Afraid of Your Own Culture

Welcome to the first issue of The CAIO. The newsletter you knew as The Cybernetic CxO has a new name, and the reason is on the cover of this issue. From 1 May, I am Group Chief AI Officer at Zühlke. The job is the same work I have been doing for years: getting AI past the pilot phase through transformation across organisation, process, and technology, not by tuning the model alone. The newsletter title now says that out loud.

May was the heaviest month this publication has ever had. Two posts cleared 100,000 impressions on the same theme: the model is the small part of an AI system, the discipline around it is the work. One post asked senior leaders whether they were afraid of AI and got an answer most of us did not want to read. A 100-year-old economic paper resurfaced and pulled 99 comments. A Chicago Booth art auction quietly produced the cleanest data yet on what happens when humans and AI mix in a product.

This edition covers what the audience actually opened. The fear inside the C-suite that nobody puts in a slide. The architecture lesson hidden in 512,000 lines of Claude Code. The discipline pattern across XP, Cloudflare, and Flask. The jobs question, finally with research behind it. The label question, finally with hard numbers. The hierarchy question, finally without polite framing. And, for the industrial readers, a clean cut on the cyber-physical shift.

Afraid of Your Own Culture

#

The most uncomfortable AI study of May had nothing to do with models. MIT Technology Review Insights surveyed 500 senior leaders worldwide. 22% admitted they had stepped back from leading an AI initiative because they feared the blame if it failed. Another 5% preferred not to answer, so the real number is almost certainly higher. Only 39% rated their own organization’s psychological safety as “very high”. A majority of senior leaders are pushing AI adoption on cultural conditions they themselves describe as mediocre.

That is the bottleneck. Not GPU capacity. Not model quality. A culture where the people closest to a failing pilot would rather quietly bury it than say so.

“Failure is not the opposite of success. It is a requirement for it. We have built a corporate culture, and, frankly, a society that rewards certainty and punishes iteration. That mindset suffocates innovation.” (Melanie Subin, Future Today Strategy Group, quoted in the report)

The report’s second finding is sharper. Psychological safety is not an HR program. It is a leadership behavior. Nick Schacht from SHRM put it directly: one negative word from the CEO and the best leadership anywhere else in the org is undone. You cannot delegate this to a training curriculum or a values poster. A CEO who punishes the messenger has the AI strategy of the messenger’s last sentence, and nothing more.

And the report’s third finding is the one I want every CxO to sit with. The single biggest lever for psychological safety, according to the survey, is clarity about how AI will and will not affect jobs. 60% of leaders say that is the most important thing. If you cannot articulate which part of the work stays human and why, every AI initiative looks like a staffing threat, and every honest question feels dangerous.

Your AI readiness is not your roadmap. It is whether someone on your team would tell you that this quarter’s pilot did not work, or quietly close the tab and move on. Which one would they do?

The Harness Is the Product

#

Culture is one side of the operating model. Architecture is the other. While the boardroom worries about the model, the engineers ate the rest of the iceberg.

Researchers from MBZUAI and UCL published a source-level analysis of Claude Code on arXiv last month. They went through 1,884 files and 512,000 lines of TypeScript. Their finding is precise. Only 1.6% of the codebase is AI decision logic. The other 98.4% is the harness around it: a seven-mode permission system with an ML classifier, a five-layer compaction pipeline so the model does not lose context, 54 built-in tools that execute the actual work, 27 hook event types for plugins and policy, append-only session storage, subagent delegation, recovery when things break, and one rule that ties it all together: the model never directly touches the filesystem, the shell, or the network.

The paper states the bet on page three: “The architecture invests in deterministic infrastructure (context management, tool routing, recovery) rather than decision scaffolding, on the premise that increasingly capable models benefit more from a rich operational environment than from frameworks that constrain their choices.”

In other words, Anthropic did not bet on a smarter model. They bet on a platform that lets a model be used safely at scale. The Cybernetic Platform pattern from my book is in production. It just has a different name on the repo.

Armin Ronacher, the creator of Flask, pointed at the same wall from the other side. He wrote one line that travelled further than anything I have shared this year: “When more and more people tell me they no longer know what code is in their own codebase, I feel like something is very wrong here.” He is right. The production speedup is real. The organizational capacity to manage what gets produced is not.

