No Sensory Grounding
It has read every description of the warehouse. It has never been in one.
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 sensory grounding?
Text under-describes physical reality, because writers omit the obvious. Nobody writes that the loading dock has two bays, that the third aisle floods when it rains, or that the machine needs eleven minutes to come up to temperature. These facts are load-bearing operationally and absent textually.
So the model's representation of physical work is built entirely from the documented portion of that work — which is the portion that was unusual enough to write down. The routine, constraining, physically-real substrate is missing.
The gap is invisible from inside the text. The model does not know it's reasoning about an incomplete picture, and neither does the output.
What do people assume?
That a system trained on descriptions of physical processes understands those processes well enough to reason about them operationally.
It has learned how people write about the warehouse. Writing about a process systematically omits what everyone present already knows.
Where does it show up at work?
A plant coordinator asks for a changeover sequence to minimize downtime. The output is logically sound and sequenced sensibly.
It schedules two operations requiring the same overhead crane in the same window. The crane constraint appears in no document — it's in the heads of the four people who work that floor.
Who carries the downside?
Vendor: none. Executive: none. Manager: owns throughput. You: published the sequence, and the floor supervisor now has a documented reason to distrust every process you send.
That second cost is the durable one. Credibility with operators is slow to build and fast to lose.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS AI-assisted physical process plans: 80 / year Rate violating an undocumented physical constraint: 18% (~14 / year) Rate caught by operators first: 80% Failures reaching execution: ~3 / year Cost per failure: $4,000 – $30,000 (downtime, rework, overtime)
Annualized exposure: ~$12,000 – $90,000
How do you control for it?
Operator review before execution — structurally, not as a courtesy. Any AI-generated plan touching physical work goes to someone who does that work, with one specific question: what does this plan assume that isn't true on the floor?
This is Automation Boundaries (Part III, Concept 11) applied at the physical/digital seam.
CONTROL COST Physical process plans: 80 / year Operator review: 15 minutes each Annual: 20 hours Fully loaded rate: $55 / hour
Annualized control cost: $1,100
What should you do this week?
RECOMMENDATION
Document the undocumented constraints in your operation. Walk the floor and ask the people who work it: what does everyone here know that isn't written anywhere?
You will fill a page in an hour. That page is simultaneously the highest-value input to any AI system touching your operation, the best onboarding document your team has, and a demonstration that you understand where the model's knowledge stops — which is the judgment this entire book is about.
Evidence
RESEARCH The symbol grounding problem and its application to language models.
ANALYSIS The "text omits what everyone present knows" argument is the author's and is the operationally useful part of this chapter. 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