Ask AI to expose the assumptions in a plan
The launch plan reads well and everyone has nodded at it. Somewhere inside it is a sentence starting "Assume…" that nobody has checked, and a few more that nobody wrote down at all.
What you'll make
a short assumptions log for a plan, quoting the plan’s own words, with each item labelled and a question that would settle it.
What you'll need
a plan you’re allowed to share with a work AI chat (ChatGPT, Claude, Copilot or Gemini), and ten minutes. Practise with the fictional plan in the worked example first.
Example result.
| Plan wording | Status | Question that settles it |
|---|---|---|
| “Assume our supplier can ship 500 boxes by 26 October” | Stated assumption, not checked | Has the supplier confirmed 500 by 26 October in writing? |
| “Price at £24, the same margin as last year” | Depends on figures not in the plan | What are this year’s unit costs? |
| (not in the plan) Last year’s photos show this year’s box | AI inference | Does the box look the same, and can we reuse the photos? |
1. Ask for the log
Paste the plan into a work chat and ask:
List the assumptions and dependencies in this plan. For each one, quote the plan's exact words, or write "not in the plan" if it is your own inference.
Label each as: stated fact, stated assumption, or your inference. Say what happens if it is wrong, and give one short question that would settle it.
Do not invent dates, owners, figures or approvals. Anything without evidence in the plan is unconfirmed.Why this works: quoting the plan makes every row checkable, and the three labels stop a sentence that begins “Assume…” being quietly upgraded to a fact.
What you should see: a list or table where every row either quotes the plan or says “not in the plan”.
2. Ask the smallest question that settles each one
“Can you confirm everything?” gets a nod. “Has the supplier confirmed 500 boxes by 26 October in writing?” gets an answer. Turn the unconfirmed rows into actions:
Turn the unconfirmed items into the smallest useful next actions. For each, say who would know and what evidence would settle it.
Where the plan names no owner, write "owner to agree" rather than picking one. Put the items that would change the launch date or price first.Why this works: a narrow question can be answered by one person in one message, which is the difference between a log and a worry list.
What you should see: three to six questions, each with a “who would know”, in order of how much they matter.
3. Check the log against the plan
Go down the log with the plan open. For each row: is the quote really in the plan? Is the label right? Has anything the plan calls an assumption been labelled a fact? Has the AI added an inference without saying so?
Why this works: the log is only useful if its labels are right. One mislabelled row can send the whole plan forward on something nobody checked.
What you should see: every quote found in the plan, and every “Assume…” sentence still labelled as an assumption.
Check the result
- Every quote appears word for word in the plan.
- Every sentence the plan starts with “Assume” is labelled a stated assumption.
- Inferences are marked “not in the plan”.
- No row has an owner, date or figure the plan doesn’t give.
The usual failure: an assumption is promoted to a fact. The plan says “Assume our supplier can ship 500 boxes”, and the log says “Fact: supplier capacity is 500 boxes”. It reads more confident and is less true. Spot it by searching the plan for the word “assume” and checking each match against its label.
The log labels "[item]" as [wrong label]. The plan says "[wording]", which is a stated assumption, not a confirmed fact. Correct the label and its question, and check the other rows for the same mistake. Do not add a confirmation that isn't in the plan.[item] is the row, [wrong label] is what the log called it, and [wording] is the plan’s exact words.
Worked example: the autumn gift box
Fictional example. A small tea company’s launch plan:
Hartwell Tea autumn gift box: launch plan
Launch on Monday 2 November 2026.
Assume our supplier can ship 500 boxes by 26 October.
Price at £24, the same margin as last year's box.
Ali to draft the launch email; send date to be agreed.
Product page reuses last year's photos.The first log had four rows. Row one read “Fact: supplier can ship 500 boxes by 26 October.”
- The assumption had become a fact. The plan says “Assume”. The repair prompt relabelled it “stated assumption, not checked”, with the question in the example result at the top of this page.
- The margin row was right to flag a gap. “The same margin as last year” depends on last year’s margin and this year’s costs, and neither is in the plan. The question asks for the costs, not a guess at them.
- The useful inference was labelled. The AI noticed that reusing last year’s photos assumes the box looks the same. That isn’t in the plan, so it’s marked “AI inference”, which is exactly where an inference belongs. Ali’s send date stayed “to be agreed”.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Data files hide assumptions too
Plans that depend on a spreadsheet often assume things about it: that IDs are unique, which currency an amount is in, what a blank means. Ask for those as rows in the same log, then check them against the file itself, not the plan. Field meanings belong in a data dictionary.
Keep the log current
When an answer comes back, update that row only: change the status, quote the evidence and note the date. Keep the old status in the history, so it’s clear why the plan changed. An AI rewrite of the plan isn’t a confirmation record.