Rituals after the model moves

Rituals That Keep the Human as Moral Agent

Dashboards don't keep people on the hook. Cadence does. Four short rituals that restore a name to the decision after the system has already sounded sure.

Dr. Sarah Dyson·September 7, 2026·4 min read·869 words

Share this note

Ticket frame for Rituals That Keep the Human as Moral Agent

Most AI governance I'm invited to inspect is a meeting that produces minutes. Minutes aren't agency. Agency is a person who can be asked why, who still knows they're the one who can be asked, after a system has already produced a finished paragraph that sounds like a decision.

You can't workshop your way into that. You can ritualize it. Rituals are short, repeated, slightly inconvenient practices that encode who we are when we're tired. Teams already have them: the stand-up, the close of books, the pre-brief before a hostile board. What they rarely have is a rite that restores a human subject to a model-shaped call.

Why rituals, not tools

Tools drift. Vendors rename. Models swap under you, as anyone who has lived through a silent upgrade knows. A ritual belongs to the group. It survives the license. It's also how you metabolize the attachment and status costs described in The Status Threat Nobody Named without turning every week into group therapy.

A good ritual has four properties:

  • Named. It's called the same thing every time.
  • Bounded. Fifteen minutes, not "as long as it takes."
  • Costly enough to be real. Someone has to prepare one artifact.
  • Public inside the group. A private conscience isn't a governance system.

If your current governance hour has none of these, it's a salon.

Four that fit in a real calendar

1. The four-sentence why. Once a week, one live decision the system shaped. The owner reads four sentences: what was decided, on what basis, what would have changed the call, who can stop it. No slides. If they reach for "the model determined," they start over. This is the practice behind Explainability Is a Leadership Skill, made into a cadence rather than a memo.

2. The dissent seat. In any review where a model artifact is present, one person is assigned, in advance, to argue that the output is wrong. They must produce one concrete failure, even if they believe the output is fine. Rotate the seat. If the same two seniors always occupy it, you have a performance, not a rite. Dissent latency (how long before a real challenge appears) is a number you can watch.

3. The exception name. Whenever the system is overridden, a human name goes in the log before the override is described. Not "the team." Not "ops." A name. This feels pedantic until the day you need to know who knew. It's also a fairness device: credit and risk stop pooling in the fog.

4. The retirement rehearsal. Once a quarter, for thirty minutes, ask: if this system went quiet on Monday, what would we lose and who would absorb it? Write the blanks. You're not planning a funeral. You're refusing the fantasy that competence lives in the license. When a real retirement comes, you will already have a map.

What to stop doing

Stop adding a steering committee every time a risk register looks thin. Committees are where agency goes to be shared until it's nobody's. Stop requiring people to attest that they "reviewed" a volume no calendar could review. Forced attestation is how you train lying. Stop celebrating "autonomous" as if the adjective were a virtue. Autonomy in a system is a transfer of judgment. Transfers should be explicit.

Trust Capital on the Balance Sheet is the ledger these rites write to. A four-sentence why is a deposit. A skipped dissent seat is a withdrawal. If you run the rites and never write them down, you will still feel better and you will still be unprepared for a call.

A note on tone

These practices will feel slightly ceremonial in a culture that prizes being casual and "not corporate." Good. Casual is how finished language from a system slips into the room without a subject. A little ceremony is the point. You're marking that a person is still here.

Don't dress the rites up as mindfulness, and don't outsource them to a copilot that "facilitates governance." The moment the ritual is generated, it has failed its only job, which is to put a human body in the way of a smooth process.

Keep the human as the moral agent

I won't claim a system cares whether you do this. It doesn't. Your people will. More precisely: your future self, the one who has to answer for a decision made at speed, will care whether a name was attached while there was still time.

Rituals are how a group remembers what it decided to be. If you have shipped agents and have no rite that restores a subject to the sentence, you have shipped a habit of unclaimed speech. That habit won't stay inside the tool.

This week

Pick Friday, fifteen minutes, one decision. Four sentences, one name, one stop. Put it on the calendar with the same gravity you give the cash meeting. If it feels too small to matter, that's a clue about how large you have allowed the unclaimed speech to become.

The Ethical AI Leadership Decision Toolkit is the portable form of these questions. EI Leadership Insights is the biweekly reminder to keep using them when the week gets loud.

Share this note

Related in this journal

EI Leadership Insights

Biweekly notes by email. One practice.

Privacy·Leave the list