Design human–AI teams around the work, not the org chart

Human–AI teams should be designed around a shared decision and complementary expertise, so that AI expands the perspectives available without removing the people who must challenge and own the result.

What changes when AI can cross functional boundaries?

The 2026 published P&G field experiment found that AI-supported participants produced more balanced technical and commercial proposals. It also found value in human evaluative judgment. These results concern a bounded innovation task; they do not establish that an organization can eliminate cross-functional teams or the relationships that make implementation work. PG26

The management opportunity is to reduce the distance between perspectives. A commercial lead can ask better technical questions before a meeting. An engineer can explore a buyer’s constraints without waiting for a fully formed handoff. The team gains when those preparations improve the disagreement it can resolve together.

Give the team one decision and different challenge roles

A group asked to “use AI for innovation” will produce different kinds of activity without necessarily producing a choice. Name the decision instead: which customer problem deserves the next product experiment? Then assign the commercial owner to challenge demand, the technical owner to challenge feasibility, and the decision owner to settle the bounded next commitment.

AI can assemble a common brief and propose alternative explanations. Each specialist should examine the part they are qualified to contest, including the assumptions the common brief obscures. The design preserves complementary judgment rather than asking everyone to approve the same model-generated answer.

Share the reviewed outcome, not every private exploration

Private exploration lets people ask unfinished questions and test weak ideas. A team does not need every draft or conversational detour to make a useful decision. It needs the selected argument, its evidence, the unresolved disagreement and the next action. Sharing that reviewed object creates a clearer starting point for the next person.

It also creates a permission question. A source mentioned in a summary does not become accessible merely because the summary is shared. The project owner must distinguish the visible decision from restricted inputs and verify the receiving team’s access. This article proposes a working pattern, not a claim about any platform’s security enforcement.

What should replace a status-heavy meeting?

Use the meeting for choices that require multiple perspectives, not for reading every prepared paragraph aloud. Circulate the decision brief, ask each functional owner to identify a challenge, and focus discussion on the uncertainty that changes the next commitment. Keep dissent visible when the team proceeds conditionally.

The strongest counterargument is that common preparation can create common blind spots. A polished shared brief can anchor everyone to the same framing. Counter it by asking specialists to record their initial objection before seeing the recommendation and by assigning someone to defend the strongest alternative. Do not measure success only by faster consensus.

Pilot the team design on one real question

Next week, a product or business-unit leader can select one cross-functional decision, agree its owner and compare the existing collaboration pattern with a reviewed-brief pattern on comparable work. Log what new evidence each function contributed, which disagreement changed the choice, and whether the receiving owner could act without reconstructing the discussion.

Here, one cycle means the full preparation, decision and receiving-owner handoff for one recurring decision. The business-unit leader names a pilot evaluator and agrees the comparison class, complexity bands and acceptance criteria before starting. Pause the affected handoff immediately if a required challenge or execution owner is missing; use the two-cycle review to examine the other, still-authorized work. Count rejected and unfinished decisions as well as accepted ones.

Choose the challenge roles from the people affected by the commitment, not from a fixed commercial–technical pair. Include the operations or resource owner who must execute it, and a control owner when the scope requires one. A balanced proposal is not executable if the person who must supply capacity has never agreed to the next step.

Review after two cycles. If the new pattern reduces specialist challenge, hides dissent or produces repeated follow-up clarification, change the shared object or restore direct collaboration. Fewer meetings are not the goal. Better-informed commitments that the team can execute are the test.

Questions managers ask

Should AI replace cross-functional teamwork?

The cited innovation experiment does not establish that. AI can broaden the perspectives available during preparation; specialists still challenge assumptions, choose and implement the work.

What should a team share after private AI work?

A reviewed decision brief containing the relevant evidence, alternatives, unresolved issues, owner and next review. Access to the original sources remains a separate permission question.

How can teams avoid agreeing too quickly with AI?

Record independent objections before viewing the shared recommendation, preserve dissent and examine the strongest alternative. Check whether the new process improves specialist challenge, not just meeting speed.

Sources and scope

PG26: The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork — Organization Science; 2026-06-12. Primary peer-reviewed field experiment; public full text.

MS26: Agents, human agency, and the opportunity for every organization — Microsoft WorkLab; 2026-05-05. Primary vendor-sponsored research; public report.

GAL26: AI in the Workplace: What Separates Adopters and Holdouts — Gallup; 2026-04-12. Primary survey publisher; public article.

Research checked 2026-10-27. Next review 2026-11-27. Suggested pilots are editorial proposals, not observed Upskill Pro customer results.