Tet Digital CTO Morisbak Links AI Adoption to Increased Staff Oversight

For God's sake, keep the nerds! - Stein Inge Morisbak - NDC AI 2026 — NDC Conferences

NDC Conferences

Generative AI functions as an amplifier that makes deep technical expertise and rigorous human verification more critical to enterprise stability than ever before.

AI-assisted summary of For God's sake, keep the nerds! - Stein Inge Morisbak - NDC AI 2026 — original source. Check the original for full context.

Source published:

Key ideas

  • A six-week pilot of Claude Code across 180 developers at Tet Digital resulted in 90% of engineers reporting increased productivity. (0:06)
  • Productivity gains for developers do not always translate to faster delivery, as coding speed is often not the actual bottleneck in complex organizational environments. (9:12)
  • Tet Digital implemented a mandatory three-step rollout process (training, team agreement, and signed declaration) rather than an unconstrained release. (15:22)
  • AI functions as an amplifier: it elevates the performance of already disciplined, high-performing teams while exacerbating existing dysfunctions in struggling teams. (4:33)

Detailed notes

## Background This talk, titled "For God's sake, keep the nerds!" and presented by CTO Stein Inge Morisbak, was published on the NDC Conferences YouTube channel on July 1, 2026. The presentation explores the impact of AI on software development teams, detailing a six-week pilot program at Tet Digital and arguing for a deliberate, guardrail-focused approach to adopting AI coding assistants.

## TL;DR Stein Inge Morisbak argues that AI serves as an amplifier for engineering teams, magnifying both high performance and existing dysfunctions. Through a pilot study involving 180 developers, he demonstrates that while AI significantly boosts subjective productivity, it does not solve core organizational bottlenecks like meetings or complex dependencies. Ultimately, he posits that as AI handles routine syntax, the human role shifts toward quality control, orchestration, and critical verification, making deep technical expertise more essential than ever.

## Atmosphere & style The presentation carries the candid, weary, and slightly humorous tone of a seasoned tech leader addressing his peers at the end of a long conference day. Morisbak balances executive-level pragmatism with a grounded, "in the trenches" perspective, using lighthearted imagery—like his AI-generated gymnastics photo—to contrast with the serious risks of enterprise system failure. The overall feeling is one of measured, cautious optimism that prioritizes human craftsmanship over the allure of automated speed.

## Bulletin - [00:06](https://www.youtube.com/watch?v=Fda_-l6HB18&t=6s) A six-week pilot of Claude Code across 180 developers at Tet Digital resulted in 90% of engineers reporting increased productivity. - [09:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=552s) Productivity gains for developers do not always translate to faster delivery, as coding speed is often not the actual bottleneck in complex organizational environments. - [15:22](https://www.youtube.com/watch?v=Fda_-l6HB18&t=922s) Tet Digital implemented a mandatory three-step rollout process (training, team agreement, and signed declaration) rather than an unconstrained release. - [04:33](https://www.youtube.com/watch?v=Fda_-l6HB18&t=273s) AI functions as an amplifier: it elevates the performance of already disciplined, high-performing teams while exacerbating existing dysfunctions in struggling teams.

## Key takeaways - [04:33](https://www.youtube.com/watch?v=Fda_-l6HB18&t=273s) AI is best understood as an amplifier of existing organizational practices, not as a replacement for foundational technical skill. - [01:17](https://www.youtube.com/watch?v=Fda_-l6HB18&t=77s) Deep technical understanding is the essential insurance policy for preventing system collapse when AI generates code at high speed. - [11:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=672s) Human oversight is non-negotiable; engineers must shift their role from manual coders to orchestrators who verify and critique AI-generated output. - [09:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=552s) While developers report high productivity, there is a perceptual gap where product leads may see little to no change in actual shipping velocity or quality. - [11:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=672s) Generative AI can pose a threat to junior skill acquisition if senior engineers do not prioritize active mentorship and manual code reviews. - [00:06](https://www.youtube.com/watch?v=Fda_-l6HB18&t=6s) Security must be built directly into the agent’s workflow via hooks and standards rather than relying solely on developers catching errors manually. - [17:04](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1024s) Tet Digital’s security and usage policy was distilled into three simple rules: Human Responsibility, Confidentiality, and Using Your Head. - [11:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=672s) Tooling maturation takes time; observed productivity gains increased from 83% to 90% as engineers adapted to the new workflow during the pilot. - [09:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=552s) The bottleneck for delivery is often found in meetings, team dependencies, and prioritization, which are areas where AI coding assistants provide no solution. - [17:54](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1074s) AI adoption requires a dedicated enablement team to handle training, monitor failure modes, and manage organizational guardrails. - [18:56](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1136s) Expanding AI to non-technical staff like designers requires caution, as they lack the foundational technical skills necessary to safely manage the generated output. - [20:54](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1254s) Success in AI adoption is driven by effective change management and team culture, not simply by purchasing software licenses. - [11:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=672s) The role of the developer is evolving to prioritize quality control and the design of practices that catch AI mistakes before they reach production. - [06:17](https://www.youtube.com/watch?v=Fda_-l6HB18&t=377s) AI-generated code quality can be inconsistent, with 10% of pilot participants reporting a decline in output quality, justifying the need for strict guardrails.

