Andreas Timm

Tag: AI governance

  • EU AI Act delay: 24 months for Brussels, 64× for AI

    EU AI Act delay: 24 months for Brussels, 64× for AI

    For the EU, it’s 24 months. For AI, it’s 64×.

    Last Wednesday the EU pushed the AI Act’s hardest deadlines back. Sixteen months for one piece. Twenty-four months for another. Read in regulatory time, that’s a reasonable phased rollout. Read against AI’s own pace of change, it’s something different.

    Exponential curve labeled 1× at Aug 2026 rising to 64× at Aug 2028, headline reads 'When the rules apply, AI is 64× more capable', subtitle 'EU AI Act high-risk deadline vs the AI doubling curve'.
    When the EU’s heaviest AI rules finally apply in 2028, the systems being regulated could be 64× more capable than the ones the rulebook was written for.

    What the EU just decided

    The AI Act is the world’s most demanding rulebook for artificial intelligence. It applies to any company that sells AI to European users — based in Europe or not. It was passed in 2024. Most of it was supposed to start applying in August 2026.

    Last Wednesday, the Council and Parliament agreed to push two of the heaviest pieces back.

    The “high-risk” category is the part most companies care about. It covers biometrics, hiring software, medical AI, AI in critical infrastructure — anything where a bad model output can hurt someone. Under the old timeline, these systems had to be fully compliant by August 2026. Under the new timeline, that becomes December 2027 (sixteen months later) for standalone systems, or August 2028 (twenty-four months later) for AI built into machinery, medical devices, and connected cars.

    Two-column comparison: BEFORE shows a single Aug 2026 deadline bar in grey, AFTER shows two new bars Dec 2027 plus 16 months and Aug 2028 plus 24 months in navy.
    The May 7, 2026 simplification agreement: one August 2026 deadline becomes two later deadlines, sixteen and twenty-four months out.

    What didn’t change matters too. The outright bans (social scoring, manipulative AI, untargeted face scraping) have been live since February 2025. The rules for big AI models — what most people call “frontier AI” — have been live since August 2025. The transparency obligations actually got tighter: providers of generative AI now have three months instead of six to ship watermarking. And a new ban on non-consensual sexual deepfakes lands hard on 2 December 2026.

    So the substance is intact. The triage is on the timeline.

    What METR actually measures

    METR is a research group that measures one specific thing about AI systems: how long they can keep working on a task before the workflow falls apart. Not how smart they are. Not how creative. How long they can stay on track without a human stepping in.

    The way they test it is straightforward. Give a model a real-world task — write a piece of code, run an analysis, debug a system — and measure the time-equivalent of work it can complete on its own. GPT-2 could chain together a few seconds of useful work. Claude 3 Opus held a few minutes. The frontier 2026 generation pushes past an hour.

    Plotted against time, that line is a clean exponential. From 2024 through early 2026, the time-horizon roughly doubled every four months.

    Exponential curve with three points: GPT-2 seconds at lower left, Claude 3 Opus minutes in middle, Frontier 2026 over an hour at upper right, headline 'Doubles every ~4 months', source METR.
    METR’s measurement of how long AI systems can work autonomously. The horizon roughly doubled every four months from 2024 through early 2026.

    Other measures point the same way. Reasoning depth, tool use, multi-step planning, software-engineering benchmarks — every adjacent curve has bent the same way over the same window. METR’s number is the cleanest single proxy I’ve seen, but it’s not an outlier.

    What 64× actually means

    If the doubling holds, the math on the EU’s new deadlines is uncomfortable:

    1. 16 months — four doublings — 16× more capable systems by the December 2027 deadline
    2. 24 months — six doublings — 64× more capable systems by the August 2028 deadline

    64× is not a metaphor. It’s the order-of-magnitude estimate of how much more autonomous task length AI can sustain by the time the EU’s heaviest rules apply.

    To put that in plain terms: if a 2026 model can do a one-hour task on its own, a 2028 model on the same trend can do a 64-hour task. A system that holds a workflow together for 64 hours is a different kind of object than the one the AI Act was drafted to regulate.

