The Solo Player with AI Beats the Two-Person Team Without It

See also (wiki): productivity-rcts · workflow-redesign · training-architecture

Executive Summary

• Harvard Business School’s Cybernetic Teammate study — the largest RCT on AI and teamwork (776 P&G professionals, pre-registered, May–July 2024) — finds that an individual using AI matches the output quality of a two-person team without AI. One person with AI does the work of two.
• AI-augmented teams produced the best results overall: human teams using AI were 3x more likely to produce top-10% ideas than individuals working alone without AI.
• AI erased functional silos. R&D and Commercial professionals without AI produced domain-trapped solutions (technical vs. commercial). With AI, both groups produced balanced, integrated solutions regardless of their training background.
• Less experienced workers gained the most. Without AI, junior employees underperformed even in teams. With AI, they matched the output of experienced teams.
• The silo-elimination finding has an organizational design implication executives are not yet acting on: if AI can give a generalist access to specialist-quality output, the traditional case for rigid functional boundaries weakens.

The Study

Dell’Acqua, Ayoubi, Lifshitz, Sadun, E. Mollick, L. Mollick, Han, Goldman, Nair, Taub, and Lakhani. “The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise.” NBER Working Paper 33641 (2025). HBS Working Paper 25-043.

776 professionals at Procter & Gamble working on real product innovation challenges in May–July 2024. Pre-registered randomized controlled trial. Four conditions:

• Individual without AI (baseline)
• Individual with AI
• Two-person team without AI
• Two-person team with AI

AI tool: GPT-4/GPT-4o. Participants were given access and brief training; they were not AI power users. Researchers evaluated output quality through blind expert scoring.

Credibility: HIGH for internal validity. Pre-registered RCT with clear randomization, blind evaluation, real tasks (product innovation, not hypothetical scenarios), substantial sample size. One disclosure requires flagging: P&G provided financial support to Harvard’s Digital Data Design Institute (D3) during 2023–2025, the period covering this research. The research team is multi-institutional (HBS, MIT, Wharton), the design was pre-registered, and evaluators were blind — these safeguards reduce but do not eliminate the conflict. Treat quantitative findings as directionally reliable while noting the sponsor relationship.

What the Four Conditions Found

Performance by Condition

The individual with AI (+0.37 SD) exceeds the two-person team without AI (+0.24 SD). That gap — 0.13 SD — means one person with AI outperformed two people without it, on average.

The difference between an AI individual and an AI team (+0.37 vs. +0.39 SD) was not statistically significant at average performance levels. However, for top-tier output — the kind of solutions that drive strategic differentiation — the picture changes:

Teams using AI were 3x more likely to produce top-10% quality solutions than individuals without AI.

The Dell’Acqua framing from the HBS Working Knowledge article: “If you want to empower an individual to be as effective as a team, give them AI. But if you want to be in that top 10% of performers, a full human team plus AI seems like the recipe for success.”

Time Savings

• AI-enabled individuals: 16% faster than baseline
• AI-enabled teams: 13% faster than baseline

The efficiency gain is additive to the quality gain. AI-using groups were both faster and better — the tradeoff that usually defines knowledge work (speed vs. quality) did not materialize here.

The Silo Elimination Finding

This is the non-obvious result that most coverage of this study underweights.

Without AI: R&D professionals produced technically-oriented solutions. Commercial professionals produced commercially-oriented solutions. Each group defaulted to their training. This is the functional silo pattern that costs large organizations enormous coordination effort and integration time.

With AI: Both R&D and Commercial professionals produced balanced, integrated solutions regardless of their functional background. The AI effectively gave each professional access to the other function’s domain knowledge in real time.

From Dell’Acqua: “Now we have many more parts of any given company that can contribute great ideas. Much less relevant is the traditional conception of who has task and domain expertise.” And: “If these expertise silos can have a very different shape with AI, we may want to rethink the design of organizations.”

For mid-market companies — where functional silos are frequently cited as the primary barrier to cross-departmental projects — this finding has direct operational value. AI may reduce the coordination cost of cross-functional work more than any organizational intervention.

The Experience Gradient (Again)

The Cybernetic Teammate replicates the Brynjolfsson finding in a different context:

Without AI, less experienced P&G professionals performed poorly even when paired with a teammate. The collaborative benefit of the two-person team was insufficient to bring junior workers to experienced-worker output levels.

With AI, less experienced professionals matched the performance of experienced teams. The AI compressed the skill gap — the same mechanism Brynjolfsson identified in customer support (two months with AI ≈ six months without it) appears in P&G’s product innovation context.

