The 2026 Pattern
Amazon, Alphabet, Microsoft, and Meta will spend an estimated $635 billion to $700 billion on capex in 2026, roughly 75% of that funding AI infrastructure, up from $381 billion in 2025 (Yahoo Finance). Layoffs.fyi tracked more than 92,000 tech layoffs through late April 2026. Meta announced 8,000 cuts on April 23 (CNBC). Microsoft opened the first voluntary buyout program in its 51-year history the same day, offering separation to roughly 8,750 US employees (CNBC). Amazon cut 14,000 corporate roles in October 2025 with $1.8 billion in severance, and another roughly 16,000 in January 2026. CNBC asked whether the 20,000 cuts at Meta and Microsoft signal an AI-driven labor crisis (CNBC).
Goldman Sachs estimated in March 2023 that 300 million jobs globally are exposed to AI automation, with 25% of US work hours automatable (Goldman Sachs). The IMF widened that in January 2024: 40% of global jobs exposed, 60% in advanced economies (CNBC). Anthropic CEO Dario Amodei told Axios in May 2025 that AI could "wipe out half of all entry-level white-collar jobs" and spike unemployment to 10% to 20% within five years.
The counterweight is small-team leverage. Cursor crossed $1 billion ARR by late 2025 with fewer than 300 employees and a $29.3 billion valuation (Anysphere). Midjourney reached $500 million in revenue in 2025 with around 100 employees, profitable since August 2022 (Sacra). Capital is moving from bodies to silicon, and the human layer that survives produces orders of magnitude more leverage per person.
Zuckerberg Says Teams Are Obsolete (And Klarna Already Tried)
On the January 29 2026 Meta earnings call, Mark Zuckerberg made the organizational case directly. "Projects that used to require big teams can now be accomplished by a single, very talented person," he said. "We're elevating individual contributors and flattening teams" (Axios). CFO Susan Li said output per engineer was up 30% since the start of 2025 and "power users" of AI tools were up 80% year over year. Meta's 2026 capex guidance is $115 billion or more.
That is the marquee argument for the flat-team future. The case study against it ran inside Klarna from February 2024 to May 2025.
In February 2024, Klarna announced its AI assistant had handled 2.3 million customer service conversations in its first month, equivalent to 700 full-time agents (Klarna). It accounted for 75% of chats. Repeat inquiries dropped 25%. Resolution time fell from 11 minutes to under 2. Klarna estimated $40 million in profit improvement and froze hiring (OpenAI case study). It was the marquee case for AI-replaces-humans economics.
Fifteen months later, Klarna reversed. CEO Sebastian Siemiatkowski told Fortune: "As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality" (Fortune). "Really investing in the quality of the human support is the way of the future for us. It's so critical that you are clear to your customer that there will always be a human if you want." Klarna had to rebuild a function it had dismantled.
The Zuckerberg framing identifies a real shift. AI does cut coordination overhead and let smaller groups produce work that previously required scaffolding. The framing misreads what is eliminated. Coordination overhead goes down. The meta-skill of coordinating in fluid systems, when stakes are high and information is incomplete, goes up in value. Smaller teams need more teaming capability per person. Andy Jassy walked back the AI framing of Amazon's October 2025 cuts two days later, calling them "not really financially driven, and it's not even really AI-driven, not right now at least" (Fortune). Even the CEOs running the cuts disagree about the cause.
What AI Can Do (And Where It Stops)
Cursor, Copilot, and Claude Code write production code at scale. Drafting, translation, legal research, and customer service have moved from humans to models. PwC's 2025 Hopes and Fears Survey of 49,843 workers across 48 countries found daily GenAI users reported productivity benefit at 92% versus 58% for infrequent users. Only 14% of workers use GenAI daily.
Where the capability stops is at the system level. The arXiv 2025 paper "Why Do Multi-Agent LLM Systems Fail?" finds that system-level failures dominate, emerging from "the ways agents interact, delegate, and interpret each other's outputs." The 2025 ACM literature names the same bottleneck: "coordination and comprehensibility of the collective system," with failure modes including "emergent behaviors, coordination breakdowns, looping dialogues." Gartner forecasts 30% of agentic AI projects will be abandoned at proof-of-concept by end of 2025.
