AI deliverable review: what its assumption list covers, what you read
Only numbers and dates you never supplied can rest on the AI's assumption list; a person checks the rest every time.
Key points
- List-only checks: values you never gave
- Have it show its work on derived values
- Read names and premises, flagged or not
- Empty lists first: suspect hidden gaps
Contents
The AI's list only covers numbers and dates you never gave
A&A perspective
Only where the AI filled in numbers and dates you never supplied can you let its assumption list stand in for your review. An assumption list is a list you have the AI write of every place where it guessed to fill in information missing from your materials.
From the sources
On September 10, 2026, Anthropic published a report on the half-day workshops it ran for small business owners in ten US cities. More than 1,000 owners and operators came, and the report says 80% of registrants run companies with 5 to 50 employees.
From the sources
The worry participants shared was something important falling through the cracks. The report says owners often talked about learning how to check the AI's results instead of taking them at face value.
From the sources
As one result, the report says, the owner of one company changed how they prompt the AI so that it flags when it is assuming and when it knows.
A&A perspective
Automate & Augment (A&A), the company writing this article, reads what that instruction can catch like this. It catches values the AI entered while knowing it had not been given them. On a quote, that means an acceptance period nobody mentioned (the time the client takes to check and accept a deliverable), or a working-day estimate you never supplied.
A&A perspective
Because the AI fills these in knowing something is missing, we expect them to show up on the list.
A&A perspective
Don't let the AI write the list freely. Have it give four things for every entry: which sentence in which section, the value it entered, why it chose that value, and whom to ask what in order to confirm it. With the fourth item, the list doubles as your list of questions for the client.
A&A perspective
Sorting four kinds of content (numbers, proper names, dates and premises such as scope and exclusions) by where they came from gives the table at the top of this section. Neither the report nor the research cited here measured how often each kind gets caught by such a list. The table is A&A's design proposal.
For values built from your files, have it show its work and compare
A&A perspective
Numbers taken from the files you supplied, and the totals and dates calculated from them, need their working shown and checked against those files, whether or not they appear on the list. When the AI takes a value you gave it to mean something else, it has no sense that anything went wrong, so the misreading rarely reaches the list.
From the sources
The same report tells of Tony Severino, who has run his own painting firm near Chicago since 1984. It says he now has the AI read PDF blueprints floor by floor to help him prepare bids.
From the sources
On an early bid, he caught the AI using a floor-area multiplier instead of real wall measurements. Since then, the report says, he asks the AI to show him its work.
A&A perspective
The report does not say whether the AI had been given the real wall measurements at that point. Nor does it say the AI pointed out the substitution itself. If it did have the measurements, A&A reads this as a misreading: the values it needed were there, so the AI would not notice that it had guessed.
A&A perspective
A misread value comes with the right units and a plausible number of digits, so it reads naturally on the page. Reading only what the list flags walks straight past it.
Hypothetical example
Say a solo developer who builds software for clients drafts quotes with AI (a hypothetical example). The effort breakdown carries the supplied figures correctly, but the total was filled in by a different calculation and was off by 1.5 person-days. The quote's assumption list was empty.
A&A perspective
So for numbers and dates built from your files, have the AI state the source of each one: which value on which page of the file, and what formula it used. The reviewer then lays the stated working next to the files and compares them.
A&A perspective
Having the working written out does not remove the comparison; it still has to happen every time. The report does not say this way of asking reduced what was missed, either.
Read names and premises every time, flagged or not
A&A perspective
Proper names that are not in your files, and premises such as scope and exclusions, get read every time, whether the list mentions them or not. Where the AI wrote something it believes it knows, the item rarely lands on the list.
From the sources
In 2022, Anthropic published research on whether language models can judge if their own answers are right and whether they know the answer to a question.
From the sources
According to the research, models predict fairly well whether they know an answer, but those predictions go wrong more often on new kinds of tasks. It also says the predicted probability of knowing rises appropriately when relevant source material is placed in the prompt.