History is not subtle about this. In the British Industrial Revolution, every time a bottleneck disappeared, a new one showed up downstream. Weaving sped up, yarn became the constraint. Spinning sped up, fiber became the constraint. Fiber arrived faster, cotton supply became the constraint. You never solved the problem, you moved it. Software has played this game for decades, through languages, frameworks, and cloud. Each wave removed one constraint and exposed the next. The human review layer was always next in line. Now it is here.

OpenClaw has 2,500 open pull requests sitting stale. Pi auto-closes incoming PRs from anyone not on the trust list. Senior engineers tell me, on calls and on stage, they have stopped reading their own diffs. Two paths remain. Throttle the input, restrict who can open PRs, take what Ronacher calls “OSS vacations”. Or accept the acceleration, let machines review machines, rubber-stamp in the morning, diffuse responsibility so no individual carries the signature.

Path two sounds modern. Path two is also the path Anthropic explicitly designed against. Machines cannot carry responsibility. Not legally, not socially, not ever. Someone will always have to answer for what ships. If accountability cannot transfer to AI, then humans remain the bottleneck no matter how fast generation gets. Faster code does not change who gets the 3 AM call.

The same paper cites three sobering findings. Developers in AI-assisted conditions score 17% lower on comprehension. Users approve 93% of permission prompts, so consent is no longer a meaningful safety mechanism. Cursor adoption raised code complexity by 40.7%, with velocity gains gone by month three. The harness compensates for the model’s flaws. It does not compensate for the human atrophy the model creates.

Where is your team investing this quarter, in the 1.6% or in the 98.4%? Because the gap is not closing on its own.

Discipline Beats Intelligence, Year Two

#

Last issue I argued that the discipline pattern across Amazon, Wix, and Vercel was the central story. May added three more witnesses that say the same thing.

Chad Fowler (on the XP throwback). Fowler was inside the Extreme Programming movement in 1999, before the Agile Manifesto existed. From the outside, XP looked reckless. No long-range plans. No design documents. Tests before code. Continuous releases. But XP did not remove discipline. It relocated it. Tests replaced promises. Working software replaced status reports. You could not hide behind process anymore. The same pattern repeated three times. Dynamic languages replaced static types, and rigor moved from compiler declarations into executable tests. XP replaced phase-gate development, and rigor moved from Gantt charts into continuous feedback. Continuous deployment replaced release management, and rigor moved into observability, reversibility, and automated verification. Now generative AI repeats the pattern. The fix is the same as 1999. You write the tests. The AI generates the implementation. If the tests do not pass, the code does not ship. Fowler’s line for it: probabilistic inside, deterministic at the edges. Generation can be flexible. Evaluation must be rigid. Cheap generation without strict judgment is not a new paradigm. It is abdication.

Boris Tane at Cloudflare (on slop creep). Boris coined a name for what Anthropic’s paper measured. Slop creep is the slow accumulation of individually reasonable but collectively destructive decisions, each one too small to flag, together quietly wrecking your system. Before AI agents, bad architectural decisions caused pain. That pain forced fixes before catastrophic decay. The friction was the circuit breaker that kept codebases from drifting into chaos. Agents removed the circuit breaker. They pile up questionable architecture indefinitely while still shipping features. Boris’s own data is the most chilling part. Building Baselime pre-agent, reaching the architectural breaking point took months. A recent side project built with agents reached the same breaking point in days. Same person, same skills, same kind of bad decisions, orders of magnitude faster. His diagnosis is one sentence: “Coding agents don’t see the system, they see the prompt.” His fix is the same as Fowler’s, just applied earlier. The engineer defines intent. The agent drafts a plan. The engineer annotates and corrects. Two or three iteration rounds. Then the agent implements, inside the constraints you locked down.

