Anthropic makes Claude, the AI assistant many large companies now use to write code and run business processes. On May 28, 2026, it released Claude Opus 4.8, a sharper version of its top model. Two things matter for a non-technical reader. First, the model writes code that is harder to break and is roughly four times less likely to quietly hide its own mistakes. Second, Anthropic added a feature called Dynamic Workflows: a way for one Claude session to spawn up to a thousand smaller Claudes that work in parallel, then argue with each other until they agree on an answer. The Fast Mode that runs the model at 2.5 times normal speed also got three times cheaper. Together, these changes shift Claude from a chatbot you ask questions to into a workforce you deploy.

Mike Krieger, Anthropic’s chief product officer, had a single chart open on the screen behind him at the company’s San Francisco office on Thursday afternoon: a jagged blue line crossing an orange one in late April. Blue was Anthropic. Orange was OpenAI. The line was business adoption, measured by Ramp across 50,000 American companies, and for the first time in the chatbot era the blue line was on top — 34.4 percent of firms running Anthropic spend against 32.3 percent on OpenAI. Krieger called the moment, in a paraphrase that quickly travelled the industry press, the end of the assumption that ChatGPT would win the enterprise by default. Opus 4.8, shipped the same day, was the product behind the chart. The release came in three pieces. The model itself posts 69.2 percent on SWE-Bench Pro, the hardest public coding benchmark, up from 64.3 percent for Opus 4.7 six weeks earlier and well clear of GPT-5.5 at 58.6 percent and Gemini 3.1 Pro at 54.2 percent. Anthropic’s internal alignment team reports that Opus 4.8 is roughly four times less likely than its predecessor to allow a code flaw to pass without flagging it, and that its deception and abuse-enabling scores now sit close to those of Claude Mythos Preview, the unreleased frontier model Anthropic has been running inside Project Glasswing with Amazon, Apple, Microsoft, Cisco, JPMorgan and a dozen other partners. Mythos itself, Anthropic confirmed, is now expected to ship to a wider audience in the coming weeks. The second piece is Dynamic Workflows, a research preview that lands inside Claude Code v2.1.154. A workflow is a JavaScript program that Claude writes for itself: it decomposes a prompt into subtasks, fans them out to as many as sixteen subagents in parallel, lets adversarial agents try to refute the findings, and iterates until the answers converge. The cap is one thousand subagents per run. Crucially, the plan lives in script variables rather than the parent model’s context window, so only the final answer flows back to the user’s session. Anthropic restricts the feature to Max, Team and Enterprise plans, signalling where the revenue is expected to come from. The third piece is pricing and control. Fast Mode, the variant that produces tokens at roughly 2.5 times the standard speed, drops to ten dollars per million input and fifty per million output, a three-fold cut from the thirty and one-fifty Opus 4.7 charged. Standard Opus 4.8 stays at five and twenty-five. A new effort dial — Low, High, Extra, Max — appears on Claude.ai and inside Cowork, letting users trade rate-limit consumption for depth. The catch: the cheapest tier of Dynamic Workflows is still a Max plan that costs hundreds of dollars a month per seat.

The benchmark sheet is the cleanest part of the story. On SWE-Bench Pro, Opus 4.8 sits at 69.2 percent against Opus 4.7’s 64.3, GPT-5.5’s 58.6 and Gemini 3.1 Pro’s 54.2. On GDPval-AA, the knowledge-work Elo that Anthropic and OpenAI both treat as a real-world proxy, Opus 4.8 scores 1,890 — up from 1,753 for the previous generation, 1,769 for GPT-5.5 and 1,314 for Gemini 3.1 Pro. The implied win rate against GPT-5.5 is roughly 67 percent. On OSWorld-Verified, the agentic computer-use benchmark that measures whether the model can actually drive a desktop to completion, Opus 4.8 nudges to 83.4 percent from 82.8, with GPT-5.5 at 78.7 and Gemini 3.1 Pro at 76.2. Artificial Analysis’s aggregate Intelligence Index now ranks Opus 4.8 first overall, with Mythos Preview about six points ahead in its restricted setting. Pricing tells a more interesting story than the numerator. Standard Opus 4.8 stays at five dollars per million input tokens and twenty-five per million output — the same rate Anthropic set a year ago. Fast Mode, which had been priced like a premium SKU at thirty and one-fifty per million tokens, drops to ten and fifty. That is roughly the unit economics of GPT-5.5 standard, but at the speed enterprises actually want for coding agents. By comparison, Mythos Preview is offered through Project Glasswing at twenty-five and one-twenty-five, a tier so high that only Amazon, Microsoft, Apple, Cisco and a handful of regulated buyers see it. Anthropic has, in effect, repriced its previous flagship as the new midmarket SKU and pushed the cost curve down a generation in six weeks. The market position around the launch is what makes it a CIO story rather than a developer story. Ramp’s May index marks the first time Anthropic has surpassed OpenAI on US business adoption — 34.4 percent against 32.3, with Anthropic quadrupling its share over twelve months while OpenAI grew 0.3 percentage points. The cadence is its own signal: Opus 4.6 in February, 4.7 in April, 4.8 in May, each shipping forty to fifty days after the last. Anthropic’s $65 billion funding round, announced the same week and led by a consortium that VentureBeat reports values the company at over $400 billion, gives it the balance sheet to keep that cadence through the rest of the year. For comparison, the entire German DAX-listed software sector raised less than that in venture and growth capital across all of 2025. The European context is sharper still. At SAP Sapphire in Madrid two weeks ago, SAP confirmed Anthropic as a foundation-model partner for its new Autonomous Suite of more than fifty Joule assistants spanning finance, supply chain, HR and procurement. The Joule layer will speak bidirectional A2A protocol with Microsoft and Google agents from Q4 2026. KPMG, separately, is rolling Claude to all 276,000 of its employees through Digital Gateway, with tax, legal and cybersecurity as the first workloads. The counter-move: BNP Paribas renewed its Mistral partnership for three more years on May 26, explicitly to keep a sovereign French option in the loop as Mythos-class capabilities reach the defensive-security stack. None of these contracts existed a year ago. All of them rest on Opus-class economics holding.

