Who Owns the Assistant After We Leave
We were asked to review why one assistant had degraded. We found eleven of them across the organisation, built by three suppliers and two internal teams, and the review turned into a different piece of work entirely.
What the Inventory Found
Eleven assistants. Two had a named owner, a maintenance budget and someone who would be told if they stopped. Three had an owner who had left the company. Six had never had one, because each was delivered as a project and projects end.
All eleven were running. Four were being used daily, three occasionally, and four had not been used in months but were still calling models and still costing money.
Why This Happens Without Anyone Deciding It
A project has a sponsor, a budget and an end date. A system has an owner, a running cost and no end date. Almost every assistant we have seen was procured as the first and became the second at handover, and nobody performs the conversion because it is not anybody's job.
Acharya published work in early 2026 proposing a governance maturity model for managing agent sprawl in business operations, which names the pattern and gives it stages. Having a name for it made the conversation with the customer considerably easier than describing it from scratch.
The Four Questions
Who is told when it breaks. Who decides whether it should change. Who pays for it. And what happens to it if the answer to any of the first three is nobody. Four questions per system, twenty minutes each, and the answers are usually known but have never been written next to each other.
For six of the eleven, the answer to the first question was that the users would notice and complain, which is not a monitoring strategy. For four, the answer to the fourth was that nobody had considered turning anything off.
| System state | What we recommend |
|---|---|
| Used, owned, monitored | Nothing. This is the target |
| Used, no owner | Assign one or plan a retirement |
| Unused, still running | Turn it off. Keep the corpus |
| Owner left the company | Reassign before anything else |
What We Turned Off
Four systems, after checking with the departments that had commissioned them. Two were genuinely obsolete because the process they supported had changed. Two had been replaced by a newer assistant that nobody had connected to the older one.
Turning them off saved a modest amount of money and a larger amount of confusion, because two of them held stale copies of documents and were occasionally answering questions with content the current system had superseded.
What We Now Hand Over
A document naming the owner, the monitoring, the evaluation set and its last result, the model versions and where they run, the cost per month, and a retirement procedure. It is four pages and the last item is the one customers find unusual.
A retirement procedure at handover sounds pessimistic and is the opposite. A system that can be turned off cleanly is one that can be replaced when something better exists, and its absence is why organisations keep running things nobody wants.
The Part That Is Our Responsibility
We built two of the six ownerless systems. Both were delivered with documentation, both worked, and neither had an owner named on the customer side because we did not insist and the project sponsor moved on.
That is a failure of ours rather than of theirs. We now decline to run a final acceptance meeting without a named owner present, which has delayed two handovers by a fortnight each and has not lost us any work.
Why an Unowned Assistant Is Worse Than None
Because it is trusted. Users do not know that nobody is maintaining it, so they treat its answers exactly as they treat the maintained one, and the answers get worse slowly enough that nobody attributes it to anything.
The drift is the mechanism: models change, documents are superseded, the corpus goes stale, and a system with nobody watching has no way for any of that to become visible. A system that is obviously broken gets fixed; one that is quietly degrading gets believed.
What We Do Not Claim
We do not claim eleven is unusual. It is one organisation of a few thousand people that adopted assistants early, and a smaller or later organisation would have fewer for reasons that have nothing to do with governance.
We also do not claim our four questions constitute governance. They are the minimum that makes a system attributable, and an organisation running assistants in a regulated process needs considerably more than that, which is a different piece of work than the one we did.