# The Pre-Flight Trust Check

> A free five-question decision tree for any AI output before you act on it or send it: accept with a check, add what it missed, or redo the key part.

Use this at the moment an AI output is about to leave your hands or change what someone does: a risk assessment for steerco, a status update, a scope recommendation. Five questions, in order, send it to one of three modes. Most routine outputs exit at the first question with a quick check. The ones that matter get the source check and the question about what the tool couldn't have known.

## The frame

```
Output: [what it is]
Going to: [who sees it or acts on it]

1. Will anyone outside the team see it, or act on it?
   No: Accept with a targeted check. Stop here.
   Yes: go to 2.
2. Does every warning, figure and recommendation name a source you gave the tool?
   No: ask the tool for the source of each. Anything still unsourced is marked "unverified" or removed. Then go to 3.
   Yes: go to 3.
3. Have you opened the sources for the claims that carry weight?
   No: open them now. Go to 4 when done.
   Yes: go to 4.
4. Is the situation new, or are the stakes high? (a new sponsor, a restructure, an integration the tool hasn't seen, steerco or sponsor audience, an estimate above your threshold)
   Yes: Compare or redo. Produce the key section yourself or cross-check it another way, and use the AI output as a sense-check.
   No: go to 5.
5. What couldn't the tool have known? Write one line.
   It changes the output: Augment or override. Add what's missing before it goes.
   It changes nothing: Accept with a targeted check.

Mode: [Accept with a targeted check / Augment or override / Compare or redo]
Unverified items left in: [list, or "none"]
```

## Have an AI list the claims and their sources

Question 2 is faster with this prompt, run in the same conversation that produced the output.

```
List every warning, figure, date and recommendation in the output above. For each one, give the exact source in the material I provided: the document name and the line, row or section. If a claim has no source in my material, write "NO SOURCE" and do not supply one.
```

"NO SOURCE" is a valid answer. A tool that cannot find a number may supply one anyway, so treat any source you cannot open as missing.

## Worked examples

The entries below are illustrative.

### Risk assessment for steering committee

```
Output: Top five delivery risks for the phase two gate
Going to: Steering committee pack

1. Outside the team? Yes.
2. Sources named? Four of five. The fifth, "vendor has missed two milestones on similar programmes", had no source. Asked the tool: NO SOURCE. Removed.
3. Sources opened? Yes, for the two risks rated high.
4. New or high stakes? Yes: steerco audience and a new sponsor since last month.
   Compare or redo. Rewrote the top two risks myself from the RAID log and used the AI list to check nothing was missed.
Mode: Compare or redo
Unverified items left in: none
```

### Weekly update for a stable project

```
Output: Weekly status update, week 23
Going to: Project team and the sponsor's office

1. Outside the team? Yes, the sponsor's office.
2. Sources named? Yes, every figure traces to the tracker export.
3. Sources opened? Yes, the two dates that moved.
4. New or high stakes? No. Routine update, no change in sponsor or scope.
5. What couldn't the tool have known? The sponsor asked in the corridor about the training dates. Added one line on them.
Mode: Augment or override
Unverified items left in: none
```

## Notes on the modes

**Accept with a targeted check** still means a check. Look for what the tool couldn't have known.

**Augment or override** is for an output that is technically right but misses the subtext. A clean document shouldn't become the record of a more complicated reality.

**Compare or redo** keeps the AI output as a second opinion. Stakes and novelty set the mode. How confident the output sounds tells you little, because a capable model is wrong in the same fluent voice it is right in.

## When it does not apply

Drafts for your own thinking, which nobody else sees and nothing depends on, don't need the check. Start it the moment an output is going to someone else or changing what gets done.

## What this feeds

This is the human review step for tasks marked "AI drafts, person approves" in the [AI Delegation Register](/toolkit/delegation-register/). For agents that run without a person reading each output, the [Agent QA Checklist](/toolkit/agent-qa-checklist/) does the same job before go-live. The reasoning behind the three modes is in [You Approved It, You Own It](/articles/ai-output-quality/).

---

From ProjectorPM: https://projectorpm.xyz/toolkit/pre-flight-trust-check/