Nils Brede Moe's Post

Nils Brede Moe · article · 2026-09-23 Reader 🖥️

Code Was Never the Bottleneck Guardrails Learning Culture & Slack Spec-Driven Development Technology Diffusion & Transformation Curves

Teams that genuinely pursue value from AI-assisted coding may work slower at first, because AI shortens code-writing but does not shorten the harder, team-level work of learning to specify tasks, verify quality, and build a shared practice — so the real productivity gain comes not from AI writing code faster but from the team developing a new, common way of working.

theoretical · expository — A 1-minute LinkedIn summary of a research finding, with a practical implication attached at the end. Reports a SINTEF Digital finding (why AI-accelerated coding doesn't translate into faster teams) and frames it as a paradox, but presents no methodology or data of its own — it points to a JavaZone talk and a kode24 article as the fuller sources. Signed 'Silje Rognsvåg' at the bottom despite being posted under Nils Brede Moe's byline, suggesting a communications co-author or a shared/repost account; treated here as Moe's post per the source metadata.

Structure

Part	Role
Opening hook: AI writes code fast, so why don't teams go faster?	Poses the question the rest of the post answers
First explanation: coding is only one part of the development process; when it speeds up, work shifts to task specification and quality verification	First-order answer to the hook question
SINTEF Digital research finding: it takes time to move from individual AI use to a shared team practice	Introduces the second, deeper explanation, credentialed by ongoing SINTEF research
Elaboration: individuals adapt fast on their own; the harder work is the team jointly working out process changes, task specification, guardrails/verification, when agents add value, and where to experiment vs. be careful	Unpacks what 'the harder work' concretely consists of
Requirement: this demands the team experiment, learn together, and actually change how it works	States the mechanism by which the gap gets closed
Paradox framing: teams genuinely pursuing AI value may work more slowly first	Names the paradox explicitly, turning the explanation into a memorable claim
Anchor in a real event: the paradox as discussed with Jan Henrik Gundelsby (Knowit) at JavaZone	Grounds the claim in a specific talk/conversation rather than leaving it abstract
Closing distinction: the gain doesn't come from AI writing code faster, but from the team developing a new shared way of working	Restates the unity as the post's final, most quotable line
Pointer to kode24 coverage of the talk and research, funding acknowledgment (Norges forskningsråd), and co-researcher credits (Darja Smite, Nina Haugland Andersen, Anastasiia Tkalich)	Directs the reader to the fuller source and establishes the research's institutional backing

Key terms

felles praksis (shared/common practice)
The team-level, jointly-worked-out way of using AI agents — as distinct from the sum of individual team members' own habits. Includes agreed task-specification conventions, guardrails, verification routines, and shared judgment about when agents add value. Something the team has to build together, not something that emerges automatically once individuals are proficient.
paradokset (the paradox)
That teams seriously trying to extract value from AI may initially slow down, because the visible speed gain (faster code) is not where the real bottleneck or the real payoff sits — those are in specification, verification, and shared practice formation.

Propositions

explicitAI accelerates code writing, but teams do not get proportionally faster, because writing code is only one part of the development process.conf: high
explicitWhen AI makes coding faster, more time goes into specifying tasks so agents understand what should be done, and into verifying that the result holds the right quality.conf: high
explicitIt takes time to move from individual AI use to a shared practice within the team.conf: high
supports9) Nils Brede Moe om AI-paradokset, autonomi, læringskultur og fremtidens team · p8 — 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.
supports9) Nils Brede Moe om AI-paradokset, autonomi, læringskultur og fremtidens team · p21 — 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.
explicitIndividuals find their own ways of using AI fairly quickly; the harder work begins when the team must jointly work out how work processes should change, how to specify good tasks for agents, what guardrails and verification are needed, when agents actually add value, and where to experiment versus where to be careful.conf: high
explicitClosing this gap requires the team to experiment, learn together, and actually change how it works.conf: high
explicitTeams that genuinely try to extract value from AI can actually work more slowly at first.conf: high
explicitThe productivity gain from AI does not come only from AI writing code faster; it comes when the team develops a new, shared way of working.conf: high
supports9) Nils Brede Moe om AI-paradokset, autonomi, læringskultur og fremtidens team · p14 — 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.
implicitTeam-level shared-practice formation is inherently slower to achieve than individual habit formation with a new AI tool.conf: medium

Arguments

Teams that genuinely pursue value from AI-assisted coding may work slower at first, because the real productivity gain requires the team to develop a new shared way of working, not just faster code generation. strength: medium

Explicit premises:

AI accelerates code writing, but coding is only one part of the development process (p1).
Faster coding shifts work toward task specification and quality verification (p2).
Moving from individual AI use to shared team practice takes time (p3).
The team-level work — process change, task specification, guardrails, judging agent value, deciding where to experiment — is harder and slower than individual adoption (p4).
Closing the gap requires the team to experiment and learn together (p5).

Implicit premises:

Team-level practice change is necessarily slower than individual habit formation.
Organizational/team AI value is gated by the team's slowest-forming collective practice, not by individual tool proficiency.

Inferences:

Leaders and teams should expect and budget for an initial slowdown as a sign of genuine engagement with AI, not as a failure signal.
Measuring AI's value purely by code-writing speed will miss where the real bottleneck — and the real payoff — actually sit.

Narratives

The paradox as a topic Nils Brede Moe and Jan Henrik Gundelsby (Knowit) discussed together on stage at JavaZone.
→ Anchors an otherwise abstract claim in a specific, named, real-world event and a named interlocutor from industry (Knowit), lending it a practitioner-plus-researcher framing rather than a purely academic one.

Questions addressed

✅ Why doesn't AI's coding speed translate into proportionally faster teams?
Because code-writing is only one part of development; the freed-up time shifts to task specification and quality verification, and because moving from individual AI use to a shared team practice — process change, specification conventions, guardrails, and judgment about when agents help — takes deliberate joint work, not just faster tooling.

Gaps

incompleteThis is a 1-minute LinkedIn teaser for deeper SINTEF empirical research; no methodology, sample size, or data are given for either the specification/verification time-shift claim or the individual-to-shared-practice claim. The post only links onward to a kode24 article and gestures at a JavaZone talk.
Read next: Read the linked kode24 article (https://lnkd.in/eKhzYCCt) covering the JavaZone talk, and cross-reference the fuller SINTEF findings already captured in this corpus at extract-heis-09-nils-brede-moe-ai-paradokset-autonomi-laringskultur.json — same researcher, same 'AI-paradokset' framing, with sourced propositions, arguments, and narratives (including the 9-of-10-vs-95% MIT figure and the car-factory study) that this short post compresses away.
incompleteThe post poses "når gir agentene faktisk verdi – og når gjør de ikke det?" (when do agents actually add value, and when not) as one of the open team-level questions, but does not answer it — no criteria, examples, or decision heuristic are offered.
Read next: Cross-reference corpus docs tagged measuring-ai-impact for candidate criteria (e.g. extract-dx-measuring-ai-impact.json, extract-8-engineering-metrics-ai-broke.json) to see whether any propose concrete when-agents-help heuristics this post leaves open.