9) Nils Brede Moe om AI-paradokset, autonomi, læringskultur og fremtidens team

Truls Jørgensen og Audun Fauchald Strand — Heis Podcast · podcast · 2026-01-19 source ↗ Reader 🖥️

practical · mixed — A research-grounded conversation. Brede Moe presents empirical findings from 27 years of SINTEF research on agile teams, autonomy, distributed work, and AI, but in service of prescriptive guidance to engineering leaders: how to build teams that learn under AI assistance. The hosts (Truls Jørgensen, Audun Fauchald Strand) prompt with practitioner questions; Brede Moe answers with research-shaped principles plus caveats. Argumentative in places (the AI paradox), narrative in places (the bilfabrikk study, the NAV Slack origin), expository on team mechanics.

Structure

Key terms

• AI-paradokset Code Was Never the Bottleneck
  The contradiction between individual perceived productivity gains from AI tools (9 of 10 feel more productive) and the organizational failure of AI initiatives (≈95% show no measurable bottom-line return, citing MIT). The deeper paradox: AI promises speed, but to capture the value at the organizational level, teams must slow down — invest in reflection, experimentation, joint practice. 'Du må gå saktere før du kan få til å løpe.' Months of investment, not weeks.
• autonomi Team Autonomy
  A team having (a) mandate/authority (myndighet), (b) competence to make good decisions, and (c) the right people present to make them — balanced against the principle that fewer team members is always better. Autonomy in his usage is constitutive of effective teams; it is not the same as individual freedom or org-design decentralization. It implies self-management, joint decision-making, and reflective practice.
• læringskultur Learning Culture & Slack
  An operational team practice — not a vague value — comprising: scheduled reflection time (e.g. every Friday), explicit show-and-tell sessions, workshops, dedicated Slack channels for sharing, joint discussion of tool use, and time set aside for experimentation. Distinct from individual learning. Requires psychological safety as substrate and slack (organizational room) as resource.
• psykologisk trygghet Psychological Safety
  The group-level condition where every member feels safe to admit mistakes, ask for feedback, and challenge how the work is done. Critically, it is collective — 'hvis ikke alle føler det, så er du ikke psykologisk trygger til stede.' Single dissonant members can poison it ('giftdråper'). Cites Google's Project Aristotle finding that this was the single factor distinguishing high-performing teams.
• enkelt- og dobbelkrekklæring (single- and double-loop learning) Learning Culture & Slack
  Single-loop learning = doing the same things better (incremental improvement). Double-loop learning = questioning whether you should be doing those things at all — examining assumptions, work processes, who you work with, whether you have the right problem. The latter is what distinguishes teams that actually learn; it requires psychological safety. (Argyris/Schön provenance, though Brede Moe does not name them.)
• augmented teams Human as Orchestrator of Agents
  His emerging research frame: teams that work with a digital assistant or AI as a team member, not just as individual tooling. The interesting question is whether AI can let teams be smaller while preserving autonomy, and how juniors are integrated when a data scientist plus an AI agent could replace headcount. Open question, not yet a finding.
• slakk Learning Culture & Slack
  Organizational room — time and capacity not consumed by delivery pressure — that must exist for learning, reflection, and experimentation to happen. Without slack the ildsjel (champion) burns out. Borrowed from systems thinking / DeMarco.
• ildsjel Learning Culture & Slack
  An individual within a team who burns for improvement, experimentation, or new tool adoption. Indispensable but insufficient — ildsjeler burn out if the broader culture and structure do not support them.
• augmentere / augmentet team Human as Orchestrator of Agents
  Same as 'augmented teams' above — a team whose work is shaped by collaboration with AI agents as quasi-members, not merely as individual tools each developer uses privately.

