The AI Order Nobody Signed: Trump's Flip-Flop, Örebro's Empty Factories

Here's what nobody in Washington or Brussels wants to say out loud: the AI order got signed, delayed, rewritten, and signed again this week, and it barely matters. Not because AI policy is unimportant. Because the real crisis is already happening, and it has nothing to do with what's written in an executive order. It's happening in Örebro, in factories that bought the robots, installed the software, and now sit there with machines nobody can operate.

That's the story. Not Trump. Not the EU AI Act. The gap between having the technology and having humans who can run it. That gap is the actual bottleneck of the next five years, and almost nobody in policy circles is talking about it.

What Actually Happened in Washington This Week

Trump signed an AI oversight executive order. Then it got delayed. Then he signed a different version. Reporting from the New York Times, the Guardian, and the Washington Post all point to the same thing: the final terms shifted because big tech pushed back and got what it wanted. Read that again. The companies the order was supposed to govern helped write the version that got signed.

This isn't new. It's just more obvious than usual. US AI policy right now is a negotiation between the federal government and five or six companies with market caps bigger than most G20 economies. The order exists mostly as theater, a signal to markets and voters that "someone is doing something." The actual constraints on frontier AI development are still being decided in boardrooms in Mountain View and Redmond, not in the Oval Office.

I'm not saying this to be cynical for its own sake. I'm saying it because if you're a European founder or CTO looking at this and thinking "at least we have real regulation," you need to sit with a harder question: does the EU AI Act actually solve anything that matters right now, or does it just make compliance the main event while the operational stuff quietly breaks?

Örebro Is the Preview, Not the Exception

While Washington performs oversight theater, Swedish manufacturers in Örebro already lived the real AI story. They automated. They invested. They bought the systems, integrated the robotics, upgraded the lines. And now, according to reporting from KiTalent, they can't find people to operate what they built. Not engineers to design new automation. Operators. People to run and maintain the systems already sitting on the floor. This is not a skills gap you close with a weekend course. It's a structural mismatch between the speed of technology adoption and the speed of workforce development, and Sweden hit it first because Swedish manufacturing moved fast on automation while the labor pipeline, vocational training, immigration policy for technical trades, regional population balance, moved at government speed.

Meanwhile Microsoft's CEO is in London this week doing enterprise tours, pushing Copilot adoption at firms like EY. Same pattern, different layer of the stack. The tool exists. The rollout exists. What doesn't exist yet, at scale, is an organization that knows how to actually restructure work around it. Software adoption in big consulting and professional services firms is running into the same wall Örebro hit with hardware: having the tech isn't the constraint anymore. Having people who know what to do with it is.

Two Headlines, One Root Cause

Put these side by side. Trump's order got shaped by the companies it regulates because the government doesn't have the technical depth or speed to write real constraints on frontier AI without industry input. Örebro's factories sit under-operated because Sweden invested in the machines faster than in the people. Both are the same failure mode: institutions optimizing for the appearance of control (a signed order, a completed automation project) instead of the operational reality underneath (who's actually running this, and are they ready).

Sweden vs the World: We're Not Behind, We're Early

Here's the thing I keep telling people in Jönköping and Stockholm both: Sweden is not behind on AI. We're early into a problem the rest of the West hasn't hit yet. American tech companies are still in the "ship the model, ship the copilot, book the enterprise deal" phase. Swedish industry already automated at scale, years ago in some sectors, and is now living the second-order consequence: a labor market that can't absorb the shift fast enough.

Compare that to the US, where the fight right now is still mostly about who gets to write the rules for model behavior and data use. Compare it to the EU, where Brussels keeps producing frameworks, the AI Act, guidance documents, sector codes, that assume the main risk is misuse of AI capability. Compare it to Asia, where China and increasingly India are scaling deployment with far less friction and far more state coordination between industrial policy and workforce planning.

Sweden sits in a strange spot. We have the industrial base to automate aggressively (see: Volvo, SKF, Ericsson supply chains, the entire Örebro/Jönköping manufacturing corridor). We have strong digital infrastructure and high trust in institutions. But our vocational and technical training pipeline, and our immigration policy for skilled technical labor, hasn't caught up to the pace our industry is automating at. SVT and Dagens Industri have both covered pieces of this, but nobody's connecting it to the global AI policy conversation, because it doesn't fit the narrative either side wants. The AI safety crowd wants to talk about frontier model risk. The AI accelerationist crowd wants to talk about GDP upside. Neither wants to talk about the boring, unsexy fact that automation without operational readiness just creates idle capital and unfilled jobs at the same time.