    That’s not an argument the rules are wrong. It’s an argument the gap between what the rulebook describes and what the system can actually do widens fast — faster than any 2-3 year drafting cycle can keep up with.

    My read

    My read on this: the headlines called May 7 a Brussels cave to industry pressure. I don’t think that’s the right frame. The substance of the Act is intact — the Commission could have used the simplification to weaken the high-risk classification or gut the impact-assessment requirement. They didn’t. They tightened transparency and added a new prohibition. The triage is on the timeline, not the rules.

    By 2028, the AI Act could be regulating systems 64× more capable than what existed when its rules were written.

    My expectation is that the August 2026 cliff was always going to slip. What’s more interesting is what the slip exposes: regulators and AI now run on incompatible clocks, and there’s no obvious mechanism to reconcile them. The Act assumed a 2-3 year drafting cycle would land on systems recognisably similar to the ones it described. That assumption broke somewhere between GPT-4 and the agentic generation that followed.

    Three things I’m watching

    • The 2 August 2026 deadline for national authorities. That date didn’t move. If most countries still don’t have working AI authorities by August, December 2027 becomes the next deadline at risk.
    • The European technical standards. Without finalised standards from the standards bodies, “high-risk” is a definition without a benchmark. Whether the Commission publishes them before the new deadline is the gating item.
    • The EU-US-UK divergence. The same week the EU softened its timeline, the US signed pre-launch testing agreements with the five frontier labs through CAISI. These two regulatory paths now point in different directions, and that gap is where the next year of this story plays out.

    One last thought

    Sixteen months. Twenty-four months. In any other regulatory context, those numbers feel reasonable. In AI they feel like an era. That’s not a problem the Commission can solve in a single omnibus.

    To be clear I am not asking for more regulation, I am asking for more decision speed!

  • Model deprecation is the new continuity risk

    Model deprecation is the new continuity risk

    Four rectangles in a row with the leftmost ghosted, simple connecting arrows
    A — model lifecycle row.

    OpenAI announced the discontinuation of the Sora web and app experiences on April 26, with the Sora API following on September 24. The first deprecation triggers in two weeks. Enterprises that built workflows on Sora since launch are not facing a model upgrade — they are facing a workflow rebuild on a four-month timeline. This is the first prominent enterprise-facing AI deprecation event of the cycle, and the precedent it sets matters more than the specific product involved.

    Model deprecation is no longer a developer-tier concern. It is an enterprise governance question that deserves a place on the risk committee agenda. The real shift is happening here: AI dependency without continuity is becoming a board-level risk in 2026.

    The shift: dependency without continuity guarantees

    The pattern of the past two years has been to build agent workflows on whichever foundation model was demonstrably best at the time, with little contractual commitment from the model provider about how long that model would remain available. Provider terms have improved — Azure OpenAI’s twelve-plus-six-month commitment for generally available models is the strongest standard in market — but most enterprises have not negotiated equivalent terms with their chosen providers. They built on capability, not on continuity.

    When the provider sunsets the model, the enterprise’s options are bad. Migrate to a successor model that may behave differently in subtle ways — requiring re-validation of every governed use case. Renegotiate at the eleventh hour for extended access at unfavorable terms. Or absorb the operational disruption of the workflow simply not working until rebuilt.

    The Sora event is small in dollar terms but large in precedent. The next deprecation will involve a more enterprise-critical model, and the enterprises that did not see this one coming are not going to see that one coming either.

    A single thread connecting a workflow box to a model box, the thread visibly fraying near the model with a clock above
    Built on capability. Not on continuity.

    The role change is the addition of an AI continuity discipline

    Inside enterprises that take this seriously, a discipline is emerging that did not exist in 2024 — AI continuity management. The work overlaps with vendor management, with disaster recovery, with model risk management, and with regulatory compliance, but it is structurally distinct from all of them. The discipline involves maintaining an inventory of model dependencies by workflow, negotiating continuity commitments at procurement, running successor-model regression tests on a regular cadence, and ensuring that the documentation chain meets the rebuild-readiness standard.

