How to Run a Human-and-AI Team Meeting That Ends in a Decision
How to Run a Human-and-AI Team Meeting That Ends in a Decision
The most useful thing an AI agent can do for your next team meeting is finish its part before anyone sits down.
Say your team launches a new import tool next month and you still have to decide how to position it. One option is bold: "Bring your whole history over in an afternoon." The other is narrower: "Bring your active projects over in an afternoon." Marketing likes the first one, sales isn't so sure, and engineering knows something about archived records that nobody else has heard yet.
The usual way to settle this is a meeting where everyone gives an update, someone pieces together where things stand, the group debates for a while, and the last few minutes go to working out who does what. If your agents spent the week reading competitor pages and drafting copy, most of that hour goes to hearing things people could have read beforehand.
There's a better order. Let the agents gather the evidence, lay out the options and name what's blocking each one before the meeting. Then use the room for the part people have to do: bring the context the agents don't have, argue about the tradeoff, make the call and own it.
Run that way, a decision meeting takes about 25 minutes and ends when your people and your agents both know what they're doing next.
Why the old meeting doesn't fit
Most meetings assume the important information lives in the people attending. Once agents are doing the research, a lot of it lives in their work instead. One fix is to have an agent answer questions live, and sometimes that helps. But Google's People + AI Guidebook warns against creating unrealistic expectations by presenting AI as human-like. In a meeting, the risk is that an answer delivered fluently, under time pressure, sounds more settled than the evidence behind it. If an agent tells the room the bold position tested well, that can land as fact before anyone asks which interviews it means.
So judge the agents by their work, and put it where the room can see it: the sources they found, the options they compared and the questions they couldn't answer. If nobody can tell the bold line came from an agent's reading of the interviews, nobody can check it. The people in the room are there to inspect that work and make the decisions their roles allow.
What the agents prepare before the meeting
Before the meeting, the agents put together a one-page decision packet. For the import tool, it might look like this:
| What the packet covers | In the launch example |
|---|---|
| The decision requested | Which position do we launch with? |
| Evidence | Competitor pages, the customer interviews the team supplied, the product brief and the draft copy, each linked, with the places they disagree marked |
| Options | The bold position, the narrow position, or waiting a week for engineering's answer |
| Tradeoffs | The bold line is the stronger hook. The narrow line is easier to stand behind. Waiting holds up the launch copy. |
| What's unknown or blocked | Nobody has confirmed that archived records will import at launch |
| Proposed next work | What the agents and the team would each do under each option |
Whoever owns the decision, here the product lead, reads the packet before it goes to the group. They're checking the scope, anything that shouldn't go to everyone, and obvious holes in the evidence.
Keep it to a page. Google's guidance on explainability says not to explain everything behind an AI system's output, just the parts that affect people's trust and decisions. Here, that's the claims the launch choice rests on. The rest of the research can sit behind a link.
It also has to show where the evidence runs out. "We couldn't confirm that archived records import" is more useful than a smooth paragraph that writes around the gap. If the gap is on the page, engineering can speak to it in the meeting. If it's hidden, the room is more likely to go along with the bold line without knowing what it rests on.
A 25-minute agenda
The meeting has five parts: three minutes to frame the decision, seven to inspect the evidence, then five each for dissent, the call and the handoff. Move the minutes around if you need to, but keep the order, so the opening isn't spent rebuilding status and the ending isn't spent guessing who owns the next step.
Minutes 0 to 3: frame the decision
The person who owns the decision states it in one sentence, along with the deadline, the constraints and what the group can and can't decide. In the example, the product lead says the group is choosing the launch position today, not the whole campaign. If someone can't influence or carry out that decision, they probably don't need to be there.
Minutes 3 to 10: inspect the evidence
Go through the packet. Which sources support the bold position? What contradicts it? How current is that evidence? Which claims are the agents' inference rather than something a source actually says? Open the detailed research only when it could change the decision. In the launch, marketing walks the group through the competitor pages and the interview quotes.
Minutes 10 to 15: force the dissent
Ask what would make the preferred option wrong. Which team or customer isn't represented here? What failure would look obvious in hindsight? The agents can prepare a list of objections ahead of time, but people have to add what the agents had no way of knowing.
This is where the launch turns. Engineering says archived-record import won't be confirmed before launch. Sales says customers on recent calls ask about their active projects far more than their history. The packet could only call the import unconfirmed, and it had nothing from sales calls. Now both point to the narrower line.
Minutes 15 to 20: make the call
The decision owner chooses, narrows the decision or puts it off. Putting it off only counts if you name the missing fact, who will get it and when the decision comes back. "More research" isn't a decision.
The product lead picks the narrower position and writes down why: the bold claim depends on something engineering can't promise yet.
