When the Digital Teammate Goes Quiet

The ticket said "decommission assistant v2." It took fourteen minutes. The vendor called it a clean cutover. By Thursday, two analysts had rebuilt the old prompts in a personal account, one project manager was writing updates "the way it used to phrase them," and a lead who had never been sentimental about software asked, in a voice I recognized from other kinds of loss, whether we could roll it back "just for the quarter-end pack."

We don't need to claim the system felt anything. It didn't. The people did. Attachment research is about what humans do with a stable other, a figure who answers, who is available, who reduces the cost of being uncertain. A well-tuned agent, used daily, occupies some of that slot. When it goes quiet, the team doesn't merely lose a feature. They lose a pattern of regulation.

Attachment without sentimentality

Bowlby's claim, stripped of the nursery, is that we organize our working lives around figures we can predict. In knowledge work the figures used to be people: the operations lead who always knew which client would explode, the architect who could smell a bad interface. Some of that predictability has been offloaded onto systems that answer at 1:14 a.m. without a mood.

That offload is a regulation strategy. People who work under load will attach to whatever reliably narrows the next ten minutes. If you mock that, you won't get honesty when you retire the thing. You will get a shadow copy and a story about "personal productivity tools."

This is adjacent to the status dynamics in The Status Threat Nobody Named. A person who felt demoted by an agent can still panic when it disappears, because both events confirm they're not holding the work. Grief and status threat aren't opposites. They're two readings of the same displacement.

What the grief actually is

It's not mourning a mind. It's mourning:

• A voice. The cadence of summaries, the adjectives the system preferred, the way it framed risk as a list of three.
• A witness. Someone (something) that had seen the project's entire messy history and could be asked at 10 p.m.
• A competence prosthesis. The parts of the job the team had stopped practicing because the agent was faster.

The third is the operational one. Skills atrophy in the shadow of a reliable other. When you retire the agent, you don't return the team to the competence they had in 2024. You return them to a gap. Change-management theater that says "we're going back to the old way" is a lie the body doesn't believe.

A continuity rite, not a town hall

I don't recommend a funeral for software. I recommend a rite that's short, named and written down, because rituals are how groups metabolize transitions without pretending they were never attached.

A working version, one hour, no slides:

1. What it carried. Three people name a task they had stopped holding in their own working memory.
2. What we will hold instead. One named human for each of those tasks, for the next thirty days, even if a new system is coming.
3. What we won't pretend. A sentence in the decision log: We're retiring a regulation strategy. The license is the small part.

That last line matters. It keeps the human as the moral agent of the transition. The system isn't being "laid off." A practice is being moved. If you can't name the practice, you're not ready to cut the license.

As The Human Overhead of Agentic Systems argues, the labor that kept the old agent trustworthy (the exception handling, the quiet corrections) doesn't retire with it. It looks for a new body. If you don't assign it, it assigns itself to the most conscientious person in the room, which is how burnout presents as loyalty.

Near-misses I keep seeing

A hospital operations group swapped a scheduling assistant over a weekend. Monday looked fine. Wednesday, two charge nurses had reconstructed the old logic in a spreadsheet because the new one "didn't know our unit." The reconstruction was competent and completely unaudited. The near-miss wasn't the spreadsheet. It was that no one had been invited to say we will miss how it knew the unit. The unofficial copy was the grief, routed around the official story.

A professional-services firm did better. They kept the old agent in read-only for two weeks, not as a fallback, as a reference for tone. People could look up how a difficult client had been summarized last quarter. Then they turned it off, in a meeting, with the new owner of client-memory sitting in the chair. It felt slightly ceremonial. It also meant nobody opened a pirate instance.

Keep the human as the moral agent

Don't tell the team the system "was just a tool" if you spent a year asking them to trust it. That sentence is a betrayal of your own previous ask. You may retire a system. You may not retire the history of reliance without a receipt.

If you're planning a cutover, people will "get over it" the way they get over other poorly marked losses: by building a private workaround and trusting you slightly less. Trust, as a ledger, notices.

The neighboring practice is cadence. Rituals That Keep the Human as Moral Agent is about what you do every week so that a retirement doesn't become the first time anyone talks about who is actually holding the work.

This week

Pick one system your team leans on. Write two sentences: If this went quiet on Monday, we would lose ___ and ___ would be expected to absorb it. If you can't fill both blanks, you don't have a retirement plan. You have a hope.

When you're ready to operationalize the pause, the Ethical AI Leadership Decision Toolkit holds the questions that keep a cutover from becoming a disappearance. Field notes land in EI Leadership Insights.