## Detailed walkthrough ### [00:06](https://www.youtube.com/watch?v=Fda_-l6HB18&t=6s) Mobility at Scale: Who Tet Digital Is

Taking the stage for the final session of the day at NDC AI in Oslo, Stein Inge Morisbak acknowledges the audience's fatigue, joking that he is the only thing standing between them and the evening's party and beers. Introducing himself as the Chief Technology Officer of Tet Digital, he provides context on what the company does and why its infrastructure matters. Tet Digital is a technology company developing public mobility solutions across Norway. It is jointly owned by the country’s largest regional public transport authorities: Ruter in Oslo, Skyss in Vestland, Agder Kollektivtrafikk, AtB in Trøndelag, and Østfold Kollektivtrafikk. Today, fully 69% of all public transport journeys throughout Norway run directly through Tet Digital's systems, meaning stability and correctness in their codebase are critical public infrastructure concerns.

[01:17](https://www.youtube.com/watch?v=Fda_-l6HB18&t=77s) Executive Hype vs. Developer Reality: The Insurance Policy

The media and executive landscape is flooded with euphoric claims regarding artificial intelligence. Stein Inge shares headlines from the Norwegian newspaper *VG*—featuring his own English translations placed onto the slides via AI—highlighting statements made by senior executives from prominent companies including Telenor, Schibsted, DNB, and Microsoft. While corporate leaders project massive productivity windfalls and celebrate the notion that technical barriers are evaporating, developers on the ground push back with legitimate caution regarding mounting technical debt and the surrender of code ownership. This caution is not mere pessimism or resistance to change; it is an act of engineering stewardship.

> “The developers, they aren't being negative, they are being responsible, because when AI produces code at enormous speed, deep technical understanding is the only insurance you have against total system collapse.” — Stein Inge Morisbak

- Executive optimism contrasts sharply with the lived realities of maintaining complex enterprise codebases. - The tension is not purely top-down: high-performing engineers also clamor for cutting-edge tools, fearing they will be left behind if held back by cautious leadership.

[04:33](https://www.youtube.com/watch?v=Fda_-l6HB18&t=273s) The AI Amplifier and Doing the Splits