Propositions

• explicitdevelopedThere is a paradox in AI for software development: 9 of 10 individual users feel more productive with tools like Copilot, ChatGPT, and Cursor, yet ~95% of AI projects show no measurable bottom-line return.conf: high
  refines ←Measuring AI impact, assessing readiness, and new data trends · p5 — DX's longitudinal 10-15% throughput finding puts a measured middle between the 9-of-10-feel-productive euphoria and the 95%-no-return figure in Brede Moe's paradox.
• explicitdevelopedSoftware development is fundamentally a learning process; working code is a side-effect of that process.conf: high
  extends ←Presentation: Engineering at AI Speed: Lessons from the First Agentically Accelerated Software Project · p1 — Wolff takes 'software development is fundamentally a learning process' further: once implementation cost approaches zero, speed of learning becomes the only competitive advantage.
• explicitcorroboratedThe most effective teams are small; fewer people in a team is always better.conf: high
  extends ←How to Turn Individual Talent into Organizational Excellence · p3 — Provides a concrete, general (non-AI-specific) model of systematic junior/apprenticeship development — deliberate exposure to higher-level decisions, senior-as-teacher role expectations tied to status — that fills in mechanism the corpus's apprenticeship-under-AI cluster repeatedly asks for, though without addressing the AI-specific angle.
  supports ←Adam Mosseri: AI is a tailwind for authenticity · p6 — Mosseri: smaller teams move faster / less design-by-committee; Moe: the most effective teams are small, fewer people is always better.
• explicitCross-functional teams require deliberate investment in shared practice and psychological safety; this does not happen automatically.conf: high
• explicitestablishedPsychological safety is the single distinguishing factor of high-performing teams and a prerequisite for genuine learning.conf: high
  refines ←5 Things High-Performing Teams Do Differently · p1 — Moe names psychological safety as THE single distinguishing factor of high-performing teams; HBR widens that single factor into a broader relatedness/connection driver — of which psychological safety is one facet — and specifies the everyday behaviors through which it operates.
  supports ←How to Turn Individual Talent into Organizational Excellence · p4 — Both treat psychological safety as a structural, collective precondition for team performance and learning rather than an individual trait or nicety.
  extends ←Our Favorite Management Tips on Giving Feedback · p3 — The Heis proposition names psychological safety as the single distinguishing factor of high-performing teams; this document's 'Build a Culture That Invites Feedback' tip supplies a concrete, teachable behavioral lever (teaching specific asking, modeling from the top, rewarding asking, embedding it in routines) for building that safety around feedback specifically.
  supports ←How to Build a Superteam That Keeps Getting Better · p4 — Moe asserts psychological safety is the single distinguishing factor of high-performing teams; this document cites the same Google 2012 study and supplies a concrete behavioral mechanism Moe's piece asserts but does not operationalize — leaders pairing 'I'm not sure' with 'but I know who to ask,' which models admitted uncertainty and makes it safe for others to do the same.
  supports ←TBM 428: Yes, Robot. Yes, Boss? · p2 — Naming 'low voice safety' as a blocker AI cannot fix supports the claim that psychological safety is a prerequisite AI cannot manufacture.
• explicitPsychological safety is collective, not individual — if any team member does not feel safe, the team does not have it.conf: high
• explicitdevelopedAI threatens learning because it makes it easy to skip the colleague conversation and to ship code one does not understand.conf: high
  extends ←17) Jonas Slørdahl Skjærpe om svinghjuleffekter, liminalfaser og kunsten å ta heis · p12 — Skjaerpe generalizes the AI-threatens-learning claim into a learning-vs-production mode distinction: you cannot skip the learning behind production output because the model is built to make you feel smart.
  extends ←Trust Factory · p5 — Beck extends Brede Moe's warning that AI lets developers skip the colleague conversation into a four-vector account of how single-player augmented development erodes team trust.
• explicit⚡ disputed & corroboratedTeams that explicitly discuss how to use AI together, set aside time for reflection, and share experiences succeed better than those that leave AI use to individual practice.conf: high
  supports ←Nils Brede Moe's Post · p3 — Both claim that team-level shared practice, not individual tool use, is the real driver of success; heis-09 shows teams that jointly discuss AI use succeed better, this post frames the same finding as the slower, harder half of the paradox.
  contradicts ←The Orchestration Tax is You · p18 — Osmani frames agent-workflow discipline as an individual problem solved by personal workflow design; Brede Moe's evidence says teams that treat AI practice as joint team work outperform individual practice.
