Do Not Let the AI Brief Decide the Meeting Before It Starts

At 8:56, four minutes before a customer call, an AI assistant produces a beautiful briefing. It says the account is at risk, the customer is frustrated with onboarding, and the renewal depends on a pricing concession.

Everyone enters prepared. Unfortunately, they are prepared for the same interpretation.

The delayed onboarding is documented. The frustration comes from one sharp message written six weeks ago. The pricing concession was floated internally, not requested by the customer. By compressing those details into a confident narrative, the brief has quietly done more than gather context. It has framed the meeting.

This risk is becoming more relevant as workplace AI gains access to calendars, conversations, files, and business systems. Slack recently described contextual AI that can prepare someone for an upcoming meeting using the work they are already authorized to access. The company also notes that the same question can produce different answers for different roles because each person has different available context. Slack's contextual AI overview

That is a genuine improvement over searching through five channels while the join screen counts down. But a briefing should make the room more curious, not make uncertainty disappear.

The fix is not a longer summary or a generic warning that AI can be wrong. Use a meeting brief with three visible lanes: known, inferred, and ask.

Start with what is known

The known lane holds claims that can be traced to an identifiable source. Each item should answer three questions: what happened, where the evidence lives, and how current it is.

For the customer call, the lane might read:

Known
The implementation milestone moved from October 12 to October 26. Source: approved project plan, updated October 18.
The customer opened three onboarding tickets in the last 30 days. Source: support queue, checked this morning.
The renewal date is December 15. Source: signed order form.

Notice what is missing: no mood, motive, or prediction. Those may be useful, but they are not the same kind of claim.

Link to the exact message, document section, record, or prior meeting moment whenever the tool allows it. A folder name is not a source. Neither is "based on recent activity." If two systems disagree, show the conflict instead of selecting the tidier answer without explanation.

This source trail matters even when the brief is correct. NIST's generative AI risk profile recommends tracking the origin and history of AI inputs and outputs, including sources and modifications, because provenance helps people assess trustworthiness and trace problems back to their origin. NIST AI 600-1

Freshness belongs beside provenance. A signed agreement may be authoritative but superseded. A chat message may be recent but speculative. Useful preparation shows both authority and time.

Label the interpretation as inferred

The inferred lane contains conclusions assembled from the evidence. It is where the assistant can be most helpful, and where polished language can be most misleading.

An inference should name its basis and leave room for correction:

Inferred
Onboarding delay may be contributing to renewal risk. Basis: two recent support threads mention the delayed milestone. Counter-signal: the customer sponsor described the rollout as "back on track" in the October 24 call.
Pricing may arise. Basis: the account team discussed a possible concession internally. The customer has not requested one in the sources reviewed.

The phrase "may be" is not timid writing. It accurately describes the status of the claim. The counter-signal is equally important. It prevents the brief from becoming a prosecution in which every selected fact supports the same story.

Do not assign numerical confidence unless the number has a defined method behind it. "82% likely to churn" looks rigorous, but without a calibrated model, a relevant comparison set, and an explanation of the outcome being predicted, it is decorative precision. Plain language such as weak, mixed, or strong evidence is often more honest.

This separation also gives reviewers something concrete to edit. Instead of approving an entire narrative, they can confirm the source and challenge the interpretation independently.

Turn uncertainty into questions for the room

The ask lane converts gaps and disagreements into productive prompts. It is not a list of everything the assistant failed to find. It should contain the few questions whose answers could change the meeting's decision or next step.

For example:

Ask
Which remaining onboarding issue matters most to the customer's rollout?
Does the customer see pricing as part of the renewal decision?
What would make the next 30 days feel successful to the customer sponsor?

These questions do two jobs. They keep the team from presenting an internal theory as the customer's reality, and they give the meeting a reason to exist.

Research from Microsoft explored this kind of prospective reflection with 18 employees. Its Meeting Purpose Assistant asked people to articulate the purpose, challenges, and success conditions of upcoming meetings. Participants reported clearer goals, better preparation, and, in some cases, changed plans or canceled meetings. The lesson is modest but useful: preparation improves when AI helps people examine the meeting's purpose, not only when it retrieves more material. Microsoft Research

An ask should be answerable by someone in the room. "What are the risks?" is too broad. "Which dependency could move the launch past November 20?" can produce a decision.

Think once before reading the generated brief

There is one more safeguard, and it costs about 60 seconds. Before opening the AI briefing, ask each decision-maker to write down:

• the outcome they think the meeting needs;
• the most important fact they already know;
• the question they do not want the room to avoid.

Then open the brief and compare.

This small ordering choice preserves an independent starting point. Otherwise, the first fluent account can become the structure against which every later thought is judged. In a 2025 survey of 319 knowledge workers, Microsoft researchers found that higher confidence in generative AI was associated with less self-reported critical thinking, while higher confidence in one's own task knowledge was associated with more. The researchers described verification, integration, and task stewardship as important parts of critical thinking in AI-assisted work. Microsoft Research

The minute of independent thought is not a ritual of distrust. It gives the team another signal to compare with the machine-assembled view.

For routine status meetings, the organizer can do this alone. For a hiring debrief, pricing decision, incident review, or sensitive customer call, collect a short private response from several participants before sharing the common brief. Independent observations are especially valuable when hierarchy might otherwise make the first interpretation hard to challenge.

Keep the brief short enough to question

A useful brief fits on one screen or one page. If participants need ten minutes to absorb it, the assistant has moved the preparation burden rather than reducing it.

Use a compact order:

1. Meeting purpose and the decision or outcome needed.
2. Three to five known facts with sources and dates.
3. Up to three inferences, each with its basis and a counter-signal if one exists.
4. Two or three questions that could change the outcome.
5. Missing or inaccessible sources that materially limit the brief.

Name the person responsible for reviewing it. An automatically generated brief without an owner will be treated either as disposable or authoritative, depending on the reader. Neither is healthy.

After the meeting, compare the preparation with what the room learned. Which inference was confirmed? Which source was stale? Which question unlocked the decision? Feed those corrections into the next brief format and the underlying records. Do not simply let a new summary overwrite the old assumption.

Context-aware AI can remove the scavenger hunt from meeting preparation. It should not remove the distinction between evidence, interpretation, and inquiry.

Caspi supports live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. Used with the known, inferred, and ask structure, those capabilities can help a team bring forward relevant history, challenge an assumption while there is still time, and carry the newly confirmed understanding into follow-through. The goal is not to enter the room with one perfect story. It is to arrive with enough context to ask better questions.