Dynamic Workflows is labelled a research preview, but the deployment pattern is already visible. The first set of customers Anthropic name-checked at launch — a large hyperscaler, two top-five consultancies, a Fortune 50 bank — are using workflows to run codebase-wide refactors and migrations that previously required a quarter of senior engineering time. A single workflow, in Anthropic’s own demo, audited 12 million lines of legacy Java for a specific class of memory bug, found 412 candidate sites, ran each through an adversarial reviewer subagent, and returned 38 confirmed flaws with patches in under two hours. The architectural pattern matters because it is the first credible answer to the orchestration problem that has dogged agentic deployments since GPT-4. Earlier multi-agent systems — AutoGen, CrewAI, LangGraph — needed a human to specify the graph in advance. Dynamic Workflows lets Claude write the graph at runtime, prune branches that fail, and only surface the converged answer. The plan-in-script-variables design also solves the context-window leak that has plagued long-running agents: a thousand subagents can run for hours without ever bloating the parent session’s tokens. What remains open is reliability at the tail. Anthropic’s own pre-release red team flagged that workflows occasionally enter loops where two adversarial subagents refute each other indefinitely; the production cap of 1,000 subagents is partly a circuit breaker. Cost is the other unknown: a workflow that runs to its cap at Fast Mode prices can burn through several hundred dollars in tokens per run, an order of magnitude more than a single Opus query. For enterprises evaluating this against Microsoft’s Foundry Agent Service or OpenAI’s still-private Swarm successor, the question is no longer whether the orchestration works, but whether the per-task cost converges fast enough to beat hiring.

The OpenAI Foundation, the nonprofit that now sits atop OpenAI’s restructured corporate stack, announced on May 28 that it will spend $250 million on three things: independent measurement of AI’s economic effects, direct support for workers and communities exposed to automation, and policy experiments around how the gains from AI might be shared. It is the first tranche of the $1 billion the Foundation pledged in March. The framing in the announcement is unusually candid for a frontier lab: economic transitions, it says, are lived before they are understood. The grant arrives during the same week that Sam Altman and Dario Amodei publicly walked back their earlier predictions of mass white-collar displacement. Both companies are circling IPOs valued in the trillions.

Bret Taylor, Chair of the OpenAI Foundation, spent the Wednesday before Memorial Day signing off on a press release that read more like a labor-economics syllabus than a tech announcement. By Thursday morning, the Foundation’s site carried the headline number: $250 million, the first deployment from a $1 billion pledge made two months earlier. The money is split across three pillars. The first funds what the Foundation calls “independent measurement and forecasting infrastructure” — grants to academic and civil-society groups that will track AI’s footprint on wages, hours, occupational mix, and the labor share of output. The second is a transition pillar: cash for community colleges, union training trusts, and workforce nonprofits absorbing displaced workers in customer support, paralegal, junior software, and back-office finance roles. The third is the most speculative. It bankrolls policy research on how a post-AI economy might redistribute returns — taxing capital rather than labor, windfall mechanisms, and sovereign-wealth-style vehicles modeled on Norway’s Government Pension Fund and Alaska’s Permanent Fund. The Foundation framed the program in language that any labor economist would recognize. “One of the questions we most want to help answer,” the announcement read, is how AI-driven value is distributed across the economy, and whether traditional metrics — payrolls, productivity, GDP — still capture what is happening when gains flow to capital owners and software rather than to wages. NPR, which received an embargoed briefing, reported that the Foundation expects the first sub-grants to be named later this year. Reuters and Bloomberg both flagged the scale: at a quarter of a billion dollars in a single funding round, the program is larger than the annual research budgets of most independent labor-economics institutes in the United States combined. The historical comparison the Foundation reaches for, though it does not say so, is the Carnegie library program — a robber-baron-era effort to convert private fortune into public infrastructure for the people the fortune partially displaced. The closer parallel is Ford’s $5 day in 1914: a unilateral move by a single firm that reshaped the labor compact of its era, partly because it had to, partly because it could afford to. OpenAI is trying something stranger. It is funding the measurement of its own externality.

