Compare two AI summaries of the same call and see what each one missed

Example result.

Locations: A "six" – contradicted. B "six" – contradicted.
  Source 08:20: "five locations, not six".
Figures: A drops "if finance signs off" – overstated.
  B keeps the condition – supported, but omits Priya.
Missing from both: supplier approval still outstanding, no date.

1. Start with a meeting you can check

Pick an ordinary, low-risk call with a decision, an action with a condition, and something explicitly ruled out. A status update where nothing happened produces two bland summaries that agree, and you learn nothing. Avoid anything about a person’s performance, pay, health or conduct.

Keep both summaries unedited and note how each was made. Compare like with like: Google Meet’s notes document, for instance, has a Quick notes tab with the main takeaways and a Full notes tab with the summary, decisions, action items and details. Comparing a quick recap with full notes measures length, not quality.

Then read the record yourself and write a short answer key: decisions, actions, owners, dates, conditions and explicit deferrals, each with a timestamp. Mark gaps and unclear speakers.

Why this works: the key is your fixed point. Without it, you’ll treat whichever summary you read first as the baseline.

What you should see: a five-to-ten-line key with timestamps, written before you read either summary closely.

2. Put the claims side by side

In a new chat, paste both summaries labelled A and B, without your key:

Compare these two summaries of the same meeting. Use only their text.
Do not merge them or decide which is correct.

Break them into discrete claims. For each topic, give:
- exact relevant wording in A, or absent;
- exact relevant wording in B, or absent;
- SAME, DIFFERENT or ONLY IN ONE;
- differences in decision status, owner, date, number or condition.

SAME means all material details agree, not just that the topic appears
in both. Keep conditions with the actions they qualify. A rejection
and an open question about the same option are different claims.

SUMMARY A: [text]
SUMMARY B: [text]

Each [text] is one summary, pasted unchanged.

Why this works: splitting sentences matters more than it sounds. “Priya will send the figures on Friday if finance approve” is three claims, and two summaries can agree on the first, differ on the second and lose the third entirely. Forbidding a merge keeps the disagreements visible; asked to compare two documents, a model’s instinct is to write a better third one.

What you should see: a row per claim, with exact wording from each summary and the label.

3. Check every material claim, including the shared ones

Use your key and the record to give each consequential claim an outcome:

• Supported: the record backs the wording, conditions included.
• Overstated: a suggestion, estimate or condition became something firmer.
• Contradicted: the record says something incompatible, at a timestamp you can name.
• Not found: you couldn’t locate it. That isn’t proof nobody said it.
• Unverifiable: the record is missing or unclear at that point.

Don’t skip a date, figure or commitment because both summaries agree on it.

Why this works: each label leads to a different action. “Contradicted” means correct it; “not found” means ask someone.

What you should see: an outcome and a timestamp beside every material claim.

4. Look for what neither summary captured

Now go the other way: from the record to the summaries. The classic gap is a negative decision. “We’re not doing the integration this year” is a real decision with consequences, and summarisers routinely drop it because nothing happened.

Compare this record with BOTH summaries. Find explicit decisions,
deferrals, conditions, refusals or open questions missing from BOTH.
Give the timestamp and exact wording. If an item appears in either
summary, do not list it. Flag gaps and uncertain speakers.

RECORD: [transcript]
SUMMARY A: [text]
SUMMARY B: [text]

[transcript] is the meeting record you’re allowed to use.

Why this works: the earlier pass asked whether included claims were true; this one asks what was lost. You need both directions.

What you should see: a short list of omissions, each with a timestamp you then check yourself.

Check the result

• Every material claim from both summaries has an outcome and a timestamp.
• Claims both summaries share were checked too.
• Rejections and deferrals in the record appear somewhere in your result.
• Your corrected summary comes from the record, not from combining A and B.

The usual failure: treating agreement as proof. Both summaries say the same thing, so it gets marked supported without a look. Spot it by checking that each “SAME” row has a timestamp from the record, not only a match between A and B. A merged version can keep the mistakes from both with excellent formatting.

Recheck the row for [topic]. A says [exact wording] and B says [exact
wording]. The record at [timestamp] says something different. Show the
difference without resolving it, and do not describe this row as
agreement.

[topic] is the row; the wording comes from each summary; [timestamp] is where the record differs.

Fictional example. A supplier call on 6 October 2026:

TRANSCRIPT EXCERPTS
02:10 Priya: I can send the figures by Friday 9 October if finance
signs off the revised rates. Otherwise, the middle of next week.
02:35 Omar: The supplier also needs internal approval before those
rates are confirmed. We don't have a date for that approval.
05:00 Hana: We are not doing the integration before the end of 2026.
The team does not have capacity.
08:20 Priya: The proposal covers five locations, not six.

SUMMARY A
- Priya will send the revised figures by 9 October.
- The integration will proceed in Q4 2026.
- The proposal covers six locations.

SUMMARY B
- Revised figures may be sent by 9 October, pending finance approval.
- Integration timing remains to be confirmed.
- The proposal covers six locations.

The first comparison marked locations “SAME” and moved on. Checking that row against 08:20 found the correction, and the repair prompt relabelled it. The checked result:

• Figures: A is overstated (condition dropped). B keeps the condition but omits Priya and the fallback timing.
• Integration: A is contradicted. B softens a clear “no” into an open question; no later start was agreed.
• Locations: both contradicted. The record says five, explicitly correcting six.
• Supplier approval: missing from both. It is outstanding, with no date.

What mattered:

1. The shared error. Both said six. Only the timestamp check caught it.
2. The negative decision. Hana’s “not before the end of 2026” vanished from A and blurred in B, because nothing was going to happen.
3. The hidden dependency. Omar’s supplier approval sits underneath the finance condition, and neither summary had it, because nobody agreed to do anything.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

The record has a gap at [timestamps]. Mark any claim that depends on
that passage as unverifiable. For each, draft one short, neutral
question I can send to the participants to confirm what was said.

Meeting type:            Tools compared:          Date:
Differences found (owner / date / figure / condition):
Claims the record contradicted:
Items in neither summary:
Check by hand every time, on this kind of meeting:
What this exercise could not establish:

Sources

• Take notes for me in Google Meet
• Does ChatGPT tell the truth? (OpenAI)