What AI cannot do is precise: forming trust quickly, resolving conflict, sense-making under uncertainty, deciding with incomplete information, recovering from collective failure. This is the substrate of coordinated action. It is the exact list that fails inside multi-agent systems.
The Teaming Research Foundation
Amy Edmondson named the construct in her April 2012 Harvard Business Review article "Teamwork on the Fly." Her definition: teaming is "coordinating and collaborating across boundaries, without the luxury of stable team structures." She framed it as a verb. "Teaming is teamwork on the fly. It is a dynamic activity, not a bounded, static entity."
The research underneath is deep. Hackman and Wageman documented that up to 80% of variance in team effectiveness is explained by six conditions. Frazier and colleagues' 2017 meta-analysis in Personnel Psychology pooled 136 samples covering more than 22,000 individuals and roughly 5,000 groups, confirming psychological safety as a robust predictor of learning, performance, and information sharing (Frazier et al., 2017). Google's Project Aristotle studied 180 teams. Of the five dynamics that distinguished high performers, psychological safety ranked first.
The most relevant recent evidence comes from Edmondson, Kerrissey, and Bahadurzada's November 2025 HBR piece "In Tough Times, Psychological Safety Is a Requirement, Not a Luxury." The team surveyed more than 27,000 healthcare workers across a US hospital system in May 2019 and again in May 2021. A one standard deviation increase in psychological safety in 2021 was associated with a 0.72-point drop in burnout and a 0.63-point rise in willingness to stay. Effects were strongest for physicians, women, and people of color. The construct became more protective during crisis.
The Cases Where Teaming Was the Leverage
Four cases show teaming as built infrastructure, engineered into the work itself.
Aviation Crew Resource Management. NASA hosted a workshop in San Francisco from June 26 to 28, 1979, with 70 people from 32 organizations across 9 countries. The trigger was the NTSB investigation of United Flight 173, which crashed in December 1978 because cockpit hierarchy suppressed copilot voice. NASA research showed 60% to 80% of aviation accidents involved interpersonal communication, leadership, and decision-making failures. United Airlines launched the first comprehensive CRM program in 1981 (FAA). CRM became the global standard by the 1990s. Aviation safety improved by orders of magnitude. Teaming capability can be engineered into an industry.
Surgical teams adopting new technology. Edmondson, Bohmer, and Pisano studied 16 hospitals adopting minimally invasive cardiac surgery in 2001. Only 6 of 16 succeeded. The successful teams designed for learning, framed the challenge to motivate learning, and built psychological safety. The hierarchical "surgeon runs the show" model failed (Disrupted Routines).
Formula One pit crews. McLaren executed a 1.80-second tire change at the 2023 Qatar Grand Prix, lap 27, on Lando Norris's car. It is the Guinness world record (Guinness). Twenty people, perfect coordination, repeated thousands of times.
Pixar's Braintrust. Ed Catmull described it in his September 2008 HBR article "How Pixar Fosters Collective Creativity." A trusted peer group reviews films-in-progress. The Braintrust has no authority to mandate changes. The director decides. Candor without authority is what makes it teaming infrastructure.
The Coordination Tax Inversion
Pre-AI, much of what looked like teamwork was coordination overhead: status updates, handoffs, the friction tax of moving information through people. AI cuts deeply into that overhead. What is left is pure teaming: the coordination that requires human judgment under uncertainty. The premium on that residue rises.
PwC's 29th Global CEO Survey 2026 reports "required skills are changing 66% faster in the most AI-exposed occupations." 49% of global CEOs believe junior employment will decrease in the next three years due to AI. The survey concludes: "the highest-performing organizations will be those in which people and AI co-create." Gartner's October 2025 release on CHRO 2026 priorities, based on 426 CHROs, names harness AI, mobilize leaders, address culture atrophy, and workforce redesign for the human-machine era. Nobel laureate Daron Acemoglu frames the policy version: "We currently have the wrong direction for AI. We're using it too much for automation and not enough for providing expertise and information to workers" (MIT). Organizations that win in 2026 will treat AI as augmentation for teams that already coordinate well. Substituting AI for coordination itself produces the Klarna outcome.
