The Sleep Test
One question before AI gets the last word.
Who loses sleep if this goes wrong?
If the answer is a team, a process or “the system”, you are not ready. If the answer is a name, AI can run.
The sleep test was first published in Norwegian in Dagens Perspektiv on 3 November 2026. The English version is AI Can Calculate a Risk. It Can’t Carry One.
Why
AI can now run the whole loop on its own: idea, build, run and repair. That is good, and it saves time and money. But a model carries no consequence. It does not face the customer on Monday, it does not lose its job, and the next time it runs, it starts with a blank slate.
So the responsibility has to sit with a person. And it has to be written down before something goes wrong, not after.
Where it applies
The sleep test applies at every point of no return, where a mistake cannot be undone with a click:
- Money: payments, refunds, prices and invoices
- Personal data: storing, sharing and deleting
- Anything the customer sees: emails, replies, offers and notifications
- Data that cannot be restored: deleting and overwriting
- Decisions about people: hiring, credit and access
How a leadership team runs it (30 minutes)
- List every process where AI writes, decides or executes something.
- Mark the points of no return.
- Write one name at each point. Not a team, not a role, a person.
- Agree what that person actually checks before anything goes out. A click on “approve” without reading does not count.
- Repeat every quarter, and every time AI is given a new task.
Template
| Process | What AI does | Point of no return | Name | What is checked before it goes out |
|---|---|---|---|---|
| Customer service | Drafts replies | The reply is sent to the customer | Example: Kari Nordmann | Facts, tone and price promises |
| Billing | Proposes a credit note | The money is paid back | ||
| Recruitment | Sorts applications | The candidate is rejected |
Warning signs
- The answer is “IT”, “the vendor” or “the model”.
- Approval happens with one click, without anyone reading.
- Nobody knows who was alerted the last time something went wrong.
- The same name sits at twenty points. Then nobody is really carrying anything.
From my own work
At Izy we replaced a customer dialogue tool costing close to 150,000 kroner a year with one we built ourselves with AI. AI drafts replies to customer enquiries, and a person reads every reply before it is sent. There is a name at that point.
Have you run the sleep test?
I am collecting examples from organisations that have written down who signs at their points of no return. Write to post@saurav.no.
Saurav Pandey is a Technical Product Manager at Izy AS and writes about AI and accountability. About · På norsk