The Upgrade Nobody Announced

Monday the drafting assistant answered differently. Same login. Same name on the tile. The sentences were shorter, more certain and slightly wrong in a way the old one hadn't been wrong. Two analysts spent the morning trying to prompt it back. By Wednesday they had rebuilt the old instructions in a personal account and were pasting the output into the official tool so the audit log would still look clean.

There was no ticket. Nobody had retired anything. The vendor had swapped the model over the weekend and called it an improvement.

How to Retire an Agent Without Retiring the Team is the work you do when you mean to take a system out. This is the work you skip when the system takes itself out and keeps the name. The login stayed. The teammate didn't. People organized their competence around a voice. You changed the voice and left the icon.

A swap is a decommission

A decommission has a date, a name and a rite, even a bad one. A silent swap has none of those. The team still clicks the same button. The regulation pattern they had built, what to trust, what to rewrite, which phrases meant the model was guessing, no longer applies. They discover the change the way you discover a colleague has been replaced: the reply is off, and nobody mentioned a departure.

When the Digital Teammate Goes Quiet is grief after a cutover you can point at. Here there's nothing to point at. The system didn't go quiet. It kept talking, in a voice that's almost the old one. Almost is worse. People don't mourn a tile that still answers. They compensate. Compensation, in a governed environment, looks like a personal account, a copied prompt and a paste that makes the official log lie.

The Attachment Question asked who would grieve if you turned the agent off. Ask the cousin: who already rebuilt it because you changed it without saying so. If you can name them, attachment was real and you spent it on a release note nobody received. If you can't name them, the shadow process is already ahead of you.

What to announce

You don't need the weights. You need one sentence, before Monday, in the channel the team already reads:

• What changed. Model, prompt or policy, not "performance improvements."
• What will feel different. Length, certainty, refusal, tone. The thing a person will notice in the first draft.
• Who can pause it. A name, not a vendor status page. If the new voice is wrong for the work, someone in the building can roll the prompt or the model back this week.

If you can't write that sentence, you don't know what shipped. Then you're not governing an agent. You're renting a voice that can be replaced under you, and you will learn about it from the personal accounts.

The Override Without a Name is who may stop a path. A silent upgrade needs the same sentence pointed at the vendor's calendar: who is told, in what channel, before the voice changes. If that person doesn't exist, the next swap will also be a surprise and the next paste into the official tool will also look like compliance.

This week

For one live agent, find the date of the last model or prompt change and the list of people who were told before they used it. If the date is missing, write that down where the owner can see it. Then draft the one sentence you will send before the next change: what will feel different, and who can pause it.

Don't wait for a shutdown ticket. The shutdown already happened. It kept the icon.

The person remains the one who can be asked why. A login that survived the weekend isn't evidence the teammate did.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.