Trust — What You Check Before It Ships

The Human Check: Which Claims You Verify Before Anything Goes Out

AI writes confident sentences about things it cannot know. Here is the short list of claim types that always get your eyes, and how to make checking fast.

Confidence is not accuracy

The single most important thing to understand about these tools: the output sounds equally certain whether it is right or invented. There is no wobble in the voice when it guesses. A fabricated licence number is formatted exactly like a real one.

So you cannot rely on noticing that something looks wrong. You need a standing list of claim types that always get checked, regardless of how right they look.

The always-check list

Any of these appearing in generated text stops it from going out until you have verified it against a real source:

Claim typeWhere the truth livesCost of getting it wrong
A price or totalYour rate sheet + your arithmeticYou are bound to it, or you lose the job
Licence / registration numbersYour actual licence documentMisrepresentation; regulatory exposure
Insurance types and limitsYour current certificateA claim you are not covered for
Warranty or guarantee termsYour written termsA promise you must honour
Turnaround or arrival timesYour actual scheduleThe complaint in the review lesson
Certifications and set-aside statusYour registration recordsOn a federal bid, this is serious
Years in business, staff count, job countsRealitySmall lie, large credibility cost
Any claim about a competitorDon'tLegal exposure, no upside
"Removes 99% of…" style efficacy claimsA source you can citeAdvertising standards

Make the check fast

The list is long enough that checking everything word by word would be slow. Two habits make it quick.

First: ask for the claims to be marked. Add this to any prompt producing customer-facing text:

> After the draft, list every factual claim it contains that I should verify before sending — prices, numbers, dates, credentials, guarantees, and any efficacy claim. If the draft contains none, say so.

Now you are checking a short list instead of re-reading prose.

Second: forbid invention up front.

> If you don't have a fact you need, write [CONFIRM: what's needed] rather than filling it in. Do not estimate, approximate, or use a placeholder that looks like a real value.

The bracket convention matters. A visible [CONFIRM: licence number] cannot be sent by accident. An invented number that looks plausible absolutely can.

Watch for the confident middle

The genuinely dangerous outputs are not obviously wrong. They are plausible. A model that has read a great deal about exterior cleaning will produce sentences like:

> Our soft-wash process removes 99.9% of mould spores and is safe for all siding types.

Both halves are the problem. The percentage is invented. "All siding types" is a claim you would not make out loud — you would ask what the siding is first.

Nothing in the sentence *looks* wrong. That is exactly why it needs the list rather than your instincts.

When you are the last check

On the business-running side, there is no review step between you and the customer. You are it. That is the reason this list exists, and the reason the two-tool split from the first lesson matters: system changes get a review gate; emails do not.

So build one habit and keep it: nothing customer-facing goes out on its first draft without one deliberate pass against the always-check list. Not a re-read for tone — a specific hunt for claims.

It takes ninety seconds. It is the cheapest ninety seconds in this entire track.