    Most enterprises have not staffed this discipline. The accountabilities are scattered across teams that do not coordinate. The procurement team negotiated the model contract a year ago without a continuity clause. The deployment team is building production dependencies on the model without thinking about migration cost. The risk team has not flagged model deprecation as a category. When the deprecation announcement lands, the company finds out it has no plan.

    The fix is straightforward in concept and slow in practice. Add continuity commitments to the procurement template. Build a model-dependency inventory. Designate an owner for AI continuity at the executive level. Run quarterly successor-model tests. None of this is hard. It is just unglamorous work that does not get done unless someone owns it.

    The strategic consequence is renewed buy-versus-build math

    Continuity risk changes the calculus of where to deploy AI capability. For workflows where the cost of unplanned migration is high — regulated workflows, mission-critical operations, customer-facing experiences with high switching costs — the case for either fine-tuning a frontier model into a controlled deployment, partnering with a vendor offering enterprise-grade continuity commitments, or building on open-weight models the enterprise can host indefinitely is stronger than it was in 2024. The case for relying on whichever model is best on a benchmark this quarter is weaker.

    The math is not simple. Open-weight models lag the frontier, sometimes meaningfully. Self-hosting carries operational cost that the proprietary providers absorb. The vendor lock-in to a single proprietary provider, even with the best continuity terms, is a different kind of risk than open-weight self-hosting carries. Each enterprise has to make this trade-off based on the workflow’s tolerance for capability lag versus its tolerance for continuity disruption.

    What is no longer defensible in 2026 is treating model continuity as someone else’s problem. The Sora sunset is small. The next one will not be.

    So what boards should do this quarter

    Add model deprecation to the risk committee agenda. The first deprecation event lands in two weeks. The board should at minimum understand which workflows are exposed and what the migration plans are.

    Demand a model-dependency inventory. Which workflows depend on which models from which providers, with which contractual continuity commitments. If this inventory does not exist, building it is the priority.

    Reconsider the buy-versus-build posture for mission-critical AI workflows. The 2024 default — use whichever proprietary model is best — was rational at the time. In 2026, with the deprecation precedent now visible, that default deserves an explicit reconsideration. Continuity is becoming a form of resilience. The boards that price it in this quarter will not be the ones rebuilding workflows under deadline.

    References and links

  • MCP became infrastructure and Apple decided to rent cognition

    MCP became infrastructure and Apple decided to rent cognition

    What was announced

    Two announcements in the week of March 2–8, 2026 redrew the agent landscape. Anthropic’s Model Context Protocol crossed 97 million installs, with every major AI provider now shipping MCP-compatible tooling — moving the protocol from experiment to default infrastructure for tool-calling agents. Apple confirmed that the redesigned, AI-powered Siri targeted for release alongside iOS 26.4 will be powered by Google’s Gemini model running on Apple’s Private Cloud Compute. In parallel, Anthropic rolled out memory features to all Claude users and deployed Opus 4.6 as an add-in inside Microsoft PowerPoint and Excel.

    What it means

    The MCP install count makes the connectivity layer between agents and tools a solved problem at the standards level. That is a meaningful shift. For two years, the friction in shipping agents was that every tool integration was bespoke; the integration debt scaled linearly with the number of tools and the number of agents. With MCP at default-infrastructure scale, the integration cost is closer to fixed than linear, and the bottleneck moves from connectivity to orchestration and governance.

    Apple’s decision to rent cognition from Google for Siri is the more strategically loaded story. It signals that even the most vertically integrated consumer-tech company in the world has concluded that building competitive frontier-model capability inside the company is not the right capital allocation. The Private Cloud Compute envelope handles the data-sovereignty argument. The Gemini choice handles the capability argument. The combination is an explicit acknowledgment that frontier-model capability has consolidated at a tier of providers most companies will rent from, not build alongside.

