The Verify-Before-You-Send Rule for AI Output
Why this matters
The cost of an AI tool is not the subscription. It is the one wrong sentence that goes out under your shop's name: a code section that does not exist, a warranty term you never offered, a part number that ships the wrong igniter, a "safe to keep using it" that was not safe. Every one of those read fine when somebody glanced at it. The skill is not reading harder. It is knowing which parts of an output need which kind of check, and refusing to send the parts nobody checked.
This card owns the verification standard for the AI shelf. Other cards point here instead of re-deriving it. The reason invention happens at all is explained in: What a Language Model Is Actually Doing When It Answers.
The rule in one line
Every specific in an AI output is checked against a source at the level its rung demands, before it leaves the building, by a named person. "Specific" means anything that could be wrong in a way a reader would act on: a number, a name, a date, a citation, a promise, an instruction.
Prose is not on the ladder. Tone, word choice and structure you judge by reading, and if you would not say it, you change it. The ladder is for the specifics buried in the prose.
The ladder: four rungs by what it takes to check
Sort each specific by what it costs you to confirm it, not by how important it feels. Importance decides whether it goes out at all; checking cost decides how you confirm it.
| Rung | What sits here | How you check it | Typical time |
|---|---|---|---|
| 1. Readable | Tone, clarity, whether it answers the question asked, whether it says anything you do not mean | Read it as the customer would | Seconds |
| 2. Your own records | Customer name, address, appointment date and window, job number, invoice total, what the tech actually found, your price for a line item | Open the job, invoice or price book and compare line by line | Under a minute per item |
| 3. Outside source | Code sections, part numbers, model specs, torque and clearance values, statute names, manufacturer warranty terms, chemical label directions | The code book or the jurisdiction's adopted edition, the manufacturer document, the label, the agency's own text | Minutes, sometimes a phone call |
| 4. Never on AI alone | Safety instructions to a customer, any promise the shop is bound by (price, timeline, warranty coverage, refund), any legal or medical statement | A human who owns that decision writes or approves it from their own knowledge, then rung 2 or 3 confirms the details | However long the decision takes |
Two things make the ladder work. First, rung 4 is not "check harder", it is "a person decides." The AI may draft the sentence; it may not be the reason the sentence is true. Second, rung 3 means the actual source, not a second AI answer. Asking another chat tool, or the same one in a new window, is checking the model against itself. The two agree often enough to feel like confirmation and are wrong together often enough to matter.
Why each rung sits where it does
Rung 1 is cheap because you are the source. You know what the shop sounds like and what you meant to say. A sentence promising to "prioritize your concerns" is wrong in a way you can see.
Rung 2 is cheap because the truth is already on your screen. The model never saw your job records unless you pasted them in, so anything it says about a specific customer is either copied from your paste or invented. Invented appointment windows are common: ask for a reschedule email and the draft will helpfully offer "Tuesday between 9 and 11" because that is a typical-looking window. Your schedule may not have that slot.
Rung 3 is expensive because the source lives elsewhere. This is where the model is least reliable and the check costs the most, which is the worst combination and the reason most bad AI output that reaches a customer comes from this rung. A code section has the right shape and the wrong digits. A part number has the right format and belongs to nothing. The temptation is to skip the check because the answer looks so specific. Specific-looking is exactly the signature of the risk.
Rung 4 is closed because the cost of being wrong is not yours to choose alone. Telling a homeowner "it is fine to run the furnace until we get there" after they mentioned a smell, or quoting a price the tool computed, or describing what the warranty covers, commits the shop or puts a person at risk. Those sentences get written by whoever owns the decision. The AI can tidy the grammar afterwards.
Worked example: one draft, every specific sorted
Your office has asked a chat assistant for a follow-up email after a water heater visit. The tech's note said: anode badly corroded, tank 11 years old, minor weep at the drain valve, recommended replacement, customer wants a quote. Here is the draft, with each specific numbered.
Hi Ms. Okafor, thanks for having us out on (1) Thursday. (2) Our technician found that your anode rod is heavily corroded and there is a small leak at the drain valve. (3) At 11 years old, your tank is past the typical 8 to 12 year lifespan. (4) Under (5) IPC Section 504.6, a leaking tank must be replaced within 30 days. We recommend replacement and (6) can install as soon as next Monday. (7) Your new heater will carry a 10-year warranty. (8) In the meantime, it is safe to continue using the heater normally. We will send your quote shortly.
Sort them:
- "Thursday" - rung 2. Open the job. The visit was Wednesday. The model guessed. Fixed in ten seconds.