Stein Inge invokes DORA’s research on software teams to define the core operational dynamic: AI functions strictly as an amplifier. It magnifies an organization's existing baseline—elevating the strengths of high-performing, well-disciplined teams while violently exacerbating the dysfunction of teams that are already struggling. Displaying an AI-generated image of himself wearing a festive Christmas sweater while physically stretched in a gymnastics split, he explains that it captures his exact role as CTO when kicking off their pilot just before the holidays. He is not split between management and engineers, but rather balanced between a shared organizational ambition to move faster and a shared responsibility to avoid driving the company over a cliff.

```mermaid flowchart TD accTitle: The DORA AI Amplifier Effect accDescr: Diagram illustrating how AI amplifies both high-performing teams and struggling engineering teams.

A["AI Tooling Adoption"] --> B["Amplifier Effect"]

subgraph Outcomes ["Organizational Outcomes"] B -->|"Strong practices & discipline"| C["High-Performing Teams<br/>(Strengths Amplified)"] B -->|"Poor practices & debt"| D["Struggling Teams<br/>(Problems Exacerbated)"] end ```

To resolve this tension, Tet Digital targeted their core asset: their code and the developers who work directly at the production boundary. Mistakes made here propagate immediately to end-users and represent the most expensive failures to remediate. Establishing robust guardrails at this critical juncture creates a safe foundation that the rest of the organization can build upon.

[06:17](https://www.youtube.com/watch?v=Fda_-l6HB18&t=377s) The Pilot Design: Scope, Productivity, and Quality

Tet Digital structured a six-week pilot utilizing Claude Code across 6 teams and 30 engineers, examining three core vectors: the impact on people and daily work, the concrete utility of Claude Code, and the requisite safety guardrails. In practice, Claude Code excelled at large-scale, repetitive tasks: bulk refactoring, test generation, documentation, and debugging. Conversely, it performed poorly on highly complex assignments tangled up in intricate cross-system and cross-team dependencies typical of enterprise mobility architectures.

Measuring the human and technical outcomes yielded clear shifts across the six weeks: - **Productivity:** A striking 90% of participating developers reported an increase in productivity, 10% observed no change, and zero developers reported a decline. - **Wellbeing:** 60% experienced improved wellbeing, while 40% saw no change. Significantly, multiple engineers had initially reported lower wellbeing at the outset before finding their footing. - **Mastery:** 65% reported a stronger sense of personal mastery, 25% saw no change, and 10% noted a decline. - **Code Quality:** Results were split. While 45% reported higher quality and 45% reported no change, 10% reported a distinct decline in quality—a statistic that underscores why rigid guardrails remain non-negotiable.

[09:12](https://www.youtube.com/watch?v=Fda_-l6HB18&t=552s) The Delivery Paradox: Writing Code Isn't the Bottleneck

When product leads were surveyed to determine if they observed the same productivity leap reported by their developers, the results revealed a stark perceptual gap. Product managers reported no noticeable change in overall delivery speed and very little difference in code quality, though they conceded that team quality was already quite high. What product leads did observe was simply that their developers felt significantly more productive and were visibly enjoying their day-to-day work.

This disconnect carried over into delivery confidence. Despite 90% of developers feeling more productive, many remained uncertain about whether their teams would hit their quarterly OKRs on schedule. Stein Inge emphasizes that this apparent contradiction exposes a fundamental truth about software engineering: producing lines of code is rarely the actual bottleneck in delivery. The actual delays in modern organizations stem from endless meetings, complex inter-team dependencies, ambiguous priority setting, and sluggish decision-making—none of which an AI coding assistant can solve.

[11:20](https://www.youtube.com/watch?v=Fda_-l6HB18&t=680s) Four Hard Lessons: Why We Keep the Nerds

Reflecting on the deeper takeaways from the pilot data, Stein Inge outlines four lessons that require human engineering expertise rather than algorithmic automation:

1. **AI makes mistakes:** Because generative models inevitably introduce errors, team discipline, code review practices, and verification processes matter more than ever. Rolling AI out to a team that already struggles does not fix its foundational issues; it compounds them. 2. **Reduced learning for juniors:** Heavy reliance on AI generation poses a serious threat to junior skill acquisition if senior engineers do not actively mentor them. Organizations must be deliberate about pairing, rigorous code reviews, and structured study to ensure foundational engineering skills still develop. 3. **Task transformation:** The engineer's day-to-day role is shifting away from manual syntax typing toward orchestrating, evaluating, and critically verifying AI outputs. Hands-on coding skill is no longer just the job itself; it is the prerequisite judgment required to determine if the generated output is safe to deploy. 4. **Tool maturation takes time:** Measuring developers at multiple intervals showed maturation in real time. Early on, 83% reported higher productivity, 13% saw no change, and 4% reported a decline. By the close of the pilot, that 4% decline dropped to zero, while higher productivity climbed to 90%.