My own Continuous Delivery chapter (on fear of releasing). Continuous Delivery is a chapter in The Cybernetic Enterprise. The line I am proudest of is also the bluntest: the real bottleneck is not your code, it is your fear of releasing it. CI catches developers lying to themselves. Continuous Deployment goes one step further: when every push delivers value, you do not need roadmaps, you need runway. These are not new ideas. They are the ideas that got drowned out in 18 months of “let the agent ship”. Nothing in the agent era invalidates them. The opposite. The agent era exposes every team that adopted velocity without adopting the discipline that velocity requires.

Across all six pieces (April’s Amazon, Wix, Vercel; May’s Fowler, Tane, and the DevOps cluster) the pattern is one sentence. Today’s models are more capable than the release processes around them. Discipline is now the bottleneck. The teams that pull ahead are the ones treating governance as a product they build and maintain, not a policy document they file away.

Lump of Labour, 100 Years Later

#

The “what is AI doing to jobs?” question dominated March. In May, it came back with research behind it.

The lump of labour fallacy is the 100-year-old name for the mistaken belief that the amount of work in an economy is fixed. Like a pie. If machines do more, humans must do less. Every technological shift has triggered this exact fear. Every time, the fear was wrong. When ATMs were invented, banks were going to fire all the tellers. What actually happened: ATMs made branches cheaper to run, banks opened more branches, the number of tellers went up for decades, and the job shifted from counting cash to customer service and sales. The work did not disappear. It transformed.

Now look at AI. Billions invested, unemployment has not spiked, millions are using AI daily, and entire new job categories are emerging that did not exist five years ago. This does not mean everything is fine. Entry-level workers and new graduates are feeling real pressure. The transition is never painless. But the premise that work is a fixed pie is false, and acting on that false premise is what generates the worst leadership moves of any AI rollout. The question is not whether there will be jobs. It is whether you are prepared for the new ones.

A paper from LSE and the University of Hong Kong, titled Weak Bundle, Strong Bundle, gave us the math for the next layer of the question. The question is not which tasks AI can do. The question is how much value is destroyed when those tasks are done separately.

A radiologist does not sell scan reading. They sell a bundle. Triage, conversations with referring physicians, training residents, and a signature on a call that other people will act on. The market buys the bundle, not the classification.

The paper names three reasons a bundle is expensive to split. Shared context: the person who read the scan also talked to the referring physician. Liability: the person who signs the diagnosis cannot outsource the judgment because they cannot outsource the consequences. Cross-task spillovers: what you learn doing one task makes you better at the other. Picture one meeting, two jobs. An AI notetaker captures who said what and what was decided. That task splits cleanly from being in the room. Weak bundle. Now picture a project manager in the same meeting. She hears the engineering lead say “sure, we can try that” in a tone that means “this will fail and I am not going to fight about it”. An hour later she takes him for coffee, gets the real objection on the table, and saves the project. Same meeting. Completely different job. Strong bundle.

The framing maps directly to the Cybernetic Enterprise. AI is the nervous system. Humans are the soul. The nervous system moves signals fast. The soul holds context, liability, and judgment. You do not replace a soul with a better nerve. If you are planning an AI rollout, do not start with a list of tasks AI can do. Start with the bundles you cannot afford to break.

What the Label Says

#

If the jobs question is about what humans do, the next question is about how the market values what they do with AI in the room.

The Chicago Booth contamination study ran two preregistered art auctions with real money. 351 participants bidding on actual prints. Each description varied randomly in how much was made by a human, how much by AI, and how exclusive the piece was. Three numbers stuck with me. A painting described as 99% human and 1% AI lost 16.12% of perceived value compared to fully human-made. Going from 25% AI to 50% AI only moved value by 3.13%. Adding 1% human touch to a 100% AI work boosted value by 27.92%. The curve has kinks at 1% and 99%. The middle is almost flat.

This is not a gradient. It is a contamination response. The same psychological pattern Rozin documented decades ago with a sterilized cockroach touching juice. Nothing about the juice changed. The juice was no longer the same category of thing. The disclosure wisdom most companies follow is backwards. Small AI involvement is not the safe thing to disclose. The first 1% is the most expensive 1%. Once you have crossed the line, you have already paid most of the price. Fully human, or AI as a named tool, may be cleaner market positions than mostly human with a little AI help.