• explicitdevelopedSeniors and juniors benefit from AI tools differently — seniors can judge quality and impact and so derive genuine help; juniors lack that calibration and risk shipping code they cannot explain.conf: high
  refines ←Presentation: Leadership in AI-Assisted Engineering · p9 — Adds an adoption-speed dimension to the senior/junior asymmetry: juniors adopt AI faster, but seniors ultimately extract more value from it.
• explicitcorroboratedThe slow-down-to-speed-up requirement for AI adoption is measured in months, not weeks, especially in environments without continuous delivery infrastructure.conf: high
  supports ←Trust Factory · p7 — Beck's 'slowing augmented development is the path to actually going faster' converges with Brede Moe's months-long slow-down-to-speed-up requirement for AI adoption.
• explicitThe good practices of the continuous delivery era — automated tests, monitoring, observability, deployment pipelines — become more important under AI assistance, not less.conf: high
• explicitcorroboratedAI projects fail at organizational level largely because they are treated as technical or use-case problems rather than as organizational change projects requiring new workflows.conf: high
  supports ←Your AI strategy has a trust problem, not a tooling problem · p1 — Both locate AI's missing organizational returns in the surrounding org system rather than the tooling — Brede Moe blames missing joint practice change, Elena blames permission and trust structures.
• explicitStructured, scheduled learning practices (Friday reflection meetings, workshops, dedicated Slack channels) outperform reliance on individual champions (ildsjeler) because champions burn out without supporting culture.conf: high
• explicitcorroboratedTelling teams to be '20% more productive' fails; telling teams to 'learn how to use this well' wins (citing a car-factory study).conf: medium
  supports ←Nils Brede Moe's Post · p7 — Echoes the car-factory finding cited in heis-09 (the 'learn to use it well' group beat the '20% more productive' group): the payoff comes from the team learning a new way of working together, not from faster raw output.
• explicitcorroboratedWorking agreements that originate from peer agreement among team members are stronger control mechanisms than rules imposed by management.conf: high
  supports ←New Skills to Navigate Continuous Change · p2 — Same claim from different domains: peer-agreed, co-decided norms persist and control behavior better than externally imposed rules.
• explicitHybrid meetings (some physical, some remote) are worse than either fully co-located or fully remote meetings.conf: high
• explicitAI agents may make teams smaller without making them less effective, but if every member just works alongside their own AI it ceases to be a team.conf: high
• explicitcorroboratedThe conversational tone of AI agents (constant praise, 'you're brilliant') creates a misleading sense of competence that substitutes for collegial calibration.conf: high
  supports ←Research: Why You Shouldn't Treat AI Agents Like Employees · p7 — Experimental evidence that AI-as-colleague framing lowers manager scrutiny (18% fewer errors caught) backs the podcast claim that AI's social/agreeable tone substitutes for real collegial calibration.
• explicitIndividuals tightly supervising AI output (Audun: 'I keep Claude on a tight leash, massive tests, guard rails') experience AI as comprehension-enhancing, not comprehension-degrading.conf: high
• explicitSoftware development practices documented in books from the 1960s still apply; AI changes the daily process but not the underlying craft.conf: medium
• implicitestablishedProductivity gains from AI accrue to individuals on individual tasks but do not translate to organizational gains without joint practice change.conf: high
  supports ←Nils Brede Moe's Post · p3 — Makes explicit, in accessible LinkedIn language, what heis-09 states only implicitly: individual AI adoption is fast, but proportional organizational gains require team-level joint practice change.
  supports ←Software Development As Collective Learning With Hazel Weakly · p1 — Both make the same structural claim in different domains: individual-level gains (a force multiplier's personal effectiveness; AI-driven individual productivity) do not translate into durable organizational gains without a change to the system/practice itself, and vanish once the individual factor (the person, or unaccompanied individual AI use) is removed.
  supports ←11) AI-spesial med Vilde Opsal og Anders Sveen · p14 — The panel explicitly cites Brede Moe in affirming that individual-level AI gains are well-attested while team-level augmentation patterns remain unsolved.
• explicitThe current AI wave, unlike blockchain, is genuinely transformative — even after the hype subsides, something will remain that changes how we think about technology.conf: high
• explicitMestring (mastery of one's work and private life) is the real driver of well-being in hybrid work, not the specific in-office/remote ratio.conf: high