> “People who can design the practices that catch AI mistakes, people whose technical debt doesn't erode, people who can verify AI output because they can read it, people who can drive maturation—and this is exactly the reason why we keep the nerds.” — Stein Inge Morisbak

[15:22](https://www.youtube.com/watch?v=Fda_-l6HB18&t=922s) A Deliberate Three-Step Rollout and Three Simple Rules

When expanding Claude Code from the 30-person pilot group to all 180 developers across the company, Tet Digital deliberately rejected an unconstrained "open tap" approach. Access was gated behind a formal three-step onboarding framework:

1. **Mandatory Training:** A targeted session centered on AI as an amplifier, security boundaries, and practical usage patterns (leaving specific tool syntax and prompting techniques to developer communities of practice). 2. **Team Leadership Agreement:** A team-wide consensus where leaders and engineers mutually commit to adopting the tool, acknowledging that the entire team shares collective responsibility for business and architectural outcomes. 3. **Signed Declaration:** An explicit confirmation that the developer has read, understood, and agreed to adhere to organizational usage guidelines.

Rather than authoring an unreadable corporate document, Stein Inge and his team distilled their entire security and quality policy into three memorable directives designed to fit comfortably in a single Slack message:

- **Rule 1: Human Responsibility.** You remain fully responsible for every single line of code generated by an AI; human review and output verification are strictly mandatory. - **Rule 2: Confidentiality.** Never input sensitive data, personal identifiable information (PII), credentials, secrets, or business-confidential logic into the model. - **Rule 3: Use Your Head.** If you are ever uncertain whether an action aligns with policy, pause and ask for guidance rather than guessing.

### [17:54](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1074s) Establishing the AI Enablement Team Stein Inge Morisbak explains that Tet Digital has formed a dedicated AI enablement team. The primary responsibilities of this team include conducting training, maintaining organizational guidelines, providing hands-on assistance to teams that get stuck, and monitoring for the specific failure modes identified during the pilot phase. Measuring the success of these initiatives remains a complex challenge. While the company has implemented necessary guardrails, Morisbak notes that determining whether productivity is truly increasing—rather than just relying on the subjective feelings of the staff—remains a persistent struggle.

* The goal is to ensure that AI adoption remains a "deliberate, not accidental" process. * The team acts as a central hub for measuring organizational performance across various dimensions.

### [18:57](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1137s) Expanding the Scope: Designers and Product Managers Beyond software developers, the company is now testing AI tools with designers and product managers. This shift highlights a critical tension: these new user groups are often uncomfortable with technical environments, such as the command-line interface or terminal, which they find intimidating. Morisbak argues that this discomfort reinforces the necessity of having technical experts available to act as guardrails. If non-technical groups are to use tools like Claude Code effectively, the company must first ensure they possess a foundational understanding of developer skills.

* "AI doesn't remove the need for technical debt just shows you where the debt is missing," according to Morisbak. * The pilot phase is transitioning from a small test to a company-wide initiative, which requires a focus on sustainable, long-term value rather than abstract, all-encompassing strategy.

### [20:56](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1256s) Scaling from Pilot to Practice To scale successfully, Tet Digital is maintaining a controlled and responsible approach. The AI enablement team remains the bedrock for support, and the company views legal, security, and ethics considerations as absolute foundations. Morisbak emphasizes the importance of community-of-practice groups, suggesting that in this era of rapid change, everyone has essentially become a "newbie." He strongly encourages industry collaboration and sharing learnings between companies and academia to keep pace with the technology.

* "Value isn't created when you buy licenses," Morisbak observes, noting that success is fundamentally about effective change management rather than the tooling itself. * The human element is paramount; the expertise of the people doing the work remains more vital today than it ever has been.

### [22:54](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1374s) Final Takeaways and the "Nerds" In his closing remarks, Morisbak urges the audience to start experimenting with AI, emphasizing that "you can't learn when you're standing on the sidelines." However, he underscores that this must be done with discipline and proper oversight. He reiterates his central thesis with a final, firm plea to the audience: "And remember, for God's sake, keep the nerds." He concludes by clarifying that rather than eliminating the need for experts, AI actually makes their specialized knowledge and technical judgment more important than ever before.

* AI is framed not as a replacement for human skill, but as an amplifier that necessitates even higher levels of expertise to govern and verify the output. * The talk concludes with an invitation for further discussion, as Morisbak stays on to field questions from the conference attendees.