A&A perspective
The research covers 2022 models on question-answering tasks; it did not measure assumption lists on deliverables.
A&A perspective
From this, A&A reads the AI's list as a list of the places it knows it does not know; what it got wrong while believing it knew does not appear. And since the AI is more likely to judge that it knows where you supplied material, we think misreadings there are also less likely to show up on the list.
A&A perspective
Proper names can come from what the AI already knows. A plausible name of an internal policy that does not exist, for instance, can slip in. Check every name that is not in your files against those files or an official source.
A&A perspective
Premises sometimes show up on the list and sometimes do not. Scope and provisos are also easy to skim past even when you read the whole document from the top. So have the AI write one line per chapter stating what that chapter assumes, and read just those lines side by side.
A&A perspective
The list itself is text the AI wrote, with no guarantee that its contents are right. Use the list to add places to read, and decide which places you may skip by the kinds in the earlier table.
Read deliverables with an empty list first
A&A perspective
When you review several deliverables, start with the ones whose list is empty or shorter than usual. An empty list does not mean nothing was filled in.
A&A perspective
An empty list mixes three cases you cannot tell apart: nothing was actually filled in, the AI never looked for gaps, or only misreadings happened.
A&A perspective
The last two are riskier than a deliverable with a long list. A long list at least tells you the AI looked for gaps.
Hypothetical example
Back to the hypothetical developer: say they sent six quotes in two weeks. Four came with lists, all covering an acceptance period nobody had mentioned or a working-day estimate nobody had supplied. Reading the two lists with zero entries first turned up the mismatched total in one of them.
Hypothetical example
For the four quotes with lists, they read only the flagged places and the places they had decided to read every time: where the totals and delivery dates came from, names not in the files, and each chapter's one-line premise.
A&A perspective
Changing the reading order does not change the total time spent reviewing. What changes is that the time goes first to the places that are easiest to miss.
A&A perspective
In day-to-day use, record how many entries each deliverable's list has and compare it with your past average. When a count is unusually low, check whether you gave the AI too much material or too short an instruction.
A&A perspective
Short documents that are quicker to read in full do not need any of this. Short deliverables with no numbers or dates get read in full, as before.
Summary
A&A perspective
Let the AI's list stand in for review only where it filled in numbers and dates you never supplied.
A&A perspective
For numbers and dates built from your files, have the AI show its working and compare it with the files.
A&A perspective
Read proper names not in your files, and premises, every time, whether the list mentions them or not.
A&A perspective
The emptier a deliverable's list, the sooner you read it, suspecting gaps that were filled in.
A&A perspective
Where the line falls between what the AI's list covers and what a person reads moves with what the AI can do today.
Only numbers and dates you never supplied can rest on the AI's assumption list. Have it show its working for numbers and dates built from your files and compare that with the files, and read proper names not in your files and premises every time, whatever the list says. Read deliverables with an empty list first, because an empty list is no proof that nothing was filled in.
FAQ
Can I have the AI put its assumption list inside the deliverable itself?
Have it output the list as a separate file or a separate section. Mixed into the body, it adds the work of deleting it before the client sees the deliverable, and some of it gets left in.
If I settle the facts before drafting, do I still need the list?
Yes. Even with the facts settled, misreadings and filling-in can still happen while the text is being written. The step before writing is covered in “Clashing numbers in AI-made documents: 3 conditions to settle facts first.”
What if I still lack time to review after narrowing what I read?
The places you read every time cannot be skipped, so cut how many deliverables you have the AI produce at once, or how much material you hand it for each one. Having it make the next one only after you finish reading the last keeps review from piling up.
Sources & editorial note
Primary pages read for this article. Publication dates below belong to the sources; access dates record our research.
- What 1,000 small business owners taught us about AI
Anthropic · 2026-09-10
Accessed 2026-10-22 - Language models (mostly) know what they know
Anthropic · 2022-07-11
Accessed 2026-10-22