- The findings - rung 2. Compare to the tech's note. "Small leak at the drain valve" overstates "minor weep at the drain valve" slightly; a weep at a valve is not a tank leak and the customer will read it as one. Reword to match the note.
- "8 to 12 year lifespan" - rung 3, and soft. It is a commonly quoted range for tank heaters and varies with water quality and type. It is defensible as "typical" but it adds nothing the customer needs. Cut it, which is cheaper than checking it.
- and 5. "Under IPC Section 504.6, a leaking tank must be replaced within 30 days" - rung 3, and it fails twice. Section 504 of the International Plumbing Code does deal with water heater safety devices, so the prefix looks right, but no model plumbing code orders a homeowner to replace a weeping tank within 30 days. The citation is invented and the rule is invented. Even if a local rule existed, it would come from the jurisdiction's adopted code, not the model code. Delete the sentence. See related: The AI Quoted a Code Section That Does Not Exist.
- "Next Monday" - rung 4 dressed as rung 2. It is a scheduling promise. The office manager checks the board, and if Monday is not open, it does not go out.
- "10-year warranty" - rung 4. The warranty depends on the unit you quote, and manufacturer terms vary by model and by whether it is registered. Nobody has picked the unit yet. Delete it; the quote will state the real terms.
- "Safe to continue using the heater normally" - rung 4, and the most dangerous line in the draft. A weep at the drain valve is usually not an emergency, but the model has no idea whether the tech also checked the relief valve, the venting, or the gas connection. Only the tech who saw the unit can say what the customer should do in the meantime. Replace it with what the tech actually advised, or call the tech.
Count the result. Of the eight flagged items, two were cheap fixes from your own records (1 and 2), one was cut rather than checked (3), one was an invented citation carrying an invented rule (4 and 5, deleted as one sentence), and three were rung-4 items the AI had no business deciding (6, 7, 8). Total checking time was about four minutes. The email that went out is shorter and every sentence in it is true.
Now notice what would have happened with a skim. The draft reads warm and competent. A tired office manager at 4:40 pm sends it. The customer forwards the code sentence to their brother-in-law, who is a plumber, and your credibility on the replacement quote is gone before you send it.
What changes the rung
The rung belongs to the specific and to where it came from. Two conditions move an item down a rung:
- You pasted the source in. If the manufacturer's spec sheet was in the prompt, a value copied from it drops from rung 3 toward rung 2: you still check it against the sheet, but the sheet is right there and the check is a glance. It does not drop to rung 1, because the model can copy the wrong row.
- The tool is grounded on your own records. Some field software features draft messages from the actual job record. Then the appointment date is probably right, but you still confirm it, because a draft built from the wrong job looks identical to one built from the right job.
One condition moves everything up: the reader will act on it without asking you. A text to a customer at 9 pm, a reply posted publicly, an instruction a new tech will follow on a roof. Where nobody gets a second chance to catch it, treat rung 2 items as rung 3 and rung 3 as rung 4.
Who does the checking
The person who sends it checks it. That sounds obvious and is the part shops skip, because the person who asked the AI for the draft assumes the person who sends it checked, and vice versa. Name it in the shop policy: the sender owns every specific in the message. See related: The Shop AI Use Policy SOP.
For rung 4, the sender is not enough unless the sender owns the decision. An office coordinator can send a scheduling promise they made from the board. They cannot send a safety instruction or a warranty description; those go to the tech or the owner first.
The failure modes this ladder exists to catch
These are the specific ways shops get burned that the table above does not already name:
- Checking the prose instead of the specifics. Reviewers fix the awkward sentence and sail past the number in it. Read the specifics first, in a separate pass, before you read for tone.
- The cut you did not make. Most failed specifics did not need to be in the message at all. The fastest verification is deletion; a reply that says less is easier to keep true.
- Verification drift. A prompt that produced clean output for three months starts producing a new kind of specific after the tool updates. The ladder applies every time, not until the tool has earned trust. Tools change underneath you; see related: When to Turn an AI Feature Off.
How to know the rule is working
Pull five AI-assisted messages your office sent last month. For each, list every specific and mark the rung and whether there is evidence it was checked (a matching job record, a manual page, a note). If any rung-3 item has no source and any rung-4 item has no named human owner, the rule is on paper but not in practice. The target is zero unchecked rung-3 and rung-4 items, and it is achievable because there are usually only one or two per message.
References
- See related: What a Language Model Is Actually Doing When It Answers
- See related: The Shop AI Use Policy SOP
- See related: The AI Quoted a Code Section That Does Not Exist
- See related: The AI-Drafted Customer Message Review SOP
- International Code Council, International Plumbing Code (check the edition your jurisdiction has adopted, which binds work there through the permit)