Arguments

Narratives

• Brede Moe's career arc: 27 years at SINTEF, started as developer realizing he was not the best, moved to research on team dynamics in software development, did his doctorate on autonomous teams when 'no one thought it was fun' — took 5-8 years before the field caught up.
  → Establishes credibility and signals that he has been ahead of the curve before; subtly suggests current AI-and-teams research will follow the same delayed-recognition pattern.
• Cross-functional team around vulnerable children referenced from Smidigpodden — doctors, psychologists, physiotherapists who agreed on the goal but still struggled to make the team work.
  → Illustrates that even with shared mission and motivation, cross-functional teamwork is hard; supports the proposition that tverrfaglig teams require deliberate investment.
• Recent interview with a junior developer who admitted he didn't fully understand the code he had written with AI assistance — he 'should have spent more time understanding it, but I had to ship.'
  → Concrete instantiation of the AI-as-threat-to-learning argument; makes the abstract risk tangible.
• Audun's own practice: keeping Claude Code on a tight leash with extensive tests and guardrails, claiming it makes him understand the codebase better, not worse.
  → Counter-example that complicates the threat narrative — Brede Moe immediately codes this as a senior practice, reinforcing the senior/junior asymmetry argument.
• The car factory study (15-20 years ago): one division told to be 20% more productive, another told to 'learn how to use this well.' The learning group won.
  → Empirical anchor for the slow-down-to-speed-up argument; positions current AI exhortations as repeating a known failure pattern.
• Brede Moe's BPM/SOA dissertation history: spent a year afterwards 'unlearning' the framework once he left the research bubble — admits the industry is fad-driven and that he has been a participant in fads.
  → Epistemic humility — establishes Brede Moe as someone who can distinguish hype from substance because he has fallen for hype before.
• Pandemic distributed-work research: when COVID hit, his team already had years of data from Ericsson, Spotify, and Norwegian companies on virtual teams — 'lucky' to have findings ready.
  → Demonstrates that the team-level research lens generalizes across contexts; primes credibility for current AI-and-teams claims.
• NAV Slack origin story: started in 2017 with five people; #general full history visible; when newcomers complained about excessive joking, they tried splitting into 'serious' and 'joke' channels, but developers and designers always migrated to the joke channel and re-merged the 50/50 mix. This happened repeatedly.
  → Illustrates that learning culture is emergent — top-down structural attempts to separate work from play failed because the actual culture is the mix. Implicitly supports the argument that working agreements must come from teams themselves.
• A colleague set up multiple AI agents to collaborate on a problem and recorded the conversation between them; Brede Moe found it 'both fascinating and very scary.'
  → Concretizes the augmented teams research direction with an example he is wrestling with; openness about not yet knowing what to make of it.

Questions addressed

• ✅ What is the AI paradox, and why do organizational AI initiatives fail despite individual productivity gains?
  Because software development is fundamentally a learning process, and AI accelerates output while threatening the learning. Without team-level practice change, slack, and psychological safety, individual gains do not aggregate to organizational gains.
• ✅ What does effective team autonomy look like?
  Small teams with mandate, competence, and the right people for decisions; balanced against the principle of having as few people as possible; supported by psychological safety and time for reflection.
• ✅ How should teams adapt their practices for AI?
  Scheduled reflection (e.g. Friday learning meetings), explicit show-and-tell, workshops, dedicated Slack channels, robust delivery/test/monitoring infrastructure, and new specification practices. Recognize that AI adoption is an organizational change process, not a tool rollout.
• ⭕ What becomes of teams when AI is in the loop?
  Open research question. Augmented teams might be smaller; AI agents might be quasi-members; but the autonomy implications and especially junior development pathways are unresolved.
• ✅ How does the senior/junior dynamic change under AI?
  Seniors gain genuine help because they can judge quality and impact. Juniors lack calibration and risk shipping code they cannot explain. The team's calibration-mediating role becomes more important.
• ⭕ How should AI agents' conversational tone (constant flattery) be managed?
  Raised by Truls; flagged as a concern (substitutes flattery for collegial calibration); not resolved.
• ✅ What hybrid work setup actually works?
  There is no universal answer; teams must agree among themselves. Hybrid meetings are the worst case; default-digital meeting start helps; mastery on both work and home sides predicts well-being more than ratio does.
• ✅ Is leadership style relevant to building learning culture?
  Yes — horizontal, peer-oriented leadership ('I think of everyone as horizontal') works better than hierarchical control. Brede Moe demonstrates this in his own group: no fixed RTO rule, but high office attendance because the culture is fun.