The Apprenticeship Crisis
The hardest version of the teaming case is generational. Human teams reproduce coordination capability through apprenticeship: juniors absorb the tacit knowledge of how seniors decide, escalate, and recover by being present in real situations. AI is now eating the bottom rung of that ladder.
The NC State ERM initiative surveyed 1,540 board members and C-suite executives in December 2025 with Fortune. Skills shortage now ranks 5th among long-term risks. NC State's Mark Beasley: "AI is way different in the sense of it is now replacing the job. Knowledge is sort of now free in some ways. Thinking now has to really kick in" (Fortune). Economist Julia Coronado named the structural problem: "If AI is sort of replacing the entry-level typical positions, how do I prepare the future middle if I don't give them that ability at the base?"
If senior teaming capability is built through years of doing the work alongside more experienced peers, and AI now does that work, the next decade's senior teaming capability has nowhere to compound. Institutional knowledge of how teams team is at risk of breaking at exactly the moment it is most needed.
Counter-Arguments
Three objections deserve direct engagement.
"AI agents will eventually do team-level work too." The frontier research argues otherwise. The arXiv 2025 multi-agent failure paper finds system-level failures dominate, originating in "the ways agents interact, delegate, and interpret each other's outputs." Gartner expects 30% of agentic AI projects abandoned at proof-of-concept. Human social trust evolved over 200,000 years; it remains the only known coordination substrate that survives ambiguity, recovery from failure, and conflict at the same time. Coordination across agents is precisely the unsolved problem the literature names.
"Teaming is just soft skills rebranded." The data say otherwise. Soft skills are individual attributes: empathy, communication, adaptability. Teaming is a coordination property of a system. Hackman and Wageman's up-to-80% variance, Frazier's 22,000-individual meta-analysis, Edmondson's 27,000-worker study, Project Aristotle's 180 teams: half a century of empirical work distinguishes the construct from interpersonal warmth. Treating teaming as soft skills produces budget for a yearly offsite and a Slack bot, then wonders why the team cannot decide under pressure.
"Big companies don't need teaming. They have AI." Klarna paid for that lesson. Microsoft's $145 billion 2026 capex paired with voluntary buyouts is restructuring what teams do; teams remain. The Edmondson 27,000-worker study showed psychological safety became more protective during crisis. PwC's data shows productivity benefit concentrated among the 14% of daily GenAI users, who are disproportionately embedded in teams that share know-how. Companies treating teaming as overhead will discover, on a longer lag, that they cut the wrong thing.
The QuestWorks Approach
The teaming case has been clear in the research for fifteen years. What changed in 2026 is the urgency. AI absorbed enough individual cognitive work that the coordination layer is now the bottleneck. Yuval Noah Harari's framing in Nexus applies: "Cooperation and trust, built through shared stories and narratives, are the foundation of human societies and economies."
QuestWorks is the Team Intelligence Engine. It runs on its own cinematic, voice-controlled platform and works with Slack as the integration layer. Teams of 2 to 5 enter a 25-minute weekly quest: a multiplayer scenario built from the simulation principles that produced Crew Resource Management for aviation. Inside the quest, the team coordinates, decides, recovers from failure, and debriefs. Behavioral signals populate QuestDash, a strengths-based leaderboard visible to everyone. A separate Weekly Team Health Report goes to leaders. HeroGPT, the Slack-hosted AI coach, runs private conversations that never share upstream. The 9 HeroTypes are public profiles.
QuestWorks is $20 per user per month with a 14-day free trial. Participation is voluntary, never tied to performance reviews. The brand line, Team Intelligence, Powered by Play, names the method: simulation-grade practice produces the behavioral data that makes teaming visible, and play sustains it. Related arguments: The Teaming Premium, AI Brain Fry in Engineering Teams, Why the Best Engineers Leave, and Team Intelligence as a New Category.
Sam Altman wrote in his January 2025 Reflections: "We continue to believe that iteratively putting great tools in the hands of people leads to great, broadly-distributed outcomes." The tools are arriving. The teams that learn to coordinate around them will compound.
Try it free
