Imagine asking an assistant to choose the best route for a delivery van. It returns a route that saves twelve minutes. Excellent. There is only one small complication: the van does not fit under the bridge.

This is a hypothetical example, but the failure is familiar. The answer can be internally tidy while the decision remains externally terrible. The spreadsheet balances. The presentation sparkles. Somewhere, a bridge waits patiently for reality to arrive.

The distinction matters because AI makes recommendations remarkably easy to produce. A person can now generate a plan before finishing the coffee that would once have accompanied the first paragraph. The time saved is real. So is the temptation to treat completion as comprehension.

Begin with what cannot go wrong

Many requests start with a goal: make it faster, cheaper, bigger, simpler. Goals need companions. What must remain true while the goal is pursued?

For a delivery route, that might include vehicle clearance, driver breaks, and arrival windows. For an article, it might include accurate attribution and a distinction between evidence and interpretation. For a staffing plan, it might include required skills and coverage at particular times.

These constraints are not annoying footnotes. They are part of the problem. A recommendation that ignores them has solved a different problem, often an easier and more photogenic one.

Before requesting an answer, write two sentences: “We are trying to achieve…” and “We cannot compromise…” The second sentence frequently does more useful work than the first.

Ask for alternatives that disagree

A list of three nearly identical options creates the appearance of choice. Useful alternatives expose a tradeoff. One route is faster but less predictable. Another costs more but tolerates delays. A third depends on information you do not yet have.

Ask what conditions would make each option preferable. Now the recommendation becomes something you can inspect. Instead of asking whether the answer sounds intelligent, you can ask whether its assumptions describe your situation.

This also creates a useful place for expertise. The person who knows the loading dock closes early may contribute more than the person who produces the most elaborate optimization diagram.

Give uncertainty an address

“There may be risks” is nearly useless. Which input is uncertain? How much would the conclusion change if it were wrong? Who can check it?

A workable review might have three lines: the assumption, the consequence of being mistaken, and the next verification step. Not every uncertainty deserves a committee. Some deserve a phone call. Others deserve a small experiment before a large commitment.

The aim is proportional effort. Spend more verification on decisions that are difficult to reverse or expensive to get wrong. A lunch recommendation and a hiring decision should not travel through the same process just because both arrive in a chat window.

Keep an owner at the end

The final decision should have a person attached to it. That person need not do every calculation. They do need to understand the important constraints, the evidence used, and the remaining uncertainty.

Try this before accepting your next generated plan: name one fact you checked, one alternative you considered, and one condition that would make you stop. If those answers are missing, the task is not finished. It merely has attractive punctuation.

An answer is a contribution to a decision. The useful promise of AI is that it can make that contribution faster. What happens next is still a human responsibility.

Further reading

✳

Stay curious.
Martin Lumen