## Notable quotes > [01:17](https://www.youtube.com/watch?v=Fda_-l6HB18&t=77s) The developers, they aren't being negative, they are being responsible, because when AI produces code at enormous speed, deep technical understanding is the only insurance you have against total system collapse. — Stein Inge Morisbak

> [20:54](https://www.youtube.com/watch?v=Fda_-l6HB18&t=1254s) Value isn't created when you buy licenses. — Stein Inge Morisbak

## Mentioned Stein Inge Morisbak, Tet Digital, Claude Code, NDC AI, Ruter, Skyss, Agder Kollektivtrafikk, AtB, Østfold Kollektivtrafikk, DORA, VG, Telenor, Schibsted, DNB, Microsoft

## Video description This talk was recorded at NDC AI in Oslo, Norway. #ndcai #ndcconferences #developer #softwaredeveloper

Attend the next NDC conference near you: https://ndcconferences.com https://ndc-ai.com/

Subscribe to our YouTube channel and learn every day: / @NDC

Follow our Social Media!

https://www.facebook.com/ndcconferences https://twitter.com/NDC_Conferences https://www.instagram.com/ndc_conferences/

#api #code #ai

Anyone can spin up an app, generate an API, and ship to production — without writing a line of code themselves. Top executives are declaring that technical expertise is becoming optional, that "anyone can be a technologist" without learning to code. Meanwhile, developers are reporting the opposite: FOMO, dread, and the sense that bad teams get worse with AI, not better. Both can't be right. This talk shows a better approach: shift the focus from chasing specific attacks to teaching solid defensive coding principles. We'll build security knowledge directly into the agent — using skills, hooks, and standards — so it looks up the right guidance, plans for it, and reviews its own work. You'll leave knowing how to make your agent produce secure code far more often, by default.

At Tet Digital, we ran a six-week Claude Code pilot across 6 teams and 30 developers, then rolled it out to 180. 90% reported higher productivity. But productivity isn't the whole picture: AI makes mistakes, so team practices for catching and correcting them matter more, not less; there's reduced learning, so skill development needs continued focus; the developer role is shifting toward orchestration and quality control; and like any tool, this one needs maturation, not just access.

In this talk, Stein Inge Morisbak (CTO) shares the concrete findings from the pilot, why Tet chose a deliberate three-step rollout with mandatory training and guardrails instead of an open tap, and why AI is best understood as an amplifier — one that makes deep technical understanding more valuable, not less. When anyone can ship, someone has to know what we're actually shipping.

[Source](https://www.youtube.com/watch?v=Fda_-l6HB18)

Reading summary

Stein Inge Morisbak, Chief Technology Officer of Norwegian public mobility company Tet Digital, argues that artificial intelligence amplifies rather than replaces human capability, making deep technical expertise essential to prevent catastrophic system failures as generative code output accelerates. Executive Hype Versus Technical Reality (1:17) While corporate executives across major Norwegian enterprises project massive productivity windfalls from AI adoption, frontline developers express caution regarding technical debt and loss of code ownership. At Tet Digital—which operates systems handling 69% of public transport journeys in Norway—Morisbak frames this developer hesitation not as resistance to progress, but as critical engineering stewardship. Deep technical knowledge acts as an essential insurance policy against enterprise collapse when code is generated at high volumes. The DORA Amplifier and Pilot Dynamics (4:33) Drawing on DORA research, Morisbak explains that AI operates strictly as an amplifier: it reinforces the strengths of disciplined, high-performing teams while severely worsening existing dysfunction in struggling teams. To test these dynamics, Tet Digital conducted a six-week pilot of Claude Code across 6 teams and 30 engineers. The tool proved highly effective for large-scale, repetitive tasks like refactoring, test generation, and documentation, but struggled with complex enterprise cross-system dependencies. Developer self-assessments revealed that 90% reported higher productivity by the end of the pilot, 60% noted improved wellbeing, and 65% experienced a stronger sense of mastery. However, 10% observed a decline in code quality, while another 45% saw no change, emphasizing that rigid verification guardrails are necessary to protect the production boundary. The Delivery Paradox and Organizational Bottlenecks (9:12) Despite widespread developer sentiment of increased productivity, surveying product leads exposed a significant perceptual disconnect. Product managers observed no meaningful shift in delivery speed or code quality, and developers remained uncertain about meeting quarterly OKR targets. Morisbak attributes this paradox to the reality that writing lines of code is rarely an organization's bottleneck; rather, systemic delays stem from excessive meetings, vague priorities, slow decision-making, and inter-team dependencies that AI assistants cannot fix. Hard Lessons and the Gated Rollout (11:20) The pilot demonstrated four core operational lessons: AI models routinely make mistakes requiring rigorous code reviews; junior developers risk stunted skill acquisition unless paired with senior mentors; engineering roles are shifting from manual typing to orchestrating and auditing outputs; and tool proficiency requires dedicated maturation time over several weeks. To expand Claude Code to all 180 developers safely, Tet Digital rejected open access in favor of a three-step onboarding process requiring mandatory training, team leadership agreement, and a signed declaration. Usage policy was condensed into three rules: developers retain full