The question

My team is pasting AI answers into decks. Which of them would survive the CFO’s next question?

Most generated answers read well. Fluent prose isn’t the test. The test is whether the number describes the right people, over the right dates, adds up, admits what it couldn’t see, and doesn’t claim more than the evidence supports.

Where it sits on the loop

Wherever an answer gets repeated: mostly where Pipeline turns into Won(the commit call and deal reviews) and in the board meeting. The same checks apply to a number someone built in a spreadsheet.

How to think about it

Six checks, in order, before an answer goes into a deck or a meeting.

Whose data? Is it the right organisation, team and scope? The quiet failure is a team view standing in for the whole company: the number looks right but describes the wrong people.
Which window? Are both dates shown, and are they the dates you asked for? “Last week” is sometimes whenever someone last saved a snapshot.
Does it add up? Do the parts sum to the headline? If they sum to minus two and the headline says minus three, it hasn’t been explained yet.
What can’t it see? Missing data should look missing. A zero reads like you measured something bad.
Is it earning its causal words? “Consistent with” is usually supported; “caused” usually isn’t. This matters most when the answer is about a person.
Did it change anything? Read-only by default, with any change waiting for your approval.

Common mistakes

Trusting fluent prose. A confident paragraph with no dates and no scope is still unverified.
Checking the number but not the window or the scope. The arithmetic can be right and the answer still describe the wrong team or the wrong week.
Reading a blank as a zero. “No activity” and “no data about activity” are different statements.

In Summit53

Each check has somewhere to look, both on the Revenue Engine and in answers from the Summit53 skills in Claude or Codex.

1. Whose data?

In the app: the masthead shows the widest view you’re allowed (organisation, team or your own book) and doesn’t label which; adding ?persona= changes the brief and cards, not the masthead. The first constraint is the same for every organisation-wide view. See the masthead guide.
In the skills: every skill confirms which Summit53 organisation it’s connected to before reading anything, and stops if it can’t. A forecast brief built on a team view says so beside the levels.

2. Which window?

In the app: the first-constraint card reads First constraint · as of <date>, the date of its evidence, while the list under it is dated today. The ledger’s WHEN column shows when you acted and how long until the result, for example acted 3d ago · 27d to call it.
In the skills: the forecast brief always shows the exact two dates it compares, and won’t quietly swap a weekly window for a daily or monthly one.

3. Does it add up?

In the app: WHY DID CONFIDENCE MOVE? ends with Each driver’s share adds up to the total move. Drivers we can’t measure yet are shown as still building. Forecast gap diagnosis splits the gap into RAMP CAPACITY, PRODUCTIVITY and RESIDUAL.
In the skills: the forecast brief’s chart sums to the headline, or shows the part no measured driver explains as a remainder.

4. What can’t it see?

In the app: Not enough data yet when a stage is under its sample floor; an estimated tag on dollar figures resting on a typical value; — for stage counts that haven’t loaded; Pending tiles; building… drivers; and an Excluded from totals warning for amounts with no exchange rate.
In the skills: missing evidence is reported as unavailable or not yet measured, never as zero, and a source it couldn’t read is listed rather than treated as healthy.

5. Is it earning its causal words?

In the app: the Loop says Based on what changed after each action — an association, not proof of cause. and Correlational — the metric moved after we acted; we don’t claim we caused it. Fix playbook results add This is what changed after the run, not proof the run made it happen. Other things moved too. Blueprint checks are shown as LIKELY CAUSES, not as the cause.
In the skills: deals are named as consistent with a movement, not its cause. The competitor response card keeps Verified evidence, with its sources, separate from Suggested response language.

6. Did it change anything?

In the app: opening Diagnostics, the Field guide or a Fix playbook preview changes nothing. Start tracking records a starting point and nothing else. Starting a Fix playbook is a deliberate step for a leader.
In the skills: skills don’t change CRM records. Most only read; a few save something along the way, such as account research for a first-meeting deck, and any change beyond that waits for you to ask and approve. The skills catalogue says which. Tools that prepare an action, such as a rescue plan for a deal that’s gone quiet, return a draft: nothing is saved, sent or scheduled. On a demo workspace, every output carries a demo label.

From Claude or Codex

Before using any answer, ask the follow-up for each check. With the Summit53 connector, for example:

Which Summit53 organisation am I connected to? (whose data)
What exact dates are you comparing? (which window)
Do the parts sum to the total? (does it add up)
What was missing or not measured? (what it can’t see)
Is this a correlation, or did you establish cause? (causal words)
Did you write, send or update anything? (did it change anything)

The first one is answered by the summit53 skill, which reports the organisation, your user and whether it’s a demo workspace. See the skills catalogue.

When the data is thin

An honest blank beats a confident zero. Every surface above prefers Not enough data yet, unavailable or a dash to a number it can’t support.
Unknown isn’t failure. Results that can’t be judged stay in Watching and are left out of the worked rate.
The checks still apply to us. If a Summit53 answer fails one of them, don’t repeat it; ask the follow-up, or check the screen it came from.

Do it this week

Print the six checks on one page.
Take the last generated answer someone pasted into a deck or a meeting note.
Ask the six follow-ups of it, and write down which it passes.
For any it fails, either get the missing piece (the dates, the scope, the parts) or remove the claim.
Agree with the team that nothing goes into the commit call or the board pack without passing all six.

Troubleshooting

My numbers differ from my manager’s. The masthead shows each person their own scope: organisation, team or own book.
The answer has no dates. Ask for them before using it.
The skill says it couldn’t confirm the organisation. Check the connector is signed in to the right Summit53 account. See the MCP setup guide.
An output says Demo data. You’re connected to a demo workspace; don’t use it for real decisions.

Can I repeat this AI answer in the room? Six checks before you do