Temporal Reasoning
It knows the dates. It does not know what order they happened in, or which one is still in force.
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 temporal reasoning?
Dates appear in text as tokens. Relationships between them — supersession, precedence, expiry, concurrency — must be inferred from language, and that language is often implicit. A contract does not usually say "this amendment supersedes the prior one"; it says "as amended," and the reader is expected to construct the timeline.
Three specific weak points:
1. Relative references. "Last quarter," "the prior agreement," "within 30 days of notice" all require an anchor the model may not have. 2. Supersession. When multiple versions of a fact exist, the model does not reliably prefer the most recent. It prefers the most prominent in context. 3. Currency. Whether something is still in effect is rarely stated in the document. It's a fact about the world as of today, and the model has no reliable representation of today.
What do people assume?
That a system handling dates correctly in isolation handles them correctly in sequence — that if it can parse "effective March 1," it can determine which of three amendments controls today.
Parsing a date and reasoning about a timeline are different operations, and only the first is reliable.
Where does it show up at work?
A procurement analyst asks which payment terms apply to a vendor. The file contains an original MSA at Net 60, a 2024 amendment moving to Net 30, and a 2025 side letter reverting to Net 60 for one product line only.
The model returns Net 30 — accurate for the amendment it weighted most heavily, wrong for the product line in question.
Who carries the downside?
Vendor: benefits from the error, has no reason to correct it. Executive: none. Manager: owns working capital. You: answered the question.
What does it cost?
[MODELED — not reported]
ASSUMPTIONS Multi-version document queries: 100 / year Rate w/ incorrect version applied: 12% (~12 / year) Rate caught before acting: 60% Errors reaching a decision: ~5 / year Cost per error: $2,000 – $35,000 (working capital, missed right, compliance with wrong terms)
Annualized exposure: ~$10,000 – $175,000
How do you control for it?
Build an explicit version timeline before querying. List every instrument, its date, its scope, and what it modifies. Then ask the question against the timeline rather than against the document pile.
The control is not smarter prompting. It is constructing the temporal structure the documents don't contain.
CONTROL COST Multi-version matters: 100 / year Timeline construction: 15 minutes each Annual: 25 hours Fully loaded rate: $70 / hour
Annualized control cost: $1,750
What should you do this week?
RECOMMENDATION
Pick your most-referenced multi-version relationship — a key vendor, a major customer, a governing policy. Build the timeline: instrument, date, scope, what it supersedes.
Nobody has done this, everybody needs it, and it takes an afternoon. It also makes you the person who can answer the question in ten seconds when it comes up in a meeting — which is a small, repeated, highly visible form of authority.
Evidence
RESEARCH Temporal reasoning as a documented weakness relative to other reasoning tasks.
ANALYSIS The three weak points and the supersession-versus-prominence claim are the author's framing. 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