No Common-Sense Grounding
It will produce a schedule that requires a truck to be in two states at once, and the formatting will be immaculate.
This page covers one specific way AI gets things wrong at work, and what to do about it.
It runs in order. What goes wrong, why it happens, where you'd notice it on an ordinary day, who takes the blame, roughly what it costs, and the check that catches it. Then one thing to try this week.
The dollar figures are estimates, not measurements. The assumptions behind each one are printed right there, so you can swap in numbers that fit your job. Anything actually measured carries an OBSERVED tag.
What is no common-sense grounding?
The model has no representation of physical constraint, resource exclusivity, or causal ordering except as those things appear in language. It has read that trucks take time to travel and that people can't be in two meetings simultaneously. It has not represented why, so it cannot check a candidate output against the constraint.
Where the failure lands: outputs that require satisfying multiple real-world constraints at once. Any single constraint is usually respected, because respecting it is the common pattern in training text. The combination is where it breaks, because the combination is novel and there's no simulator underneath to test it against.
What do people assume?
That obvious physical or practical impossibilities get caught — that some floor of basic sense filters output before it reaches you.
There is no such floor. Plausibility in text is not plausibility in the world.
Where does it show up at work?
An ops coordinator asks for a delivery schedule optimizing eleven stops across two days with driver-hours constraints. The output is clean, formatted, confident, and assigns a driver a route requiring 14 hours of driving inside an 11-hour legal limit.
It looks like a schedule. It is a well-formatted description of something that cannot happen.
Who carries the downside?
Vendor: none. Executive: none. Manager: owns the service failure. You: published the schedule. Also: if the driver runs it, this is now a regulatory violation with your name on the dispatch.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS AI-assisted operational plans: 120 / year Rate w/ infeasible constraint combo: 10% (~12 / year) Rate caught before execution: 75% Failures reaching execution: ~3 / year Cost per failure: $3,000 – $25,000 (expedite, service credit, overtime, compliance exposure)
Annualized exposure: ~$9,000 – $75,000
How do you control for it?
An explicit feasibility pass against a written constraint list. Not a read-through — a checklist. Enumerate the hard constraints for your domain once (hours, capacity, sequence, exclusivity, regulatory limits) and check every generated plan against the list mechanically.
CONTROL COST Operational plans: 120 / year Feasibility check: 8 minutes each Annual: 16 hours Fully loaded rate: $55 / hour
Annualized control cost: $880
What should you do this week?
RECOMMENDATION
Write your function's hard constraints on one page. Physical, legal, contractual, capacity. Most people have never written these down — they're carried as experience, which means they leave when the person does and they don't survive a handoff.
Then use it as a checklist on AI-generated plans.
You have converted tacit expertise into a documented control. That is a promotable artifact, and it is worth noting that it is valuable independent of AI — which is exactly the kind of contribution that survives the next reorganization.
Evidence
RESEARCH Language model performance on physical and commonsense reasoning benchmarks.
ANALYSIS The single-constraint-versus-combination framing is the author's. Retain the label.
This is one of 24 failure modes. The book gives you all of them — plus the controls that catch each one and a 90-day plan to prove you ran them.
Preorder the Book