Almost every company that lets AI agents act on its behalf has some way of checking what they do. Far fewer can say who answers when one of them gets something wrong.
A study published in October 2026 puts numbers on both. Wharton's Human-AI Research group and GBK Collective surveyed 1,007 managers and executives at US companies with more than a thousand employees. Among the 636 whose departments let AI act on its own, almost all had some formal way of checking an agent's work: fewer than one percent said they had none. Then everyone was asked who is held responsible when an AI agent makes a mistake with real consequences. Twenty percent of managers and directors said nobody is clearly accountable, or that they were not sure. Among vice presidents and above in the same companies, only three percent said so. Nor was there agreement on who it should be: no single answer, from the governance team to the employee who set the agent up, was given by more than three in ten.
Both groups are answering honestly. They are looking at the same company from different heights. I described this effect in "Your AI programme looks better from where you are sitting": the higher up you ask, the tidier the picture, because from up there you see a summary of the work, not the work itself.
Ownership, unlike perception, can be checked on paper. Many companies already keep a list of the AI tools they use, sometimes called an AI register: one line per tool, with the department that uses it, the data it touches, the supplier, the cost and the renewal date. Building that list is the first step I describe in "The first four weeks of AI governance". When a company fills it in honestly, every column gets an answer except one, the owner. It stays empty, or it holds the name of a working group.
That empty column matters more than it looks. Last week, in "What AI level is your company at?", I wrote that most AI projects do not fail; they lose the person who was holding them. Losing that person usually starts with exactly this empty column. So what does holding an AI tool actually involve?
An owner of an AI tool carries three duties at the same time. You can check each one by asking a single person a single question.
First, can this person switch the tool off on their own authority, without asking anyone's permission? Raising a concern or putting it on an agenda does not count. Switching it off means the work the tool was doing stops until somebody decides otherwise.
Second, is the tool paid from a budget this person answers for, both the subscription and the usage charges? If the money disappears into a general software budget that nobody reads line by line, nobody is deciding whether this particular tool is worth what it costs.
Third, when the tool gets something wrong, a wrong figure in a quote or a letter sent to the wrong customer, is this person the one who gets the call? Reading about it a week later in a report is a different job.
Someone who carries all three is the owner. A name that carries none of them is just a name on a list. In most companies the three duties end up with three different people, and nobody plans it that way.
Why the owner column stays empty
In a company of fifty to five hundred people, the three duties come apart on their own. Nobody decides that they should. It happens step by step, usually in the same order.
It starts with something useful. Someone in marketing builds a small AI tool on software the company already pays for, and it works.
At the top, the managing director has told everyone that AI is a priority and feels in charge of it. That is common well beyond small companies. In a survey Harris Poll ran for the software company Dataiku in February and March 2026, 70 percent of 900 chief executives at companies with more than 500 million dollars in revenue named themselves the main driver of their company's AI strategy; 16 percent named the CIO. Dataiku summed up the finding as AI ownership that is claimed, but not exercised. Those companies are far larger than the ones this piece is about, and I would expect the pattern to be at least as strong at fifty people, where the person at the top is close enough to everything to believe nothing gets past them.
Meanwhile the licences are handled by whoever manages software, often one person in IT or an outside provider on a support contract. And the obvious person to look after the tool, the colleague who built it, cannot really take it on: looking after it was never part of their job, and they already have a full-time one.
Put together, the duties have split three ways. The managing director sets the direction but cannot switch the tool off. The person who manages software can switch it off but does not know what it is used for. The colleague who built it knows exactly what it does, but has no budget and no authority over it.
None of this matters while the builder is the only one using the tool. It starts to matter the day the builder shares it with two colleagues and they begin to use what it produces in their own work. From that day it is no longer a personal shortcut. Other people's work depends on it.
Nothing technical changes on the day a colleague starts relying on a tool. That is why the day passes without anyone giving the tool an owner.
If nobody is named by then, the question of ownership comes up for the first time when something goes wrong. It falls to whoever is nearest to the problem, usually without the budget or the authority to fix its cause.
Larger organisations face the same problem with agents. For a 2026 paper, the OECD interviewed 25 organisations that already run AI agents, and several of them described agents that employees had set up without approval; they call them shadow agents. Their safeguards include registries of agents and dashboards that show what each agent is doing. Both come down to the same thing: a list with a name against each line. Why employees start using tools nobody approved is the subject of "Shadow AI is a demand signal".
Where to write the name
The obvious place is the list of AI tools described above. In Germany, most mid-sized companies do not have one yet. A 2025 study of German Mittelstand companies, the KI-Studie Mittelstand (reported in the ROI-Kompass 2026 by Telekom MMS and mind digital), found that 76 percent have no AI governance framework at all. According to the study, such a framework has four building blocks: training staff in AI, a list of where AI is used in the company, human review with a clear route for escalating problems, and the ISO/IEC 42001 standard for managing AI. Of the four, the list is the only one that costs nothing and can start before the other three exist. It is also the one that answers the ownership question, as long as the owner column holds a person's name.
That name should not be a committee. I am not arguing against governance committees: a committee is the right body to set the rules, to decide what needs approval before it starts, and to say where AI is not allowed at all. But a committee cannot take the phone call when a tool has sent the wrong figure to a customer, and it cannot switch anything off, because its authority only exists while it is meeting. It can suggest who the owner should be. The owner itself has to be one person.
Not every tool needs a line on the list. A tool one person uses for their own work, whose output never reaches anyone else, can stay off it. It goes on the list the day a colleague starts relying on it, and writing that line takes two minutes.
The ten-minute test
Take the AI tool or agent your company used most last month, and ask three questions about that one tool.
- Who can stop it without asking anyone?
- Whose budget pays for its subscription and its usage?
- Who got the call the last time it got something wrong?
If you get three different names, the duties have come apart, and each person assumes one of the others has it covered. If one answer is blank, nobody is doing that part, and you will find out what was missing the first time the tool gets something wrong. A policy document closes neither gap.
The repair takes two weeks and no budget, because it is not a project. One name per tool, carrying all three duties, written in the owner column. If nobody is willing to put their name against a tool, switch it off for a week and see who asks for it back. Whoever does is doing work the list missed, and is the obvious owner. If you would rather have someone from outside run this check across every workflow at once, it is one of the first things a workflow audit looks at.
Companies already write down an owner for almost everything else they have. Nobody would hand out a company car without noting whose it is: who can take the keys back, which budget pays for the fuel, whose phone rings after an accident. Your AI tools need exactly the same three lines. Most of them have been on the road for months without them.
Questions this raises
It can propose the owner, but it cannot be one. The owner has to be a single person who can act without waiting for the next meeting.
Only the ones whose output reaches someone else. A tool one person uses for their own work can stay off the list until a colleague starts depending on it.
What if nobody is willing to put their name against a tool?
Switch it off for a week. Whoever asks for it back is doing work the list missed, and is the obvious owner.
Original article: https://unlockedconsulting.ai/blog/nobody-owns-the-ai-tools-your-departments-built