    Andreas’s view

    My read on this: the agent stack is settling into a recognizable shape. Standards layer (MCP, becoming generic). Frontier-model layer (a small number of providers — OpenAI, Anthropic, Google, with regional players underneath). Application layer (where most enterprise value is created). The interesting strategic action for the next 24 months is in the application layer, where the questions are which workflows to embed, which data to expose, and which orchestration logic to own.

    I don’t think Apple’s choice is anomalous. It is the start of a wave. Companies that have been building internal frontier-model capabilities will increasingly find that the math does not work — the capex is consumer-internet scale, the talent is concentrated at three or four employers, and the capability gap to “good enough internal model” widens every six months. The economically rational answer for almost everyone is: rent the cognition, own the integration and the data envelope around it. Apple has now made that a defensible board-level position.

    The way I see it: the most important architectural question right now is whether the cognition layer (rented, frontier-model, expensive but improving exponentially) is clearly distinguished from the integration layer (owned, workflow-specific, where the moat actually lives). Where those layers are blurred, I’d expect companies to find themselves overpaying on one side and under-investing on the other. The Apple-Google deal is the clean reference architecture for how that separation can look.

    Three things I’m watching

    Three things I’m watching as this plays out:

    1. I’ll be watching whether companies architect the cognition layer and the integration layer separately — treating frontier-model providers as utilities while building proprietary infrastructure around workflow integration and the data envelope.
    2. The companies that preserve optionality will be the ones that default to MCP-compatible tooling for new agent integrations. The standards layer is no longer a strategic differentiator — the question is how quickly organizations stop treating it as one.
    3. I’ll be watching how internal frontier-model build efforts hold up against the Apple-Gemini reference case. Where differentiation rests on owning the model, I’m interested to see whether those bets come with a credible 36-month capex and capability projection — and what happens when they don’t.

    References and related signals

    • Crescendo AI: latest AI news and developments
    • Related signal: Anthropic’s Opus 4.6 PowerPoint and Excel integrations move frontier-model capability deeper into the enterprise default tooling, accelerating the rented-cognition pattern.
    • Related signal: NVIDIA GTC 2026 (March) emphasized agentic frameworks and Fortune 500 production deployments — the application layer is where the next wave of enterprise AI value is being created.
    • Related signal: 95% of generative AI pilots still fail to reach production. The connectivity layer being solved does not solve the operating-model layer.
    • Related signal: Apple choosing Gemini over OpenAI for Siri changes the competitive math for every enterprise still scoping a frontier-model partnership.
  • The pilot-to-production gap is an execution problem, not a model problem

    The pilot-to-production gap is an execution problem, not a model problem

    What was announced

    Through the week of February 9–15, 2026, the enterprise AI deployment story sharpened around a paradox: 95% of generative AI pilots still fail to reach production, yet 42% of enterprises now run agentic AI in production and 72% have agentic systems live in production or pilot. Microsoft’s February enterprise update reframed Copilot from “assistant” to “governance-first agent” capable of completing entire workflows. Oracle introduced Fusion Agentic Applications for finance, supply chain, and HR. OutSystems research released the same week reported that 94% of enterprises adopting agentic AI now flag agent sprawl as a primary concern.

    What it means

    The two statistics are not in conflict. They describe two different populations of organizations. The 95%-pilot-failure number describes how the average enterprise treats generative AI: a proof-of-concept budget, a small team, and a handoff to operations that never happens. The 42%-in-production number describes a smaller cohort that has done the operational work — governance, identity, runtime monitoring, rollback procedures, and explicit ownership of the agent fleet. The gap between the two cohorts is not technical. It is procedural.

    Microsoft’s “governance-first agent” framing acknowledges this directly. The next phase of enterprise AI is not better models. It is the operating discipline around models — who deploys them, who owns them when they misbehave, who pays for the inference, and how the organization rolls back a bad agent without disrupting downstream work. That is a CIO problem, not a CTO problem.