The Compliance Trap

The EU's instinct, and it's a good instinct in principle, is to build guardrails before things break. But guardrails built for the wrong risk don't help. If the AI Act and its national implementations spend the next two years focused on model transparency, bias audits, and high-risk classification systems, while the actual crisis is "we automated our factories and now have a hollowed-out middle layer of technical operators," Europe will have won the compliance argument and lost the economic one. That's the trap. Optimizing for the paperwork you can measure while the operational gap you can't easily measure becomes the thing that actually slows growth.

Where This Goes: 2026 to 2031

Here's my honest read on the next five years, and I'll separate what I think is likely from what I think is a real tail risk worth preparing for.

First, the labor-technology gap gets worse before it gets better. Every sector that automates, manufacturing now, logistics next, professional services and knowledge work after that as agentic AI tools mature, will hit its own version of Örebro. The tech will outrun the operational capacity to run it. This is not an AI capability problem. GPT-5 class models, agent frameworks, and robotics are already good enough for most of what's being deployed. It's a change management and workforce problem, and those move at human speed, not model training speed.

Second, regulation stays reactive and mostly symbolic at the frontier level, because governments genuinely cannot move fast enough to write binding technical constraints on systems that change every few months, and because the companies building those systems have more technical staff than most regulatory bodies combined. Trump's order is the clearest evidence of this pattern, but the EU will hit the same wall eventually, just with more paperwork attached.

Third, and this is the part that matters for AGI trajectory: if and when we get systems that can genuinely operate with agentic autonomy across long horizons (not just chat, but actually running workflows, managing other agents, making judgment calls), the operational gap doesn't close. It moves. Instead of "we have the tech but not the humans to run it," it becomes "we have the tech but not the humans who understand what the tech is deciding." That's a governance problem inside companies, not just a labor market problem. Boards and leadership teams that don't build internal literacy on how their AI agents actually work will be flying blind, regardless of what any executive order says.

Fourth: the countries and companies that win this decade won't be the ones with the best model access. Model access is becoming commoditized fast, look at how quickly open frameworks like AutoGPT and Langflow have made agent orchestration accessible to any team with an engineer and a weekend. The winners will be the ones who solved the operational and organizational problem first. Sweden, precisely because it's living the pain early, has a real shot at figuring this out before the US and most of the EU even admit it's the actual problem.

What to Actually Look At

If you're a founder or technical leader trying to get ahead of this instead of reacting to it, here's where I'd point you right now:

• Claude Code — if your engineering org isn't already testing agentic coding tools inside real workflows, you're behind on the thing that actually matters, which is building internal muscle memory for how agents fit into daily operations, not just marveling at demos.
• Langflow — for teams that want to prototype agent workflows without building orchestration from scratch. Good way to test whether your org actually has the operational maturity to run agents before you commit budget to a full build.
• awesome-llm-apps — a solid map of what's already shipping in production. Useful for benchmarking whether your "AI strategy" is actually ahead of the curve or just catching up to what a solo developer built in a weekend.
• Your own org chart. Seriously. Before you buy more AI tooling, map who on your team can actually operate what you already have. If the answer is "nobody, we'll hire for it later," you're building your own Örebro.

What to Actually Do About It

Stop treating AI adoption as a procurement decision and start treating it as an operational readiness project. That means three things, concretely.

One: before you deploy agentic systems into a workflow, identify who owns that workflow's outcomes and make sure they understand what the agent is doing well enough to catch it when it's wrong. This is the same discipline we bring into every AI agent development engagement at HEIMLANDR. The agent is never the hard part. The handoff to a human who can supervise it is.

Two: if you're building or buying custom AI solutions, build the training and operational documentation into the project scope from day one, not as an afterthought after go-live. We treat this as part of every serious rapid MVP build, because a system nobody can run six months after launch isn't a system, it's a liability sitting on your balance sheet.

Three: if you're a manufacturer or industrial company reading this and recognizing Örebro in your own operation, don't wait for the labor market to fix itself. It won't. Build automation and workforce planning as one project, not two separate budgets that never talk to each other. That's the actual lesson from this month's headlines, and it's cheaper to learn it now than after you've sunk capital into machines nobody can run.

The Real Signal

Trump's order and Örebro's factories look unrelated. Different continents, different scale, different actors. But they're the same failure, told twice. Institutions and companies keep optimizing for the visible win, a signed order, a completed automation rollout, while the invisible bottleneck, human operational capacity, quietly becomes the thing that actually decides who wins this decade. Washington will keep debating oversight. Brussels will keep drafting frameworks. Neither of those fixes what's happening in Örebro. Only building the operational layer, the people, the training, the internal literacy, fixes that. That's the work. Everything else is noise.

Fredrik Brunnberg is the CEO of HEIMLANDR.IO, building AI and software solutions from Jönköping, Sweden. This is the daily HEIMLANDR briefing. If you found this valuable, share it with someone who builds things.