personal responsibility for all code, confidential data and secrets must never enter the model, and engineers must ask for guidance when uncertain. Scaling Beyond Developers and Ongoing Measurement (17:54) To govern adoption sustainably, Tet Digital established a dedicated AI enablement team to monitor failure modes, run training, and maintain guidelines. As the organization extends AI pilots to non-technical staff such as designers and product managers, Morisbak notes that tools like command-line interfaces can be intimidating, reinforcing the need for technical experts to provide structural guardrails. He also acknowledges an ongoing operational struggle: distinguishing genuine organizational productivity gains from subjective developer satisfaction. Morisbak concludes that buying software licenses does not automatically generate value without deliberate change management and industry knowledge-sharing. Because AI makes foundational errors and amplifies baseline practices, retaining experienced technical experts remains the primary safeguard for ensuring sustainable, secure software delivery.

Source quotes

[1:17] Stein Inge Morisbak: The developers, they aren't being negative, they are being responsible, because when AI produces code at enormous speed, deep technical understanding is the only insurance you have against total system collapse. [20:54] Stein Inge Morisbak: Value isn't created when you buy licenses.

Original transcript

[0:06] Hello everyone. You can hear me well. So [0:10] last talk of the day. It's me holding [0:13] you back from the party and the beers. I [0:16] hope you can bear with me a little bit [0:18] more. Hope you still have the spirit [0:21] going for listening to one more talk. [0:25] Ehm, my name is Stein Inge and I'm CTO [0:29] at Tet digital. [0:32] Eh, and for those of you who don't know [0:36] digital, eh, we're a technology company [0:40] on mobility solutions. We're owned by [0:43] the largest [0:45] eh public transport authorities in [0:48] Norway which is Ruter her in Oslow eh [0:52] skys eh inn Vestland Agder [0:56] kollektivtrafikk [0:58] at AT til B in Trøndelag and Østfold [1:01] kollektivtrafikk. [1:04] Today eh 69% of all public transport [1:08] journeys eh in Norway go through our [1:12] digital systems. [1:17] The media is flooded with content about [1:21] AI eh and top executives they have very [1:27] high expectations [1:29] and let's look at what some of them say. [1:33] I came across this article eh and just [1:37] want to say that this is from VG and the [1:40] translations I have done myself eh and [1:45] inserting them on to eh the images is [1:48] done by AI. [1:55] Oh, this is the headline [2:01] and this is the good news. [2:06] This is some of the stuff they saying. [2:09] These are top executives in top [2:11] companies in Norway, Telenor OSer, DNB [2:16] and Microsoft. [2:19] Hope well maybe some of you are in the [2:21] audience today. So but this is this is [2:26] public something that these people have [2:28] said. Så I think it's eh I think it's eh [2:33] okay to put it up there. [2:56] So and now this is the counterweight eh [3:00] the fair or you may call it the reality [3:05] eh from those who know what technical [3:07] depth is and the loss of code ownership [3:12] actually means. Really liked the last [3:16] talk in this scene. Yngve eh talked [3:22] something about ja bit some of this ehm [3:28] and the developers the aren't being [3:30] negative [3:32] they are being responsible [3:34] because when AI produces [3:38] code at enormous speed deep technical [3:41] understanding is the only insurance you [3:44] have against total system collapse. [4:08] But eagernness for AI is of course not [4:12] only coming from above, it's coming from [4:16] our best people the ones who can see the [4:19] superpers on the horizon and they feel [4:23] held back. I think many companies, many [4:26] CTOs have experienced [4:30] messages like that. [4:33] And as many has said today, AI works as [4:36] an amplifier for better and for worse. [4:40] And the framing comes from Dora's [4:43] research on AI in software teams. [4:47] It amplifies the strengths of high [4:50] performing organizations and it [4:53] amplifies the weaknesses of struggling [4:56] ones. [5:00] This photo [5:01] of course AI generated was taken just [5:04] before Christmas [5:06] which explains the Christmas sweater. [5:09] That's when we kicked off the pilot and [5:11] it sums up my role. [5:14] I'm doing the splits not between leaders [5:18] and developers but between a shared hope [5:22] of moving faster and with higher quality [5:25] and shared responsibility for making [5:28] sure we don't run off the cliff. [5:33] So how did we resolve the splits? [5:37] Eh we went straight to where the stakes [5:40] are highest. eh is our main asset or [5:45] code the developers [5:48] because they are the ones that are close [5:49] to production so where mistakes and most [5:54] yeah where mistakes can happen and where [5:57] most expensive to fix. [6:00] So if we can make it work there with the [6:03] right guard rails, you have a foundation [6:06] to build on. [6:08] And here's how we set it up and some of [6:11] our eh findings. [6:17] First a little bit about the pilot. We [6:19] ran the pilot with six teams, 30 people [6:23] over six weeks. [6:26] And the slide shows the three areas that [6:29] we focused on. [6:31] How people and the work itself were [6:33] affected, [6:35] what cloud code is actually useful for, [6:38] which is the AI implementation that we [6:41] used and the guard rails we needed for [6:44] safe use. [6:47] And the short version on use cases, it [6:51] worked very well for repetitive work at [6:54] scale, refactoring, testing, [6:57] documentation and