    Andreas’s view

    My read on this: the production cohort is pulling away from the pilot cohort, and the gap is widening every quarter. The companies in production are accumulating an operational learning curve — what governance looks like, how to staff agent operations, how to track agent behavior in production, how to compose agents into workflows without losing accountability. The companies still iterating on pilots are accumulating learnings about prompts and demos. Those are different skill sets and they compound at different rates.

    I don’t think the next 12 months reward the companies that pick the best model. They reward the companies that figured out how to operate any reasonable model at production scale, with controls, with monitoring, and with an explicit chain of accountability when an agent does the wrong thing. Agent sprawl is the leading indicator that the operations layer is missing — when 94% of practitioners flag it as a top concern, the conversation has moved past whether agents work and onto whether they are manageable.

    The way I see it: the clearest signal a board can get on where an organization actually stands is whether the CIO can produce a production agent inventory — by name, by owner, by usage volume, by incident count. If the question produces a list, the organization is in the production cohort. If it produces “we are still piloting,” it is in the failure cohort, and the strategic gap to peers will be visible in operating costs by mid-2027.

    Three things I’m watching

    Three things I’m watching:

    1. I’ll be watching whether companies can produce a named, owned, monitored agent inventory with rollback procedures on demand — that capability is the clearest proxy I have for whether a real agent operating model exists or not.
    2. The organizations that interest me are the ones shifting pilot evaluation from “did the demo work” to “did the agent ship to production with controls in place” — and backing that shift by defunding pilots that stay in demo mode past a fixed time-box.
    3. The question I’d be asking myself is whether a dedicated agent-operations lead — with explicit authority over the production fleet and seniority equivalent to the head of enterprise systems — is in place. Without single ownership, sprawl is the default outcome, and I expect that to show up clearly in incident and cost data over the next several quarters.

    References and related signals

  • Davos 2026 made AI sovereignty the policy line — and the corporate one

    Davos 2026 made AI sovereignty the policy line — and the corporate one

    What was announced

    The World Economic Forum 2026 met in Davos January 19–23 with AI as the dominant agenda item. The conversation converged on three themes: risk-proportionate governance, runtime governance for multi-agent systems, and what Microsoft CEO Satya Nadella framed as “corporate AI sovereignty” — firms owning the intelligence layer that encodes their distinctive capabilities. Anthropic CEO Dario Amodei warned the forum that frontier AI is uniquely well-suited to autocracy, calling for targeted chip-export controls. A WEF press release on the same week reported leading organizations are shifting from “potential” to “performance” — measuring AI by realized output rather than pilot count.

    What it means

    The vocabulary shift is the substantive event. For two years, AI policy discussion at this forum was framed as risk management — what to restrict, what to monitor, what to ban. The 2026 framing is different. It treats AI as critical infrastructure where the governance question is who owns it, not whether it should exist. “Sovereignty” applied to AI is a deliberate echo of “data sovereignty” — a recognition that the layer of intelligence inside an organization is becoming as load-bearing as its data layer was a decade ago.

    For governments, this redirects policy from rule-writing to capability-building: domestic compute, domestic foundation models, controlled exports. For corporations, it redirects strategy from procurement to capability ownership: which models do you fine-tune yourself, which workflows encode your tacit knowledge, and which partners do you let inside the trust boundary. Both translations point to the same architectural question: where does the irreducible cognitive core of your organization live, and who can take it from you.

    Andreas’s view

    My read on this: Davos is a leading indicator of where C-suite vocabulary moves over the next 12 months. “Corporate AI sovereignty” is not a slogan — it is a framing that makes specific decisions easier to defend in a board meeting. Building your own model fine-tunes is sovereignty. Choosing not to send your customer interactions through a third-party model API is sovereignty. Maintaining a private inference cluster is sovereignty. The vocabulary justifies budgets that previously read as duplicative or paranoid.

    I don’t think the sovereignty framing is purely defensive. There is a competitive argument inside it: organizations that operate as pure consumers of frontier models are paying rent on the cognitive layer of their own business. Organizations that operate as owner-operators of a fine-tuned, workflow-embedded intelligence layer pay less rent and accumulate a moat that compounds with their data. The Davos talking points are starting to reflect that distinction.