Minutes 20 to 25: send both workforces back to work
Write down the decision and why, the dissent, who owns each next step, what the agents may and may not do, and when people look at the work again. Before anyone leaves, put that record in the brief the agents work from next.
For the launch, the agents are cleared to revise the launch article around the narrower position, write a few variations and look into two questions from the dissent. How many customers have archived records they'd actually want to move? And what do competitors promise about bringing history over? Nothing is published until product and legal have reviewed the claims. The next review happens when that work is ready, not at the next weekly status meeting.
Who should have the final say?
A named person whose role matches the decision and who will live with what happens. Agents can recommend, test options and carry out work within limits someone approved. But "human oversight" doesn't mean much unless that person can look at the evidence, reject the recommendation, change the plan and stop the work.
The NIST AI Risk Management Framework asks organizations to spell out who is responsible for what when people and AI systems work together, and to define and document how human oversight works. Its appendix on human-AI interaction points out that an AI system can make decisions on its own, defer to a human expert, or give a person an additional opinion. In the launch, the agents are doing the third: they give the product lead an opinion, and the product lead decides.
So write down who holds authority before you write the agenda. For the launch, that's five answers:
- Who owns the decision? The product lead.
- What can the agents do without another approval? Draft and revise copy, and research the open questions.
- What needs review before it happens? Anything published, and any claim about what the product does.
- Who can stop the work? The product lead, and legal for the claims.
- Who answers for it if the position turns out wrong? The product lead.
The answers will change with the task and the risk. They shouldn't change depending on who talks most confidently in the meeting.
How to keep people from rubber-stamping the agents
The launch already shows how: the packet kept linked evidence apart from the agents' pick, engineering and sales got their turn to object, and the product lead wrote down why.
The research doesn't say the combination always wins. In a preregistered field experiment with 791 professionals at Procter & Gamble, published in Organization Science in 2026, people working alone with AI on real product-innovation challenges matched the performance of two-person teams working without it. That's one company and one kind of task.
A study by Eli Ben-Michael and colleagues came out differently. In their randomized trial of a pretrial risk assessment tool, the recommendations didn't improve the accuracy of judges' decisions to impose cash bail, and replacing the judges with algorithms, including a large language model, did worse. Their framework compares decisions made by people alone, people with AI and AI alone. Whether the combination helps your team is something to measure, not assume.
The decision record gives you a way to measure it. Say the packet favored the bold position because customers in the interviews responded to it. The record shows the agents' recommendation, the product lead's decision, the objections from engineering and sales, and later, how the launch went. Over a few months you'll see when the agents' preparation helped, when people overruled it for good reasons, and when the meeting added nothing but ceremony.
When to skip the meeting
Routine agent work should keep going inside the brief and review points you've already agreed, and a clear yes-or-no choice can be approved without anyone meeting. Call a decision meeting when talking it through will settle things faster than another round of documents. That usually means:
- the evidence creates a real choice;
- a constraint needs someone with authority to decide;
- teams read the same evidence differently;
- the next step is costly or hard to undo; or
- the limits on the agents' work need to change.
Don't hold one because an agent produced an update.
After the launch decision, the agents' revised copy comes back inside the narrower position. That needs a review, not a meeting. If engineering later confirms that archived records will import, the limits have changed, and that's worth getting the group together again.
Google's guidance on feedback and control says that when people give feedback, a product should ideally let them know what the system will do next or how their input will influence it. A meeting should work the same way. When sales pointed out what customers actually ask about, that became a narrower position and a revised brief for the agents. If the feedback ends up in minutes nobody reads, the team went through the motions of oversight without controlling anything.
Where Supanova fits
Say the launch prep ran on Supanova. The product lead writes a project brief with the decision, the two positions to test and the questions the team needs answered. The atoms do the work, and the finished work comes back to the product lead for review. So the packet has been read by the person who owns the decision before the group ever sees it, which is the order the meeting depends on.
After the call, the handoff becomes the next brief, and the revised launch article comes back for review the same way before product and legal sign off on the claims.
The meeting is still the team's. The product lead frames the choice, engineering and sales add what the atoms couldn't know, and the product lead makes the call.
Before your next team meeting, ask what decision it's for. If you can't say it in one sentence, the agents' work isn't ready for the room yet, and the meeting may not need to happen at all.
Sources
- AI Risk Management Framework Core, NIST
- Appendix C: AI Risk Management and Human-AI Interaction, NIST AI RMF
- People + AI Guidebook: Mental Models, Google PAIR
- People + AI Guidebook: Explainability + Trust, Google PAIR
- People + AI Guidebook: Feedback + Control, Google PAIR
- The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork, Dell'Acqua and colleagues, Organization Science (2026)
- Does AI help humans make better decisions? A statistical evaluation framework for experimental and observational studies, Ben-Michael and colleagues, arXiv (2024)