← All posts

AI sorted a vendor pitch. Two sales lines passed as fact.

October 9, 2026 · View on Instagram · View on Facebook

The plain request added its own number. Before sorting, we asked for a summary and a buy decision. GPT-5.5 added its own $20/hour manager rate, which the pitch never gives, though it did flag the payroll claim as possibly just a CSV upload. — The plain request added its own number
Real output · GPT-5.5 (OpenAI API) · Oct 2026
If managers save 5 hours per week total across the business, that is 260 hours per year. Even at $20/hour of manager time… … They say it “works with every major payroll system,” but the specific feature listed is a CSV export.
Before sorting, we asked for a summary and a buy decision. GPT-5.5 added its own $20/hour manager rate, which the pitch never gives, though it did flag the payroll claim as possibly just a CSV upload.

Three lists, one quote each. Facts are things you could check right now. Assumptions only hold if events go the vendor's way. Recommendations are what the vendor wants you to do. — Three lists, one quote each
Facts are things you could check right now. Assumptions only hold if events go the vendor's way. Recommendations are what the vendor wants you to do.

The sample pitch we sorted. Our fictional pitch from RotaLoop to a 3-location, 42-staff cafe group claims 5 hours a week saved and a 12% overtime drop. It costs $6 per staff member per month and says to sign before October 31. — The sample pitch we sorted
Our fictional pitch from RotaLoop to a 3-location, 42-staff cafe group claims 5 hours a week saved and a 12% overtime drop. It costs $6 per staff member per month and says to sign before October 31.
Source: Fictional sample pitch written for this test

The prompt that sorts it. Paste this above the pitch. It defines each list, asks for the exact line behind every item, and sends unsourced results to assumptions. — The prompt that sorts it
Sort every claim in the vendor pitch below into three lists: FACTS: things stated as true that I could check right now, like prices, terms or features. ASSUMPTIONS: things that only hold if events go the way the vendor expects, including results, savings and predictions. RECOMMENDATIONS: what the vendor wants me to do. For each item, quote the exact line it came from. If a result or number has no source in the pitch, it goes under ASSUMPTIONS, not FACTS. Do not add anything the pitch does not say.
Tested in GPT-5.5 (OpenAI API), Oct 2026
Paste this above the pitch. It defines each list, asks for the exact line behind every item, and sends unsourced results to assumptions.
**Try it:** Run it on the last proposal in your inbox.

All 19 lines sorted, every quote exact. In our test, the savings and overtime claims landed under assumptions, and the October 31 deadline under recommendations. We checked each quote against the pitch by hand. — All 19 lines sorted, every quote exact
Real output · GPT-5.5 (OpenAI API) · Oct 2026
## ASSUMPTIONS … - “Managers using RotaLoop save an average of 5 hours a week on scheduling.” - “Overall, cafes like yours see overtime costs fall by 12% in the first quarter.” … - “Setup takes one afternoon.” … ## RECOMMENDATIONS … - “Sign before October 31 to lock in the current price.”
In our test, the savings and overtime claims landed under assumptions, and the October 31 deadline under recommendations. We checked each quote against the pitch by hand.

The quote was exact. The label wasn't.. The sorting prompt filed “works with every major payroll system” and “built for busy hospitality teams” under FACTS. The pitch's only payroll feature is a CSV export. A second run with a check per fact kept both. — The quote was exact. The label wasn't.
Real output · GPT-5.5 (OpenAI API) · Oct 2026
## FACTS … - “RotaLoop is the scheduling app built for busy hospitality teams.” … - “RotaLoop works with every major payroll system.”
The sorting prompt filed “works with every major payroll system” and “built for busy hospitality teams” under FACTS. The pitch's only payroll feature is a CSV export. A second run with a check per fact kept both.

Treat the facts list as a to-check list. Look hard at any 'every', 'all' or 'built for'. Before pasting a real pitch, remove names, account numbers and confidential details. Also check your employer's approved tools before uploading real documents. — Treat the facts list as a to-check list
Look hard at any 'every', 'all' or 'built for'. Before pasting a real pitch, remove names, account numbers and confidential details.
Also check your employer's approved tools before uploading real documents.
**Try it:** Add: for each FACT, write how I would check it. It won't fix labels.

Sources and assumptions

Anthropic, Reduce hallucinations (Claude Platform Docs): Anthropic recommends having the model cite quotes for each claim so the response is auditable ('Make Claude's response auditable by having it cite quotes and sources for each of its claims'), recommends word-for-word quote extraction specifically 'For tasks involving long documents (>20k tokens)', and notes 'while these techniques significantly reduce hallucinations, they don't eliminate them entirely. Always validate critical information.' The post applies the citing principle to a short document, and the fact/assumption/recommendation split is the post's own technique. (checked 2026-10-08)

Assumptions:

The vendor pitch, the company 'RotaLoop', 'Harbor Street Cafes' and every number in them are fictional sample material written for this test. No real product or company is meant.
Only GPT-5.5 via the OpenAI API was run successfully (tests before-gpt, after-gpt, after-v2-gpt, Oct 2026). The Claude runs (tests 'before' and 'after') failed in this research session with no output, so the slides must not name Claude as tested; prompt_tools = GPT-5.5 (OpenAI API), Oct 2026.
Each test was one run. Another run or another model may sort the lines differently.
My hand check: the pitch has 19 claim lines (not counting headings and the title). after-gpt sorted all 19 (10 facts, 6 assumptions, 3 recommendations), every quote matches the pitch word for word, and nothing was added.
Mislabels found by hand: 'RotaLoop works with every major payroll system.' (a broad sales claim that can't be checked from the pitch, which only lists a CSV export) and 'RotaLoop is the scheduling app built for busy hospitality teams.' (marketing positioning) were both listed as FACTS in after-gpt and after-v2-gpt. 'Scheduling shouldn’t eat your week.' is a slogan, listed under ASSUMPTIONS (minor).
In the 'before' output, GPT-5.5 brought in its own '$20/hour' manager rate to value the savings, read the 5 hours as '5 hours per week total across the business' (the pitch says 'Managers using RotaLoop save an average of 5 hours a week'), and listed '99.9% uptime over the past year' without the 'Claimed' label it gave the other numbers. Its arithmetic is correct: 42 × $6 × 12 = $3,024; 42 × $8 × 12 = $4,032; difference $1,008.
To be fair to the 'before' run: it flagged on its own that 'works with every major payroll system' might mean only a CSV upload. The post's point is that the sorted version is auditable line by line, not that the summary missed everything.

The short version

Sort the pitch into facts, assumptions and recommendations.
Make the AI quote every line.
A quote proves the line exists, not that it's true.
Check every fact before you sign.

Follow @todaysaitip Save this post on Instagram

AI sorted a vendor pitch. Two sales lines passed as fact.

The plain request added its own number

Three lists, one quote each

The sample pitch we sorted

The prompt that sorts it

All 19 lines sorted, every quote exact

The quote was exact. The label wasn't.

Treat the facts list as a to-check list

Sources and assumptions

The short version

More posts

Paste [CLIENT], not your client's name. Then make AI keep it.

GPT-5.5 put my private note into a customer email. Twice.

Tell ChatGPT about your job once. Then stop retyping it.