Most AI portfolios are designed to start work. Almost none are designed to stop it. A use case is already staffed, already in a steering pack, already mentioned in a board update. The model is not in production. The business owner is tired. Nobody has written the sentence that would end the work this month.
That is a different failure from chasing every new announcement. AI FOMO is the habit of starting work the portfolio cannot finish. Stop rules are the missing control on work that has already started. Without them, in-flight initiatives become permanent residents. They keep consuming scarce owners, data access, and leadership attention long after the original bet has failed.
A stop rule is a kill criterion, not a status colour
A stop rule names the evidence that ends an in-flight initiative, the owner who is allowed to end it, and the date those facts are reviewed. Red, amber, and green on a slide are not stop rules. They are colours. If the same work can sit in amber for two quarters with no change in staffing, the organisation has a reporting habit, not a kill criterion.
| What usually happens | What a stop rule requires |
|---|---|
| The team reports progress and asks for more time | The owner tests the work against a written evidence gate |
| Steering notes that the initiative is “still promising” | Someone with authority records stop, continue, or change-scope |
| A new adjacent use case is added to keep the team busy | Scope cannot expand until the original kill test is passed or the work is ended |
| The model is still in a notebook, so “it is too early to judge” | Elapsed calendar time, missing owner, and unused output are valid reasons to stop |
Why in-flight work is harder to kill than a new idea
New ideas can be declined in a ranking session. In-flight work has names attached to it. A vendor contract is open. A data scientist’s year-end story depends on the demo. The sponsor already told an exec that “we are doing AI in operations.” Stopping that work feels like a personal failure, so the organisation invents reasons to continue. The cost is not the model licence. The cost is the attention that never returns to the few initiatives that still have a path.
Who is allowed to stop the work
If only the delivery team can recommend a stop, the work will not stop. They are measured on activity. If only a quarterly committee can stop it, the work will wait for the next pack. A usable stop rule names a business owner who can end the initiative between meetings, and a second person who must be told. That is closer to governance that names who owns the output and the rollback than it is to a new pilot charter. Ownership of the kill is still not the same job as ownership of the model change.
- Write the evidence that would end this initiative before the next funding slice, not after the demo is late.
- Name the person who can stop the work between steering cycles, and the person who must be informed the same day.
- Treat missing business use, missing owner, and unchanged operating KPIs as valid stop evidence — not as reasons to wait.
- When the work stops, free the people and the data access. A zombie initiative that keeps a Slack channel is still in-flight.
- Do not replace a stopped initiative with a “phase two” of the same bet unless the kill test for phase two is already written.
If nobody in the room can name the evidence that would end this initiative this quarter, it is not a bet. It is occupancy.
Stop rules are not a substitute for readiness
Kill criteria do not replace the four organisational conditions required before AI spend can show in the P&L. Those conditions decide whether more spend should start. Stop rules decide whether spend that already started should continue. Organisations that skip both keep adding pilots because starting is the only decision they know how to make.
The commercial follow-on is the strategy work that installs kill criteria on in-flight AI, not another intake form for new use cases.
Limitations and corrections
Company operating argument about kill criteria for in-flight AI work. It does not report a measured stop-rate study or name a client portfolio.
To request a correction, contact Luminaracorp.