Two independent RCTs, different companies, different task types, same gradient. The pattern is strengthening toward a generalizable finding: AI’s biggest gains accrue where skill gaps are largest.

The Emotional / Social Dimension

Less studied, but directly relevant to change management:

Employees using AI reported significantly higher levels of excitement, energy, and enthusiasm — and lower levels of anxiety and frustration — compared to individuals working without AI.

The authors interpret this as AI fulfilling part of the motivational role traditionally provided by human teammates: providing a responsive interlocutor, validating ideas, and reducing the loneliness of solo knowledge work. This has implications for remote and hybrid organizations where isolation-driven disengagement is an ongoing problem.

This is not a primary finding and should not be treated as causal evidence. But it provides a directional signal worth tracking: AI may improve employee experience in knowledge work settings, not just output.

What This Study Does Not Cover

Single-company, single-task type. P&G product innovation is a bounded, well-defined task with scoreable output. Legal analysis, financial modeling, and client-relationship work have different task structures and less scoreable outputs. The 0.37 SD gain does not translate automatically.

Short-duration tasks. The experiment covered specific innovation challenges within a defined timeframe. Longer-horizon projects — multi-week strategy work, annual planning, deal execution — may show different dynamics.

Participant AI proficiency was limited. Participants had brief training but were not prompt-engineering experts. Mollick notes the results are “likely a lower bound” — more experienced AI users would probably show larger gains. This is a credible argument, but it also means the exact numbers should be treated as a floor, not a ceiling.

Teamwork is more than task output. The study measures idea quality and time. It does not measure the relationship capital, trust development, or organizational learning that happen inside human teams. Replacing team collaboration with AI-individual work may have costs that do not show up in 90-minute innovation challenges.

Key Data Points

What This Means for Your Organization

The most direct implication: if your headcount decisions are based partly on the assumption that two people produce better work than one, the Cybernetic Teammate data suggests that assumption is now conditional on AI access. An AI-equipped individual at average performance levels matches a two-person team. The calculus for how many people you need on a knowledge-work task has changed.

The top-10% finding adds nuance. For routine deliverables — reports, analyses, first drafts, market scans — one person with AI covers the work. For the initiatives where breakthrough quality is the actual goal — new product strategy, M&A due diligence, competitive response — a full team with AI produces materially better results than either alone. The deployment question is not “AI or humans” but “what level of output quality does this task actually require?”

The silo-elimination finding is the one most executives overlook. If AI gives every functional professional access to integrated cross-domain perspective, the coordination overhead that justified rigid functional boundaries shrinks. This does not mean flattening your organization structure. It means the assumption that a great commercial idea requires a commercial specialist is no longer always true — and the organizational design implications of that shift are worth examining before your competitors get there first.

If you are working through where these findings apply to your specific team structure — what tasks qualify for AI-individual efficiency vs. AI-team breakthrough quality — that is exactly the kind of question worth a direct conversation — brandon@brandonsneider.com.

Sources

1. Dell’Acqua, Ayoubi, Lifshitz, Sadun, E. Mollick, L. Mollick, Han, Goldman, Nair, Taub, Lakhani. “The Cybernetic Teammate.” NBER Working Paper 33641 (2025). HBS Working Paper 25-043. 776 P&G professionals, pre-registered RCT, May–July 2024. Multi-institutional team (HBS, MIT, Wharton). CREDIBILITY: HIGH for internal validity; note P&G financial support to HBS D3 Institute during study period. URL: https://www.nber.org/papers/w33641 | SSRN: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5188231

2. Dell’Acqua quote — HBS Working Knowledge. “When AI Joins the Team, Better Ideas Surface.” October 2025. URL: https://www.library.hbs.edu/working-knowledge/when-ai-joins-the-team-better-ideas-surface — CREDIBILITY: HIGH (primary author interview, institutional source).

3. Mollick, Ethan. “The Cybernetic Teammate.” One Useful Thing (Substack), 2025. Co-author summary with additional methodological commentary. URL: https://www.oneusefulthing.org/p/the-cybernetic-teammate — CREDIBILITY: HIGH as author interpretation; note potential advocacy lean.

4. Brynjolfsson, Li, Raymond. “Generative AI at Work.” QJE 140(2): 889–942 (2025). Cross-referenced for experience gradient finding replicated in both studies. See research/01-ai-native-landscape/nber-generative-ai-at-work-brynjolfsson-2023.md.

Brandon Sneider | brandon@brandonsneider.com April 2026