    The way I see it, the question that matters this quarter is not “what is our AI strategy” but “what would it take to lose access to our primary model provider, and what would happen to the business if we did.” If the answer is catastrophic, the sovereignty argument is operational, not philosophical, and it has a budget implication.

    Three things I’m watching

    1. I’ll be watching whether companies run model-dependency stress tests — simulating the operational impact of losing their primary frontier-model provider for 30, 90, and 180 days. The result is the size of their sovereignty problem, and whether they even know that number tells me a lot.
    2. The companies that preserve strategic optionality will be the ones that draw a clear line between work requiring owned cognition (fine-tuned, embedded, internal) and work that can run on rented cognition (API-served frontier models) — and treat that boundary as a capital decision, not a procurement decision.
    3. I’ll be watching how the policy direction develops across major operating jurisdictions. Sovereignty framing in Davos has a consistent track record of translating into sovereignty requirements in regulated industries within 12–24 months.

    References and related signals

  • When 88% of organizations have adopted AI, adoption stops being the question

    When 88% of organizations have adopted AI, adoption stops being the question

    What was announced

    The Stanford HAI 2026 AI Index landed in mid-January with a set of numbers that close out a debate. Organizational AI adoption reached 88% globally. Global corporate AI investment more than doubled in 2025 to $581.7 billion. Generative AI hit 53% population adoption within three years — faster than the personal computer or the internet. Four out of five university students now use generative AI as part of their coursework.

    What it means

    When adoption crosses the 80% line, the question of “should we adopt” becomes structurally uninteresting. Every relevant comparison group has already answered it. What remains is differentiation — and differentiation in a world of universal access is harder, not easier, than in a world of selective access. The strategic margin moves from access to integration depth, from licenses to workflow penetration, and from procurement decisions to operating-model decisions.

    The investment number is the more telling signal. $581.7 billion of corporate AI investment in a single year is a capital allocation that prices in a specific belief: that AI capability will compound at a rate that makes today’s spending the cheap option in retrospect. That belief either turns out to be correct, in which case the laggards face a permanent gap, or it overshoots, in which case the survivors of the correction still own infrastructure and skills the laggards do not.

    Andreas’s view

    My read on this: the AI Index numbers are not a celebration of momentum, they are a notice of obsolescence. Adoption was the entry-level metric — the one that let companies say “we are doing AI” without committing to anything that mattered. With 88% adoption, that metric is exhausted. The companies that conflate “we have AI deployed” with “we have an AI strategy” will be the ones surprised in 18 months when peers with the same headline adoption rate are operating at a fundamentally different unit-economics base.

    I don’t think the next two years will be about adopting more. They will be about routing work differently — deciding which functions become AI-native, which roles get redesigned, which middle-management layers compress, and which workflows get rebuilt from the ground up rather than augmented. The companies treating this as a tooling question will keep the org chart they had in 2024 and bolt assistants onto it. The companies treating it as a structural question will redesign for AI-native operations and harvest a different cost base.

    My expectation is that boards still reporting on adoption rates are measuring the wrong thing entirely. The number that matters is the percentage of work routed through AI-native processes versus AI-augmented legacy processes. Those are two different cost structures and two different competitive positions. The first is a step change. The second is a feature.

    Three things I’m watching

    1. I’ll be watching whether companies move away from adoption KPIs toward integration-depth KPIs — specifically, the percentage of revenue-generating workflows that are AI-native, not just AI-touched.
    2. The companies that stand out to me will be the ones that build the comparison the AI Index doesn’t make for them: how their spend per FTE on AI infrastructure and tooling stacks up against the 90th-percentile peer in their sector. If that number isn’t visible to leadership, it isn’t informing strategy.
    3. I’ll be watching whether organizations use the next 12 months as a workflow-redesign window rather than a tooling-procurement window. The structural opportunity narrows the moment competitors finish their redesign.

    References and related signals