The product review is stuck on a launch risk. The team needs a blunt question: Which assumption would make this plan fail?
Instead, the AI assistant offers three upbeat prompts about alignment, summarizes the last five minutes, and praises the group for making progress. Nothing it says is obviously wrong. All of it is wrong for the moment.
In a brainstorming session, the same warmth and expansiveness might help. In a hiring debrief, it could soften a necessary comparison. In a customer call, an overly forceful challenge could damage trust even when the underlying concern is valid.
We often evaluate an AI meeting copilot by asking whether it is accurate or useful. There is another question hiding underneath: What kind of collaborator is it being right now?
Teams do not need to give their copilot a charming fictional identity. They need to choose a working style that fits the job of the meeting.
New research suggests style changes the work
In August 2026, the Proceedings of the National Academy of Sciences published a large randomized experiment on personality pairing in human-AI collaboration. The researchers paired 1,258 participants with AI agents prompted to exhibit different levels of the Big Five personality traits. The teams spent 40 minutes creating display ads, which were later rated by independent evaluators and tested in a campaign that received nearly five million impressions.
The results were not as simple as “match the person.” Some pairings improved text or image quality, while others reduced it. Effects also varied by the kind of output. The authors describe the pattern as jagged: a behavioral pairing that helped one part of the task could hurt another. You can read the peer-reviewed PNAS paper or MIT Sloan's summary of the experiment.
The evidence deserves restraint. This was one advertising task, completed by U.S. participants in a purpose-built collaboration environment. It did not study recurring workplace meetings, and the authors are co-founders of a company developing AI personalization technology. The findings do not justify assigning every employee a psychological profile or claiming that one assistant persona will raise performance across a company.
They do support a practical observation: the social and behavioral design of an AI collaborator is not decorative. It can change how people work and what they produce.
For meetings, the safest starting point is not the personality of each participant. It is the purpose and risk of the conversation.
Build a working-style card for each meeting type
A working-style card is a four-line agreement covering pace, challenge, detail, and visibility. It describes observable behavior. Nobody needs to be labeled introverted, agreeable, anxious, or anything else.
Pace: How much space should the conversation get?
Choose whether the copilot should wait to be asked, offer prompts at natural pauses, or actively track moments that need attention.
A brainstorm usually needs more uninterrupted space. A launch review may benefit from a prompt when the group moves past an unresolved dependency. An interview panel may need a quiet reminder that a required topic has not been covered, but not a stream of suggestions while the candidate is speaking.
Pace is not simply frequency. It includes timing. Three prompts delivered at a pause feel different from one prompt that cuts across a delicate sentence.
Challenge: What kind of friction should it add?
Decide whether the copilot should mainly clarify, test assumptions, seek alternatives, or look for disconfirming evidence.
“Summarize the emerging view” is a cooperative posture. “Name the strongest evidence against the emerging view” is a challenging one. Both can be useful. Trouble begins when the team expects one and receives the other.
For a decision review, ask the copilot to probe weak assumptions and missing approval. For early ideation, ask it to widen the option set before criticizing. For a sensitive one-to-one conversation, keep interpretation of motive and emotion with the people involved.
Detail: How much evidence should travel with a contribution?
An AI assistant can offer a one-line nudge, a concise explanation, or a source-backed account. Pick the level before the meeting gets busy.
High-consequence claims deserve a source and a date. A reminder that the budget changed should point to the current budget, not merely state that there may be a conflict. A low-stakes prompt such as “Have we heard from support?” can stay brief.
More detail is not automatically more trustworthy. Long answers can consume the room and make weak evidence look substantial. The right level is the smallest amount that lets people judge the contribution.
Visibility: Who should see what the copilot notices?
Some assistance belongs in a private channel to the facilitator. Some belongs in the shared meeting view. Some should appear only in the recap.
Visibility changes the social meaning of a suggestion. A private prompt that says “The customer has not confirmed the date” lets the account owner decide how to ask. Displayed to the whole room, the same message may sound like a correction or an accusation.
Use shared visibility for facts everyone needs to evaluate together. Use private visibility for ambiguous signals that require context. Make sure consequential corrections enter the final record, even if they began as a private nudge.
Write the card in plain language
The complete card should fit on four lines. For a weekly product decision review, it might read:
Pace: Wait for pauses; surface no more than two live prompts unless a confirmed constraint is at risk.
Challenge: Test assumptions, ownership, and missing approval.
Detail: Give the source and its date for factual conflicts; keep ordinary prompts to one sentence.
Visibility: Show verified conflicts to the room; send uncertain signals to the facilitator first.
For a creative workshop, the card would be different. It might protect ten minutes of uninterrupted idea generation, favor expansion over critique, keep examples short, and send pattern observations privately until the group begins narrowing options.
This is why a universal “helpful” setting is inadequate. Helpful behavior depends on the phase of the work.
Calibrate from evidence, not preference alone
Run the card for two comparable meetings, then review three moments: one contribution that improved the discussion, one that got in the way, and one moment when the copilot stayed quiet but might have helped.
Change a single dimension for the next meeting. If interruptions were the problem, adjust pace before rewriting every instruction. If prompts felt vague, increase the evidence standard rather than making the assistant more assertive.
Judge the style against the meeting's outcome. Did the team expose an important risk, produce a broader set of options, cover the agreed interview criteria, or leave with a decision people understood? “People liked the assistant” is useful feedback, but likability is not the same as contribution.
Also ask who carried the cost. A fast, combative copilot may energize confident speakers while making it harder for others to enter. A highly agreeable one may make the meeting feel smooth while leaving a weak assumption untouched. The team needs room to say that a style hindered participation without turning the discussion into a judgment of anyone's personality.
Keep personalization accountable to the room
Behavioral tuning becomes risky when it is invisible or based on inferred traits. Do not secretly classify participants from their speech, culture, accent, job level, or previous conversations. Do not let a system decide that one colleague needs more pressure and another needs more reassurance.
Instead, make the working-style card visible, editable, and tied to the meeting type. Let the facilitator override it. Revisit it when the meeting changes from exploration to decision, or when a customer joins an internal call. Preserve a clear boundary: the copilot can adapt its contribution, but people remain responsible for reading the room and for the decisions that follow.
The point is not to make AI feel more human. It is to make its behavior more legible.
Caspi supports meetings with live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. A working-style card gives those capabilities a practical social setting: how much space to take, what kind of question to ask, how much evidence to bring, and who should see it. The result is not a copilot with a better personality. It is a team with a clearer agreement about how assistance should show up.