposal tool wrote. The meeting recorder's notes sit in one person's account. Every handover between tools runs through someone's clipboard, and whatever they forget to paste does not travel.

The cost is not mainly the licences

Owners tend to see tool sprawl as a subscription problem. It is mostly a context problem. Each of those tools has been told something about the business, in a chat, a settings box or an uploaded document. Each was told by a different person, at a different time, and none of them knows what the others were told.

So the writing assistant thinks your premium service still exists. The email add-on signs off the way the sales lead likes, which is not how the founder wants clients addressed. The chatbot the operations lead pays for knows the new pricing, because she told it last week. Nobody else's does.

Then the outputs meet. A client receives a proposal written with one tool and a follow-up written with another, and they contradict each other.

Where the context should live

The pattern we use ourselves is simple to describe. The company's rules, verified facts, voice and past corrections live in one written memory that belongs to the company, in plain files rather than in any AI product's settings. Our AI sessions read it before they work. People correct it. A correction made once reaches every session that reads the memory, instead of living inside one person's chat history. Any tool you decide to keep should be able to read from the same place.

That shifts the question from "which AI tool should we standardise on" to "where does our context live, and does every tool we keep read from it."

It does not reduce the number of tools or the bill. You might still want three tools for good reasons. It also creates a new responsibility. A memory that many tools can read is a concentration of company knowledge, so it needs the same care as anything else valuable: scope it by client or company where that matters, decide who can change it, and keep passwords and keys out of it entirely.

An inventory you can do in an hour

Make a five-column sheet and send it round:

The third column is the one that matters. Ask people to paste in, roughly, what they have told the tool: custom instructions, uploaded documents, the "about us" they typed in once. You will find duplicates, contradictions and at least one tool holding something it should not, such as a client list or a password.

When the sheet comes back, pick the five facts that appear in different versions across the third column. Those five, written once and agreed, are the start of a shared memory.

The operating system health check looks at the wider picture of how your tools and processes fit together. Book 20 minutes at pgco.world/book.