One caveat before anyone over-generalizes. This is art. Art is uniquely about authorship. Software is mostly judged by whether it works. Consulting is judged on outcomes. Knowledge work sits somewhere in between, and the paper itself says we do not know where. Do not turn a 16% finding from an art auction into a default discount for AI-augmented services. But do think about what you write on the label. The augmentation story still works. How you frame the human contribution is now a design decision, not an honest-disclosure checkbox.

From Hierarchy to Intelligence

#

The same operating-model pressure shows up at the very top of the chart.

Jack Dorsey (Block CEO) and Roelof Botha (Sequoia) published a detailed plan to replace middle management with an AI-powered world model. Their argument is older than it looks. A leader can effectively manage three to eight people. That is a biological limit, not a leadership style. Everything else in the corporate pyramid is engineering around that one constraint. They walk through 2,000 years of history to prove it. The Roman legion’s contubernium of 8, century of 80, cohort, legion: a span-of-control protocol, not a plan. The Prussian General Staff in 1806: created to support incompetent generals after Napoleon exposed the risk of relying on individual genius. American railroads in the 1850s: Daniel McCallum built the world’s first organizational chart because informal management was killing people in train collisions. Every flat-org experiment since (Spotify, Zappos, Valve) either reverted to conventional management at scale or collapsed.

The trap in one line: “narrowing span of control means adding layers of command, but more layers mean slower information flow.”

And then the sentence that should pause every org chart owner: the question was never whether you needed layers. The question was whether humans were the only option for what those layers do. They are not anymore.

This is the Cybernetic Enterprise pattern, just described by someone selling payments software. AI as nervous system. World models. Feedback loops. Humans at the edge where the model meets reality. The idea is arriving from multiple directions because the underlying reality is the same: the biological limit on span of control is finally being relaxed, and everything downstream is up for reconsideration.

What Dorsey and Botha may be underselling: this transition will not break because the technology is wrong. It will break because people will resist being told that a layer of work they built careers on was always a workaround, not a craft. That is the real adoption problem, and it lives on the same axis as the MIT psychological safety finding I opened this issue with. The org chart you defended in front of your board did not exist 200 years ago, and may not exist 200 days from now. Who would fight hardest in your organization to keep what should disappear?

Hardware to Systems, Industrial Edition

#

A short section for the industrial readers, because the operating-model question lands differently when it is also a product-definition question.

Customers no longer buy compressors. They contract guaranteed compressed air. The system predicts leaks, adjusts output, optimises energy, and ensures uptime. The motor’s peak performance is invisible to the buyer. Only the outcome is visible. This is outcomes over output, industrial edition, and Réka Leisztner, Urs Künzle, and Thomas Weber from Zühlke captured it cleanly earlier in May.

A cyber-physical system is not hardware with an app bolted on. It is a continuous feedback loop between physical and digital, engineered as one coherent product. Sensors capture, software decides, the machine adapts, insights feed back. That loop is the product. For legacy industrial players, especially in DACH, this looks like a threat. It is the opposite. The physical layer is the hardest part to replicate. Decades of domain knowledge, customer trust, production know-how, and installed base are barriers digital natives cannot buy their way around. The bottleneck is not legacy. The bottleneck is mindset. As long as hardware comes first and software follows, integration stays a painful afterthought.

The matching data point comes from the Cisco 2026 State of Industrial AI Report. 65% of German industry uses AI. Only 1 in 5 in EMEA is ready to scale it. The seam between IT and OT is where the whole thing either holds or breaks. 88% are confident in scaling AI without IT/OT alignment. 96% with it. 80% are confident in staying compliant without alignment. 98% with it. The delta is the operating model.

Three things turn industrial AI use into industrial AI scale: a shared P&L between IT and OT, a shared platform, and a single accountability for outcomes. The next decade of revenue for industrial manufacturers turns on one product-definition choice. Either the product is a component inside someone else’s system, or the product is the system.

My Current AI Stack

#

Claude Code: Still the primary tool. Reports, meeting prep, coding, everything. The AGENTS_md pattern I started in April (hard rules at project root, no commits without explicit permission, build/lint/test before pushing) is now muscle memory. New for May: I started using Claude Code subagents and skills for repeatable workflows. The harness paper from MBZUAI/UCL only made this discipline feel more right.