Gaps

• incompleteThe MIT '95% of AI projects show no return' figure is asserted as foundational but never sourced, methodologically interrogated, or compared to base rates for non-AI IT projects (where failure rates are also famously high). The framing relies heavily on this number.
  Read next: Pursue the MIT study directly (likely the MIT NANDA / Sloan AI 'GenAI Divide' report from 2025) and compare against Standish CHAOS-style baselines for IT project failure to establish whether AI is failing more than other IT initiatives, or merely the same.
• incompleteThe framing is almost entirely about knowledge-work software teams in professional services / large org settings. The argument is silent on: solo developers, small startups without team infrastructure, contracting/outsourcing relationships, non-developer roles where AI is also disrupting work, and contexts where 'team' is not the unit of production.
  Read next: Read Tabarrok / Acemoglu on AI and labor markets in non-team-shaped work; and Beane's Skill Code on apprenticeship in mixed expert/AI environments outside software, to test how portable Brede Moe's prescriptions are.
• incompleteThe senior/junior asymmetry is identified but the prescription for what to do about it stops at 'build safety so juniors can admit confusion.' There is no concrete model of how junior skill formation actually works in an AI-augmented team — what scaffolding, what kinds of exercises, what the new apprenticeship looks like.
  Read next: Beane's Precision Proactivity (extracted this session) directly addresses this — proactive skill formation under AI assistance. Brede Moe's empirical team lens and Beane's apprenticeship lens together would be more complete than either alone.
  closed byHow to Turn Individual Talent into Organizational Excellence — Supplies a concrete (though not AI-specific) systemic apprenticeship model: senior leaders are given formal role expectations and time allocation to teach; juniors are deliberately exposed to decisions two to three levels above their formal role; feedback is frequent, specific, and tied to real decisions; teaching is systematized via status and promotion rather than left to individual personality. This answers the general 'what scaffolding' question the gap raised, but does not address what changes when AI removes the on-ramp tasks juniors traditionally learned on — that AI-specific angle remains open.
• incompleteBrede Moe's account of læringskultur emphasizes practices internal to a team (reflection, show-and-tell, Slack channels). It is silent on inter-team learning at scale — how lessons from one team's AI experimentation propagate across an organization of dozens or hundreds of teams. This is precisely where the MIT 95% gap might be most operationally relevant.
  Read next: Cutler on institutional overload and operating systems; SINTEF's own work on large-scale agile (Brede Moe has co-authored on this); Lean enterprise / Forsgren-Humble-Kim on horizontal practice diffusion.
  closed by11) AI-spesial med Vilde Opsal og Anders Sveen — Describes a working organization-scale propagation pattern for AI learning across teams: surface an ordinary team that succeeded, run fast and safe Slack channels, workshops, internal blog summaries, and a no-opt-out learning culture.
• incompleteThe conversation treats AI tooling as relatively homogeneous — Copilot, ChatGPT, Cursor, Claude Code are mentioned as a cluster. But the team-level practices likely differ between chat-based assistants, autonomous agents, and in-IDE pair tools. Brede Moe gestures at this with 'augmented teams' but doesn't yet differentiate.
  Read next: Empirical studies of differential effects by tool class — DORA's 2025 AI findings, Microsoft / GitHub research on Copilot effects, ongoing studies of agentic coding tools.
• incompleteThe car-factory study (cited as 15-20 years old) is the central empirical anchor for the slow-down-to-speed-up claim, but it is cited without source — the listener cannot evaluate the comparison. Similarly the 9-of-10 figure.
  Read next: Source the original study; even if accurate, ask whether software development resembles auto-manufacturing learning curves enough for the lesson to transfer.
• incompleteThe argument assumes that team autonomy and team learning are the right units of analysis. It does not engage with alternative views — e.g. that AI might shift the unit back to the individual contractor or fragment work into smaller, more decomposable tasks, in which case the team-level prescription is the wrong response to the wrong question.
  Read next: Read agentic-work literature (Eric Ries, Naval-style 'agency leverage' arguments, Yegge's vibe-coding manifesto) for the opposing view.
• incompletePsychological safety is treated as foundational and necessary, but the conversation does not address how to actually build it in distributed, hybrid, and AI-augmented configurations — only that it is harder there. This is the most-cited construct in the conversation and the least operationalized.
  Read next: Edmondson's recent work on hybrid safety; Brede Moe's own SINTEF publications on distributed-team trust formation.
  closed byPresentation: The Human Scalability Problem: Why Your Teams Don't Scale Like Your Code — Supplies concrete mechanisms for building the safety Brede Moe treats as foundational but unoperationalized: relationship-first team formation, leaders modeling vulnerability, rituals that persist across changing team composition, and direct handling of fear.
• uninformedThe hosts and Brede Moe do not engage with the strong critique that AI agents are not yet reliable enough to be modeled as 'team members' in the augmented-team sense — they neither hold state, nor carry context across sessions in the way a colleague does, nor have stakes. The augmented-teams framing risks anthropomorphizing.
  Read next: Read critics like Gary Marcus, Emily Bender, and Murray Shanahan on the gap between LLM behavior and team-member competence.
  closed byResearch: Why You Shouldn't Treat AI Agents Like Employees — Engages the anthropomorphizing critique the podcast skipped, with experimental evidence: framing AI as a team member/employee measurably reduces human accountability and error-catching, supporting the case against modeling agents as colleagues (p1-p3, p7).

