An AI adviser that always proposes another experiment is optimizing for continuation, not for preserving your runway. Founders need a stopping rule set before the session begins, because the model cannot decide how much time or cash the next answer deserves.
In May 1996, climbers high on Mount Everest were still moving toward the summit after planned turnaround times had passed. Expedition leaders Rob Hall and Scott Fischer had set times for turning back, yet several climbers continued upward. The summit remained close enough to pursue. The margin for getting down was disappearing.
Jon Krakauer documented the expedition in Into Thin Air. Eight climbers died during the disaster, including Hall and Fischer. Several factors contributed, and it would be dishonest to reduce the outcome to one decision. The ignored turnaround times still matter because they were supposed to end the debate while turning back remained possible.
The model always has one more move
A founder opens an AI session on Friday afternoon with a practical question: should we keep testing this feature?
The model recommends a narrower segment. The founder asks how to test it. The model produces an interview script. One answer suggests changing the onboarding flow, so the founder requests a new flow. Then a pricing experiment. Then a landing page variant. Each response is coherent. None says the week should end here.
That pattern can feel like momentum, particularly when the alternatives are uncomfortable. Stopping may mean admitting the feature lacks demand. It may mean returning to sales calls, delaying the demo, or telling a small team that several days of work produced no decision.
The AI has no Friday afternoon in Accra, Lagos, Berlin or New York. It does not carry payroll into Monday. It does not lose a contract while the roadmap drifts. Within the conversation, another plausible experiment is cheap. Inside the company, it may consume the week that was meant to extend runway.
Continued engagement does not need to appear as an explicit instruction for it to shape the exchange. A conversational system responds when prompted. It remains useful by producing the next useful-looking step. Unless the founder supplies an endpoint, the session has no natural reason to stop.
Define the turnaround time before asking
A strong AI session begins with a constraint the model cannot quietly negotiate away.
For example:
“We have four engineering hours and no budget for paid traffic. Recommend one test that can disprove this assumption by Tuesday. If the available evidence cannot change our decision, tell me to stop.”
That instruction changes the job. The model must work within a decision boundary instead of generating an expanding research programme.
I would also write the stopping condition outside the chat:
- We stop after speaking with five qualified prospects.
- We stop if nobody agrees to share existing data.
- We stop if the test requires production code.
- We stop on Tuesday and make the decision with incomplete evidence.
The exact boundary depends on the company. A founder with a signed pilot in Ghana faces a different constraint from a pre-seed team selling into Germany or an African diaspora market in the US. The principle survives those differences: define what the experiment may consume before discussing what it could reveal.
This is the same discipline behind asking what Monday must prove after Friday’s demo wins the room. Interest creates options. A proof condition turns those options into a decision.
Separate advice from authority
AI can compare paths, expose assumptions and draft tests quickly. It should not receive authority over runway by accident.
Before accepting another experiment, I use three questions:
What decision will this evidence change?
What will the test cost in founder time, engineering time and delayed work?
What result ends the investigation?
If the first answer is unclear, the experiment is activity. If the second answer is missing, the recommendation has ignored the company. If the third answer is “we will see what we learn,” the loop can continue until cash or patience ends it.
This matters most when the advice sounds sensible. Bad suggestions are easy to reject. A sequence of reasonable suggestions can consume a month while leaving the original decision untouched.
Kelechi’s choice between three working features and six weeks to prove one carried the same constraint: possibility had to yield to a bounded proof. More viable directions did not create more runway.
Put the stop decision back in the room
On Everest in 1996, a turnaround time existed because conditions near the summit could make continued progress feel rational after retreat had become necessary. The rule was meant to protect the descent from the ambition of the ascent.
A founder needs the equivalent before opening the chat.
Write one sentence at the top of the working document: “We will spend no more than ___ on this decision, and we will stop when ___.” Then ask the model to challenge the plan against that boundary.
When the next polished experiment appears, do not ask whether it might work. Ask whether it earns the runway it consumes. If it does not, close the session.
Comments
No comments yet.