Perplexity: Web research with real sources. Still the only tool I trust for “find me ten primary sources on X”.

NotebookLM: Documents in, audio and video summaries out. Useful again this month for the longer arXiv papers cited above (Dive into Claude Code, Weak Bundle / Strong Bundle, the Chicago Booth contamination study).

Gemini: Image generation. Best and fastest for the cyberpunk style I use for newsletter covers.

Gamma (last appearance): I retired Gamma this month. Two reasons. First, the output never quite matched my voice without rework, so the gain over slides-from-scratch was thinner than it looked. Second, every other tool on this list is one I can shape (skills, instructions, files), and Gamma was the only black box. I built a Claude skill that knows my structure, my visual defaults, and the patterns I reuse. Less tool, more system. Same pattern as everything else in this issue: invest in the harness around the model, not in another product on top of it.

Vibe Coded This Month (Private Projects)

#

• KnowledgeBase: Continued build-out from April. Notes, reading notes (books, papers, presentations, videos, courses), and curated reference materials. Claude Code is the primary interface. Strict filing rules so the agent never has to guess where something belongs. The full system runs from one OneDrive folder. The discipline lessons in this issue (Fowler’s relocated rigor, Tane’s slop creep prevention) shape how I extend it: every new content type gets its own schema before it gets a file.
• Taskmanager: Also continued from April. Natural-language input, structured task.md files with frontmatter (priority, category, due date, subtasks) as output. No Notion, no Asana, no Todoist. The May upgrade was a stricter front-matter contract so subagents can plan against the task graph without re-reading every file. The Claude Code harness paper is, in miniature, what I am building here.
• Fitnessplaner (new this month): A personal fitness planner. Workouts, progressions, recovery, and notes, again as structured markdown that Claude Code can read, plan against, and adjust. The interesting part is not the markdown. It is what happens when the planner has access to the same knowledge base where I store research articles on training. Cross-system context is the unlock.

The pattern across all three: replace a SaaS habit with structured folders plus an AI interface. Discipline first, then the agent. The exoskeleton question (where does the lifter hurt?) drives what I build next.

Where to Find Me

#

A short closing for those who want to meet in person.

June. Sovereignty at Digicomp Zürich on 3 June with Martin Thalmann. The Momentum panel on 8 June in Frankfurt with Sara Jourdan, Peter Buxmann, and Daniel Schleidt, on the shift from copilots to autonomous agents. The MobAI Workshop at the Next Industries 7. Use Cases-Event on 11 June.

July. Industry.forward in Berlin on 1 July, moderating the panel on the Industrial AI Roadmap 2040 with Markus Himpele, Martin Bergmann, and Tim Herfurth. AI value creation at Digicomp Zürich on 2 July with Martin Thalmann.

September. DevOps human side at Digicomp Zürich on 1 September with Martin Thalmann.

December. Closing out 2026 at Digicomp Zürich on 8 December, two practitioner talks with Martin Thalmann.

January and beyond. Martin and I are already booked for 21 January 2027. And the DevOpsDays Zurich 10-year anniversary is set for 14 and 15 April 2027.

What to Watch in June

#

Three things on my radar:

1. First half-year-of-CAIO honest read. I will publish a public reflection on what I learned in the first 60 days of the role: what worked, what did not, where the operating-model thesis met reality. Subscribe to The CAIO if you want it in your inbox.
2. The “label” question reaching software. The Chicago Booth contamination finding is about art today. Knowledge work next. Watch for the first SaaS company to test what happens to pricing when they explicitly disclose AI involvement.
3. The first major lawsuit on AI-generated production code. Boris Tane’s slop creep observation will eventually be tested in court. The path from “the agent made the bad decision” to “the engineer signed off” is the path from technical debt to legal liability.

See you in June.

If this issue connected for you, the most useful thing you can do is forward it to one peer who needs the same conversation, and reply with the single sentence from this issue you would put on the wall.

Until June.

Romano

The CAIO is a monthly LinkedIn newsletter for C-level executives and senior leaders translating AI developments into operating models. Edition 5 covers posts published 24 April to 23 May 2026. Friday Fun posts excluded.