Closes gaps elsewhere

• 5) Vidar Moe: Det eneste Vidar vil er å ha det lett · g1 — Supplies the missing AI chapter for Moe's 'software development is structured learning' framework: AI threatens the learning act by displacing colleague interaction, and the mitigation is making AI use a collective, reflective team practice under tight supervision.
• 8) Christian Printzell Halvorsen om å lede på riktig nivå: Lokalt forankret, nordisk samlet · g4 — Operationalizes the psychological-safety premise behind Vend's self-reported status mechanism: safety is collective, requires deliberate investment in shared practice, and is built through structured learning rituals and peer-agreed working agreements.

Related documents

• Measuring AI impact, assessing readiness, and new data trends
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness Human as Orchestrator of Agents
• Finding Comfort in the Uncertainty
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness Human as Orchestrator of Agents
• Presentation: Leadership in AI-Assisted Engineering
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness Psychological Safety
• Your AI strategy has a trust problem, not a tooling problem
  Enterprise AI Adoption & Readiness Human as Orchestrator of Agents Team Autonomy
• Managers Are Not Overhead: They Are Infrastructure
  Enterprise AI Adoption & Readiness Human as Orchestrator of Agents Psychological Safety
• New Skills to Navigate Continuous Change
  Learning Culture & Slack Psychological Safety Team Autonomy
• Ordinary Engineers, Not Heroic Inventors
  Code Was Never the Bottleneck Enterprise AI Adoption & Readiness Learning Culture & Slack
• Mob programming: From avant-garde experimentation to established practice
  Learning Culture & Slack Psychological Safety Team Autonomy
• Research: Why You Shouldn't Treat AI Agents Like Employees
  Enterprise AI Adoption & Readiness Human as Orchestrator of Agents
• Nils Brede Moe's Post
  Code Was Never the Bottleneck Learning Culture & Slack