The error is cheap when it happens. It gets expensive when you find out.

TL;DR

The cost of an agent error is not when the error happens — it is when you discover it. Find it immediately: cheap. Find it after a day: much more expensive. Find it when a client reports back: potentially twenty times the cost. Agent output looks professional so you tend to skip verification — that gap is exactly how the distance opens. Two sentences added to the end of a prompt can close it before it widens.

You ask the agent to write an analysis. It returns nine pages, with tables, bullet points, formatted like something that would take you half a day to produce.

You skim it. Looks fine. Send it out.

Three days later a client finds that one of the numbers in the summary table was calculated incorrectly — using completely different logic from what was intended. You have to rewrite that section, explain it to the client, and spend another half-session patching the other places that relied on that number.

How much did that cost? Not the time to read nine pages initially — because you did not read them. It is the rewrite plus the explanation plus the patch plus the partial erosion of trust. All from one error that would have taken five minutes to fix if caught on delivery.

01Cost grows with the distance to discovery

Caught immediately on delivery

1× — fixed in 5 minutes

Caught after 1–2 days of use

5–8× — fix plus trace all downstream uses

Caught after others have relied on it

15–25× — fix, explain, patch consequences

The same error, three completely different price tags. The only variable you control is the distance between when the error occurred and when you found it — and that distance is entirely in your hands.

This principle is not new. In software development it has long been known: a bug caught in code review is far cheaper than one caught in staging, and both are far cheaper than one that reaches production. With agent output the principle is identical — just less obvious, because the output looks polished and rarely causes an immediately visible error.

02Which outputs carry the highest late-discovery risk

Not all output is equal. These four categories are the most common source of expensive late-caught errors:

Verify these more carefully than usual

Output with numbers — calculations, reports, data summariesagents are very good at making numbers look reasonable; that does not mean the numbers are right; verify by recalculating at least one representative figure yourself

Output that others will use as the basis for a decisionif the next reader acts on this without additional verification, your error becomes their error — and then their consequences

Output synthesized from multiple sourcesagents are good at combining information but miss contradictions between sources or flatten important nuances that actually matter

Output in a domain you know less wellwhen you lack the expertise to evaluate the output, you tend to trust what looks professional — this is where silent errors are most dangerous

Professional-looking output is not a signal of correct output. Sometimes it is the opposite — a well-written wrong answer gets through more easily than a badly-written wrong answer.

03Two sentences to close the gap at the source

The cheapest way to reduce the cost of errors is not reviewing output more carefully after you receive it. It is telling the agent upfront what to flag — so it surfaces uncertainty before you have to find it:

Ask it to mark the uncertain parts

"In the output, mark [CHECK] anywhere you are uncertain or where I should verify further. Do not let those sections look the same as the parts you are confident about."

Ask it to name its assumptions

"After the output, briefly list the key assumptions you used to produce it. If an assumption is wrong, in which way would the output be wrong?"

These two sentences do one important thing: they turn silent errors into labeled errors. Output with [CHECK] markers tells you exactly where to read carefully. A list of assumptions tells you when the output is still valid — and when it is not.

The speed an agent delivers is the sticker price. The speed you actually take home is after subtracting the time spent correcting things that were not verified at the right moment. The gap between those two numbers is entirely controllable — and it starts with the question you add before you accept the result.