Článek v tomto jazyce není k dispozici – zobrazujeme verzi v angličtině

Are Teams Still Relevant?

• Autor Andreas Schliep • 13 min read

Adaptive organisations Agile Principles Teams AI in Collaboration
Symbolbild zu Teamwork als Fussballspiel

Are Teams Still Relevant?

Revised 9 October 2026: several claims in the original version were wrong or unsupported, and the argument has been rebuilt. See the revision note at the end. The revision was prepared with AI assistance.

Two developments put the traditional team in question. Remote work has become a lasting part of working life: in the United States, full days worked from home made up 28 percent of paid workdays in mid-2023, about four times the 2019 rate (Barrero, Bloom, and Davis). And tools like Claude and ChatGPT enable new forms of collaboration between humans and machines.

As Peter Beck argues in his article “The AI-Enhanced Team of the Future”, agentic AI now lets a single engineer coordinate dozens of specialised code agents to ship production features. This is an observation from practice, not a research finding. (Disclosure: Peter Beck is my business partner, and we developed AME3 together.) Beck does not conclude that teams disappear. He expects their composition and focus to change into what he calls “AI-Enhanced Teams” and, in software, “Full Business Teams”.

My answer up front: teams remain relevant, but only under conditions. I take the five conditions that Richard Hackman identified for effective teams and ask of each what dynamic teaming and AI do to it.

The External View: What a Team Is and When It Is Worth Having

The Agile Manifesto from 2001 values “Individuals and interactions over processes and tools”, and its fourth principle asks business people and developers to work together daily. Most agile practices, such as retrospectives and daily coordination, presuppose a stable group that shares accountability. That is my reading; the Manifesto does not define a team.

Katzenbach and Smith do. In The Wisdom of Teams (1993) they define a team as “a small number of people with complementary skills who are committed to a common purpose, performance goals, and approach for which they hold themselves mutually accountable.” Whether this definition holds when AI agents take on team tasks is an open question. It would fail if mutual accountability could not be maintained, for example if nobody on the team could answer for what an agent produced.

Richard Hackman named five conditions for team effectiveness in Leading Teams (2002): a real team, a compelling direction, an enabling structure, a supportive organisational context, and expert coaching. Hackman did not treat teams as the default. Teams are worth forming when the work needs interdependent contributions, and in a 2009 interview he said there are many things individuals can do better on their own.

A caution on evidence: the first version of this article claimed large performance gains for agile ways of working. Its source, a McKinsey Global Survey of 2,190 respondents published in 2021, reports gains of around 30 percent only for highly successful transformations, based on self-reports; the average chance of success is 30 percent. It is a consultancy survey and says nothing about teams as such, so the argument below rests on team research.

The Internal View: Five Conditions Under Pressure

Dynamic teaming – forming and re-forming teams around the work – goes back to Heidi Helfand’s Dynamic Reteaming. AI pairing – people working with tools like GitHub Copilot or with coding agents – changes how work is divided between people and machines.

1. A real team. The first condition asks for clear boundaries, interdependent work, and stable membership. Dynamic teaming puts stability at risk. Hackman cites the US National Transportation Safety Board: 73 percent of the incidents in its database happened on a crew’s first day of flying together. The underlying study is narrower than that sounds: of 37 major accidents between 1978 and 1990, pairing data were available for 15, so the figure rests on 11 accidents. The base is small, but other work points the same way. Huckman, Staats, and Upton found in an Indian software services firm that team familiarity had a significant positive effect on performance. Mortensen and Haas argue that today’s teams are fluid, overlapping, and dispersed, and that the resulting blurred boundaries can undermine effectiveness. Reteaming has a price.

AI presses on interdependence instead. In a preregistered field experiment with roughly 780 professionals at Procter & Gamble, Dell’Acqua et al. found that individuals with AI matched the performance of teams without AI. Teams with AI did best, although the difference to individuals with AI was not statistically significant, and they were about three times as likely to produce top-tier solutions. The limits matter: the “teams” were pairs, in one company, on a short task. I read this with Hackman, not against him: where work no longer needs interdependent contributions, an individual with AI is the better design, and the organisation needs fewer teams, not worse ones.

2. A compelling direction. I expect this condition to gain weight. Beck’s practitioner view is that teams become responsible for complete business capabilities, with AI as a force multiplier. The 2025 DORA report supports the amplifier idea with survey data: AI magnifies an organisation’s existing strengths and weaknesses. In DORA’s own summary, throughput rises, but often at the cost of stability. My inference, not a finding: when output gets cheaper, a team without a shared direction produces more of the wrong thing, faster.

3. An enabling structure. For Hackman this covers the design of the task, the number and mix of members, and norms of conduct. On the mix, AI may help. In the Procter & Gamble experiment, R&D and commercial professionals proposed solutions typical of their function without AI and more balanced ones with it. On norms, AI adds a requirement. A randomised trial by METR with 16 experienced open-source developers and 246 tasks found that they took 19 percent longer with AI, yet believed afterwards that AI had made them about 20 percent faster. That reflects the tools of early 2025. It does not show that AI slows people down, but that we judge our own AI-assisted productivity poorly. In a survey of 319 knowledge workers, higher confidence in generative AI was associated with less self-reported critical thinking (Lee et al.). I conclude that a team working with AI needs an explicit norm for checking AI output and for measuring its results.

Which structure fits which work is what practitioner models such as Org Topologies by Alexey Krivitsky and Roland Flemm address. It maps organisational units by the breadth of their capabilities and the scope of their work, currently in 16 archetypes; the seven in the early introductory article were, by the authors’ own account, generalised from consulting experience, not derived empirically.

