AI in careful jobs: write the gate before you open the tool

After this page you can write a gate list for your job and use it to stop an unsupported sentence, using the lab's healthcare and lending boundaries.

Start with the lab's question: what must be true before anything the tool touched reaches a person? Write a gate, a check held by someone other than the drafter that runs before work goes out and can stop it, using the healthcare and lending setups below.

Open the “What stays with the owner” sections in healthcare and lending. Use “what stays with the owner”, the decisions the tool cannot make, as the first line of your list.

Apply the healthcare safety boundary to your job: keep diagnosis and dosing with the clinician. Apply the lending sign-off boundary to decisions on a person's money: keep those with the responsible professional. Require a person's sign-off before anything goes to a patient or borrower, following the lending hold on money-page changes.

You know it worked when your list names who can approve patient-facing or borrower-facing work. Use healthcare's practitioner sign-off on safety changes and Nate's sign-off on lending money pages as your check.

Put the record, the written source your reviewer can check, in front of the tool. Check whether that source is complete: in the 250-reachable audience account, the tool read only the ecommerce platform and missed consent held in the legacy system. Name the actual sources on your list so your reviewer can catch the same gap.

Use a fictional protocol for today's exercise, following the healthcare demo's no-real-records boundary; leave remembered details out of the prompt.

You know it worked when you can show which sources belong in the record and what you checked for missing information, the question that exposed an audience of roughly 250 shown as 4.

Apply the healthcare safety boundary to writing: let the tool produce drafts, summaries or checklists from the protocol. Leave diagnosis and dosing outside that task, with the clinician who owns the call. Carry lending's citation gate into your drafting instruction:

Draft an appointment-protocol letter from this record. Point to the source for each instruction. Flag anything the record does not answer. Add no medication advice.

You know it worked when the draft points back to the protocol and leaves any clinical question for the clinician, carrying healthcare's safety sign-off into the letter.

Compare the draft with its record before anyone sends it, using lending's citation gate as your check. Give that check to someone who did not draft it, following the account where a rival AI signs off first. For high-stakes output, add an out-of-family reviewer, an AI from a different model family than the drafter, using only material approved for that reviewer.

Call this review explicitly, as the lab does on demand for high-stakes changes; do not assume every lab change gets an automatic outside review. Keep the responsible person's sign-off after the AI review, as lending's money-page rule requires. Require evidence for the reviewer's findings too; the rival-AI account includes false alarms overturned by clean build runs.

You know it worked when the independent reviewer can stop the send and show why. Reopen any “done” claim whose checks cannot be reproduced, following the false-done account: a blank page was called complete and an unwired safety check was called armed.

Keep a receipt, saved evidence that lets someone verify the work. Copy the link between a change and its checks from Under the hood's receipts into your own file.

Save the draft beside the exact protocol version so you can repeat lending's source check. Attach the review findings and the sign-off that cleared it, carrying lending's money-page approval into the receipt. Keep rejected wording with its source gap, as the fictional clinic receipt below does.

You know it worked when another reviewer can open that file and repeat the check against the approved draft. Use the lab's receipt-forgery account to test it: a signature alone did not prove the claim.

Fictional exercise using the healthcare boundary and lending's source gate.

I'm using an invented clinic here to show the lab's gate on paper. The office manager gives the tool a written protocol about a changed appointment process, with no patient information. The tool drafts a patient-facing letter and adds a reassurance about a medication. The protocol says nothing about medication.

The colleague assigned to the gate compares the letter with the protocol and holds it: that sentence has no source. The manager removes it and routes the medication question to the clinician. The colleague checks the revised letter before signing off.

The manager saves both versions with the review. The receipt names the colleague who signed and records the medication question's handoff. Apply the lab's false-done check here: “drafted” cannot mean “cleared to send” while that unsupported sentence remains.

Keep this example separate from the actual DocAI staging demo: one practitioner is testing it without patient data, and it has no paying customer or signed contract.

• Checking only one source: question the missing records that made 250 reachable people look like 4.
• Letting the drafter close the job: hold the send for an independent verdict, as the rival-AI review held a release.
• Calling a checker live because it exists: trace a real draft through it, as the disconnected lie-detector account required.
• Treating a signature as proof of the claim: check the result too, as the receipt-forgery tests did.

Here's the part you can copy from the healthcare and lending boundaries. Write these five lines for your own job before you open a tool:

1. What stays with me: ___; nothing reaches a person without ___ signing.
2. What record the tool gets: ___; checked for missing sources by ___.
3. What it may draft: ___; claims outside that record stay out.
4. Who checks: ___; this person did not draft it and can stop the send.
5. Where the receipt lives: ___; save the draft and check results with the sign-off.

Test your list on the invented medication sentence above, using the missing-source check and the false-done rule. Hold the send if your list cannot say who rejects that sentence and where the rejection is saved.