Walk into a two-hundred-person company and ask the leadership team to list every AI tool currently in use. They will name three or four. Walk through the same building department by department, and the real number will be eleven. Or fifteen. Or twenty-two. We have not yet found a mid-market company where the gap between the perceived count and the actual count was less than three.
This is agent sprawl. It is the AI version of the shadow IT problem from the 2010s, except the software is now making decisions instead of storing files, and the rate of adoption is faster than anything that came before it.
Why it happens
Two things have changed at the same time. The first is that AI tools are now cheap enough that anyone with a corporate credit card can buy one without asking. The second is that the value of an AI tool is most visible to the person closest to the work, who is rarely the person on the AI committee.
The marketing manager finds a tool that writes ad copy. The sales lead subscribes to a meeting recorder. The customer support team installs an inbox assistant. None of these decisions are wrong on their own. None of them go through procurement. None of them are tracked anywhere central. Three months later, the company has a fleet of unmanaged systems making small decisions every minute of every working day.
What goes wrong
1. The same problem gets solved three times
Marketing, sales, and support each end up with a different AI tool that drafts outbound messages. The tools have different styles, different guardrails, different prices, and no awareness of each other. The brand voice quietly fragments. The procurement bill doubles, then triples. Nobody on the exec team can produce a single number for total AI spend, because the spend is scattered across thirty expense reports.
2. The handoff between agents breaks
An AI agent in one department creates a record. An AI agent in the next department reads it, misinterprets it, and acts. There is no human in the chain. Nobody knows the handoff exists, because no single person designed it. When the customer complaint arrives, the company cannot reconstruct what happened, because the trail crosses three vendors and four data stores.
3. Nobody owns it when it breaks
The tool was bought by someone who has since moved teams or left the company. The credit card is still being charged. The integration is still running. When the model drifts and the outputs start to slip, there is no named person to flag it, fix it, or turn it off. The system becomes a slow, invisible source of error that nobody is paid to notice.
Why the obvious fix is the wrong one
The instinct is to ban shadow AI. Lock down procurement, require approval for every new tool, treat unauthorised AI as a security violation. This does not work. We have watched several companies try it. The tools do not disappear. They go further underground, get bought on personal cards and expensed quietly, or get used through web interfaces that never touch corporate procurement at all.
The reason a ban fails is the same reason the sprawl started. People are using these tools because the tools work, and because their job got easier the day they installed them. Removing the tool without replacing the value is a tax on the people doing the work, and they will route around it.
What actually works
1. Take inventory honestly
Department by department, ask three questions. What AI tools are you using. What are they doing. Who pays for them. Promise no consequences for honest answers, and mean it. The inventory will surprise the exec team. That is the point. You cannot manage what you cannot see, and the goal of the first step is only to see.
2. Designate an official tool for each common workflow
If three teams are using three different AI tools to draft outbound messages, pick one. The choice matters less than the act of choosing. Cancel the other two, return the budget to the teams that lose them, and write the choice down on the AI policy page so the next hire knows what to use.
3. Name an owner for each AI surface
Not a committee. One person. Their job is to watch the tool, monitor the outputs, and call the meeting when something is off. The owner does not need to be technical. They need to be close to the work and willing to be the person who answers when something goes wrong. Most companies discover, when they do this exercise, that half their AI tools do not currently have an owner. Naming the owner is the work.
4. Review the inventory every quarter
New tools will appear. Old tools will become unnecessary. The vendor market is moving fast enough that the right tool for a workflow is rarely the same tool eighteen months later. Bake the review into the operating rhythm. Treat it the way a finance team treats the quarterly forecast.
Why an outside operator helps here
The cleanup is technically straightforward. The hard part is political. Telling the marketing team to give up the tool they like. Telling the sales lead that the recorder they expensed is no longer approved. Telling the founder that the AI tool they personally bought a year ago is one of the ones being deprecated. These conversations are easier when an outside operator is the one having them, because the outsider is not auditioning for the next promotion, and the inside relationships do not have to absorb the friction.
This is most of what an AI operating partner ends up doing in the first ninety days inside a company that has been using AI for more than a year. Not building anything new. Cleaning up what is already there, so the new work has somewhere to stand. The cleanup, done honestly, is usually the engagement's first real result.