4. A supportive context. This condition is often listed and rarely applied, as in the first version of this article. It concerns the reward, human resource, and information systems around the team. Hackman observes that systems built to guide and correct individuals are one of the main reasons teams do less well than they could. Both developments move load onto this condition. If people change teams often, what a stable team would carry itself – working agreements, standards, knowledge of who knows what – has to live in the organisation. Jurgen Appelo’s unFIX model, which names Helfand’s book as an inspiration, is one design for this. Besides the Crew, which it expects to be stable though not static, its structural patterns include Forum, Base, and Turf; the Forum exists so that people in similar roles agree on how work is done, which makes reteaming less painful. Information flows depend on context too: a study of 61,182 US Microsoft employees found that firm-wide remote work made collaboration networks more static and siloed (Yang et al.). And if AI raises individual output, I expect a reward system that measures individuals to pull people out of teamwork, however well the team itself is designed.

5. Expert coaching. Hackman found that coaching individual members does little for team performance; teams need coaching as a group, on their processes. Today’s AI tools assist individuals. I know of no evidence that they replace coaching at group level, so I treat this condition as unchanged, and as a gap.

How Do We Start the Conversation?

The four metaphors from the first version of this article can now be read as positions on these conditions:

  1. The Solid Fortress secures the first condition by keeping membership fixed. It fits work that is strongly interdependent and where familiarity is expensive to rebuild, as in a cockpit. That it reconfigures slowly is my judgement, not Katzenbach and Smith’s.

  2. The Swarm trades stability for flexibility. It works only where the fourth condition carries what the team no longer can; otherwise it pays the price Hackman documents, lost familiarity.

  3. The Isolated Islands are individuals working with AI. They fit work that does not need interdependent contributions. The evidence so far covers short tasks only.

  4. The AI Hybrid is a real team that uses AI, as Beck outlines it. In the one experiment we have, it produced top-tier solutions most often, and it amplifies whatever conditions are already in place.

The question is then not which metaphor we like. For each piece of work it is: does it need interdependent contributions? If not, an island is the right answer. If so: which of the five conditions is missing, and can the organisation supply it? Integrating AI without losing familiarity means keeping membership stable where interdependence is high, and investing in the context wherever people move.

Conclusion

Teams remain relevant where work needs interdependent contributions and the five conditions can be met. Dynamic teaming and AI make the first condition harder to meet, raise the stakes of the second, change the third, shift load onto the fourth, and leave the fifth where it was. Haas and Mortensen have begun this kind of reinterpretation for teams that are more diverse, dispersed, digital, and dynamic.

This position can be wrong. It would fail if larger teams working over longer periods did no better with AI than individuals with AI on interdependent work. It would also fail if frequently re-formed teams, given a supportive context, performed as well as familiar ones. Neither has been tested, as far as I can see.

Until then, my hypothesis is conditional: if Hackman’s conditions carry over, organisations that decide per piece of work whether a team is needed, and that build the context for the teams they keep, will adapt better than those that default to fixed teams or to individuals with AI.

Your practical experiences with modern team concepts enrich this discussion. The association looks forward to exchanging ideas with you.

What the evidence says

  • Team familiarity matters: crews and project teams that have worked together perform better (Hackman; Huckman et al. 2009).
  • On a short innovation task, individuals with AI matched two-person teams without AI; teams with AI produced top-tier solutions most often (Dell’Acqua et al. 2026).
  • Self-assessed productivity gains from AI are unreliable (METR 2025), and confidence in AI goes along with less critical thinking (Lee et al. 2025).
  • AI acts as an amplifier: higher throughput, often at the cost of stability (DORA 2025).
  • Firm-wide remote work made collaboration networks more static and siloed (Yang et al. 2022).

What we assume

  • Hackman’s five conditions carry over to AI-supported and distributed teams.
  • The findings on pairs and short tasks say something about larger teams and long-running work.
  • A supportive context can make up for part of the familiarity that dynamic teaming costs.
  • Cheaper output makes a shared direction more important, and individual rewards weaken teamwork when AI raises individual output.
  • Practitioner models such as unFIX, Org Topologies, and AME3 describe useful options; they are not empirically validated.

References

Revision note

Revised 9 October 2026 (first published 4 November 2025).

Corrected:

  • McKinsey: the original claimed gains of “30 to 50 percent” from “current studies”. The source is a consultancy survey reporting around 30 percent, for highly successful transformations only.
  • Hackman: Leading Teams names five conditions, not four; the “real team” was missing. The quotation “authority vacuum” could not be verified and was removed.
  • unFIX: there is no “Crowds” pattern. Dynamic teaming is now attributed to Heidi Helfand.
  • Org Topologies: the description of the axes and the number of archetypes were wrong.
  • Statements attributed to Peter Beck or to Katzenbach and Smith that are my own judgments are now marked as such; the business relationship with Peter Beck is disclosed.

Restructured:

  • The article now tests Hackman’s five conditions against dynamic teaming and AI, where the original surveyed frameworks.
  • The four metaphors are derived from these conditions, and the closing questions are answered.
  • The conclusion names the findings that would refute it.
  • Added: research on team familiarity, AI and teamwork, and remote work, an overview of evidence and assumptions, and a reference list.

AI assistance: this revision was prepared with the help of an AI assistant (Claude, Anthropic), which checked the sources, drafted the restructured text and produced the German and Italian versions. I reviewed the result and am responsible for its content.


About the Society

The Society for Adaptive Organisations e.V. is a coalition of leading European organisational developers, coaches, and trainers, working to create unified quality standards for education and training in organisational development and supporting companies in uncertain markets. The goal is to help organisations achieve greater customer focus, effectiveness, and resilience.