debugging. [7:01] Less well for very complex tasks or [7:05] tasks tangled up in too many [7:07] dependencies. These are enterprise [7:10] systems. There's a lot of dependencies [7:12] just not across systems but also across [7:16] teams and yeah that's hard come back to [7:20] the guard rails later. [7:27] in the work. [7:31] So most people experienced higher [7:35] wellbeing and a stronger sense of [7:39] mastery. [7:41] A full 60% reported higher wellbeing [7:45] while remaining 40% they saw no change. [7:49] It wasn't worth. [7:52] Worth noting here is that these numbers [7:55] are have shifted during the pilot. At [7:59] the start, several people actually [8:01] reported lower wellbeing. The very [8:05] start. It took time before the benefits [8:08] showed up. [8:10] Eh, on the sense of mastery, 65% [8:14] reported it went up while 25% saw no [8:18] change and 10% reported decline. [8:25] On productivity, the full 90% of [8:28] developers reported higher productivity. [8:32] 10% saw [8:34] and notably not a single developer [8:38] reported a decline. [8:44] Quality is a different story. 45% [8:48] reported higher quality, 45% saw no [8:52] change, and 10% reported a decline. [8:57] So productivity went up across the [8:59] board, quality is more mixed and that [9:03] 10% is part of why guard rails matter [9:07] [snøfter] [9:08] and we come back to those. [9:12] And then there's the product manager. [9:16] Eh we looked at whether product leads [9:19] saw the same effects as the developers [9:23] and the answer was quite interesting. [9:28] no noticeable change in productivity and [9:33] very little change in quality [9:36] that said they attributed it that the [9:39] quality [9:41] already was very high in in their teams [9:47] what they did notice was that developers [9:50] themselves feel more productive and are [9:53] having fun with the tool [9:57] That gap between how developers [9:59] experience it and what product leads [10:02] actually observe sets up the numbers on [10:05] the next slide. [10:11] Even though the majority reported higher [10:14] productivity, many were still unsure [10:17] whether theyd be able to deliver on [10:19] their OKR in time. [10:23] That might seem contradictory. [10:26] It expect higher productivity [10:29] into more confidence about delivery, [10:34] but it doesn't. [10:37] And what does this tell us? [10:42] Tells us that AI isn't a magic tool that [10:45] solves every bottleneck in delivery. [10:49] Points is something more important. [10:51] Writing code is rarely the real [10:54] bottleneck. [10:56] Previous speaker always also made a [11:00] point of. [11:02] Eh, so many of the challengers they lay [11:06] elsewhere meetings, dependencies between [11:09] teams, unclear priorities, slow [11:12] decisions and all that other stuff going [11:15] on in an organization. [11:20] And a few other things that we learned [11:24] and what they all have in common eh that [11:27] AI doesn't handle them for you [11:33] because AI makes mistakes. So good [11:38] routines, processes and team practices [11:41] matter more than ever for catching and [11:46] correcting them. And this is where the [11:48] amplifier idea really shows up. [11:51] Highperforming teams get even better [11:54] with AI [11:57] and teams that already struggle will use [12:00] those will see those struggles [12:02] amplified. [12:04] Which means rolling AI out to a [12:07] struggling team doesn't it doesn't fix [12:09] the team. [12:12] It makes the problems worse. [12:17] And then there's a matter of reduced [12:21] learning. [12:23] AI can reduce learning [12:27] and this is especially true for junior [12:31] developers if not seniors are paying [12:34] attention to them. [12:38] That doesn't mean keeping AI away from [12:41] them. It means being deliberate about [12:46] how they use it. [12:49] So active skill development matters [12:53] pairing, code reviews, deliberate [12:57] learning time, whatever it takes to make [13:00] sure juniors actually grow. [13:05] And then there is the shift in tasks. [13:09] Developers move from writing most of the [13:12] code themselves to orchestrating AI [13:15] output and verifying it. Less typing, [13:20] more judgment. [13:22] So the hands on skill that used to be [13:25] the job becomes a skill that makes you [13:28] safe to do the job. [13:34] Finally, adopting AI tools takes [13:38] maturation [13:40] because like any new tool, it takes time [13:43] and experience before you use them well [13:45] and productively which means [13:49] give it that time not declaring victory [13:53] or failure too early. [13:59] Here's what maturation looked like in [14:01] our data. [14:03] We surveyed multiple points during the [14:07] pilot, not just [snøfter] at the end. [14:09] Eh, so we could see the trajectory, not [14:12] just the destination. [14:15] These are the same developers surveyed [14:18] twice. [14:20] Early on 83% reported higher [14:22] productivity, 13% saw no change and 4% [14:27] reported a decline. By the end of the [14:29] pilot, nobody [snøfter] reported a [14:31] decline and only 10% saw no change. And [14:35] full 90% reported higher productivity. [14:43] This is why we keep the nerds. [14:46] So [14:48] four findings [14:50] to repeat them. Each one needs people. [14:55] People who can design the practices that [14:57] catch AI mistakes. [15:00] People whose technical debt doesn't erod [15:06] people who can verify AI output because [15:09] they can read it. [15:11] People who can drive maturation [15:14] and this is exactly [15:18] the reason why we keep the nerds. [15:22] How did we actually roll cloud coil out [15:24] to 180 developers? [15:29] Because we didn't just hand it out to [15:32] get access to cloud code. We have to go [15:35] through three steps. First mandatory [15:39] training focused on AI as an amplifier, [15:43] security and practical use. This is not [15:47] [snøfter] [15:48] a full on course on how to use clot or [15:52] something like that that developers can [15:55] do themselves through other practices. [15:58] But these are eh training focus on [16:01] amplification, security and practical [16:04] use. [16:10] Then agreement with the team leadership [16:14] to actually adopt the tool. So the whole [16:17] team has to agree on eh adopting the [16:21] tool because they are responsible for [16:23] the outcomes. So not just the [16:25] developers. [16:27] And finally finding a declaration that [16:31] you have read understood and will follow [16:35] the guidelines. [16:39] So the guidelines are quite easy eh so [16:44] that everyone eh can follow it and [16:47] everyone has to follow it. [16:49] And [16:51] point one is that you're responsible for [16:54] all code that AI generates. Always [16:57] verify the output with human review. [17:03] Never share sensitive information, [17:05] personal data, secrets or business [17:08] confidential. [17:10] And the third, probably the most [17:12] important, use your head. Ask for help. [17:16] Don't do anything if you're unsure [17:18] whether it's within the guidelines. [17:21] So rather you ask than guess. Three [17:25] guidelines. That's it. One rules that [17:29] people would actually remember. It's [17:31] also possible to post on Slack also eh [17:36] without posting a wall of text that no [17:40] one will read. [17:43] So what's next? [17:47] We just in the beginning. This is what [17:50] we are working on now. We are eh [17:54] building an AI enablement team [17:57] and this team's job is training, [18:01] maintain the guidelines, help teams that [18:04] get stuck, watch for the failure modes [18:08] we saw in the pilot and measure what's [18:11] actually working. Matrix is extremely [18:14] important. [18:17] from [18:23] up on [18:26] how AI is doing in our organization on [18:29] several aspects. We have the have guard [18:33] rails in place but measuring eh are we [18:38] actually going faster slower? It's only [18:42] the feeling of people that can eh that [18:45] are used for measuring that. [18:49] So we wanted AI adoption to stay [18:51] deliberate, not accidental [18:55] and [18:57] developers not the only ones eh [19:00] interested in AI. So we're also piloting [19:03] new groups. Eh we're currently piloting [19:08] designers and product managers. [19:12] And by doing that, that's a little bit [19:15] different because eh it reinforces the [19:20] main point of this talk. [19:23] Buuse designers and PMs, they don't know [19:26] the terminal, [19:28] they find it awkward [19:30] which makes it a little bit hard and [19:34] it needs experts to eh [19:39] serve. as people that can guard rail the [19:44] use. [19:46] So if we want these groups of people to [19:49] use cloud code, we need to build basic [19:52] developer skills first. [19:54] AI doesn't remove the need for technical [19:56] debt just shows you where the debt is [19:59] missing. [20:02] So we are barely at the launchpad but [20:05] lifof I would say has gone well. Eh and [20:10] the challenge now is scaling. [20:12] was pilot into that delivers value [20:16] across the whole company every team [20:19] sustainability sustainably [20:23] eh and AI strategy [20:27] it often becomes very abstract and all [20:30] compassing. [20:32] We're trying not to do that. [20:35] Our focus has been and will be for a [20:38] while learning the tools and using them [20:40] well in the organization we have before [20:43] we make big changes to the way of work, [20:47] how we organize the processes and and [20:49] stuff and before AI changes everything [20:52] if it does. [20:56] So how do we set ourselves up for eh [21:01] scaling from pilot to practice? We're [21:04] sticking with doing this in a controlled [21:06] and responsible way. [21:08] We're establishing the AI enablement [21:11] team to provide support and ensure good [21:14] guard rails. Security, legal and ethics [21:18] are an absolute foundation. [21:21] Security is is a is a different talk. [21:24] can be I know many companies struggle [21:27] with with this [21:30] eh no not security but legal eh [21:35] and when development is moving this fast [21:38] we have to learn and help each other [21:39] through a community of practice and I [21:42] encourage everyone here to be part of a [21:44] community of practice because basically [21:47] weeve all become new bees [21:50] so where even if you a senior [22:02] to not know [22:05] and I hope we share what we learn across [22:08] companies and with academia [22:14] but the most important point is at the [22:18] bottom value isn't created when you buy [22:22] licenses [22:30] AI pilot to build lasting capability [22:35] is basically change management [snøfter] [22:38] and AI is just the means. [22:42] These are the people we're talking [22:44] about, real people doing real work and [22:48] whose expertise matters more in this [22:52] age, not less. [snøfter] [22:54] So, what do we take away from all this? [23:00] Yes, you should of course get started, [23:03] test it, pilot it, deploy it. You can't [23:07] learn when you're standing on the [23:09] sidelines, [23:11] but you need to do it properly. [23:15] And remember, for God's sake, keep the [23:18] nerds. [23:20] AI doesn't remove the need for the [23:22] expert. It makes them more important [23:25] than ever. [23:28] Thank you. [23:30] [applaus] [23:35] [applaus] [23:38] Do think we have time for questions if [23:40] someone [23:42] wants to ask some questions otherwise [23:43] you can come over here. I'll be here for [23:47] a while.