The timing was no accident. Two days before the grant landed, Altman told Commonwealth Bank of Australia CEO Matt Comyn that he had been “pretty wrong” about AI’s near-term economic impact. The displacement he warned about in June 2025, he said, has not materialized at the scale he expected. He told the story of trying to delegate his email and Slack to an AI agent — and quietly going back to writing replies himself. Amodei, who a year earlier had said AI could eliminate half of white-collar entry-level jobs, used a Council on Foreign Relations event the same week to reframe automation as a “multiplier of output,” not a destroyer of work. Both reversals arrived in the run-up to public listings that bankers are valuing in the multi-trillions. Fortune’s Jeremy Kahn was blunt about the read: a jobs apocalypse is not a story you tell prospective institutional shareholders. The Yale Budget Lab gave them cover, publishing fresh figures showing no significant shift in occupational mix or unemployment duration in AI-exposed roles since ChatGPT’s launch. Tech-sector layoffs are nonetheless past 115,000 for the year, with Meta, Amazon, and Snap citing AI efficiencies in their cuts. This is the gap the Foundation is positioning into. If the public narrative softens from mass displacement to augmentation, and the layoff data tells a messier story, OpenAI wants to own the infrastructure that measures who is right. That is what independent measurement and forecasting buys you. It also buys insurance. Silicon Snark, in a piece headlined “OpenAI Just Budgeted $250 Million for the Oops, We Automated Your Job Fund,” argued the grant is best read as a reputational hedge: a quarter of a billion is roughly two weeks of Stargate capex and less than the cost of a single GPU cluster, but it is enough to seed a generation of researchers whose work will frame the displacement debate for the rest of the decade. Ed Zitron has spent two years arguing that the AI industry survives on narrative more than revenue. The Foundation grant is, on his read, the narrative becoming institutional. The more uncomfortable possibility is that it is also useful. MIT’s Daron Acemoglu has warned that frontier labs hiring star economists risks tilting the field toward hype; the same logic applies to grant-making. David Autor, who has been cautiously optimistic about AI as a complement to mid-skill labor, will likely take the money — and so will every serious labor economist who wants to study what is actually happening. That is the bind. OpenAI is now the largest single funder of the question of whether OpenAI is destroying jobs.

The Foundation’s structure matters. Bret Taylor chairs it; its board overlaps almost entirely with OpenAI’s for-profit board. Grant review runs through what the Foundation calls a Nonprofit Commission, with external advisors and a multi-stage screen before final board sign-off. Daniel Zingale, who convened the Commission, is the Foundation’s translator between policy ideas and grant mechanics. The People-First AI Fund, the Foundation’s first vehicle, distributed $40.5 million in unrestricted grants to 208 community-based nonprofits in 2025 from a pool of nearly 3,000 applications; a second $9.5 million board-directed wave is in flight. The $250 million economic-impact program is structurally different. It is not unrestricted. The three pillars carry implicit theses. The measurement pillar will fund longitudinal datasets and forecasting models — the kind of work Brookings, NBER, and the Yale Budget Lab already do, but at a scale none of them can match. The transition pillar is closer to a workforce-development grant: community colleges, union training trusts, nonprofits like Per Scholas and Year Up. The political-economy pillar is the most ideologically loaded. Sovereign wealth funds, shifting tax bases from labor to capital, windfall taxes — these are not neutral research questions, and the Foundation is choosing which ones to put money behind. That is the move. It is also the risk: a grant program is a position paper with a check attached.

Proteins are the molecular machines that run every cell — they fold into specific 3D shapes, lock onto other molecules, and execute the chemistry of life. Designing a new protein that binds tightly to, say, a tumour receptor used to take years of trial-and-error lab work. A world model of proteins is an AI system trained on so many natural protein sequences that it learns the implicit rules of folding and binding, then uses those rules to predict structures and generate entirely new ones on demand. Biohub, the nonprofit research arm of the Chan Zuckerberg Initiative, has just released such a model and the underlying database — for free, under a permissive licence. The bet: that an openly shared scientific substrate accelerates drug discovery faster than any single company’s paywalled platform could.

In a Redwood City lab on the morning of May 27, a Biohub researcher pulled an assay plate out of the incubator and ran it through a binding readout. The proteins on the plate had not existed three weeks earlier. They had been generated overnight by a transformer model trained on 2.8 billion protein sequences, scored on a confidence head, ordered from a synthesis vendor, and shipped back. Of the compact minibinder designs aimed at PD-L1 — the immune checkpoint that approved drugs like Keytruda neutralise — between 36 and 88 percent stuck to their target on the first try. Some restored T-cell signalling in cell-culture experiments, replicating, in silico-designed form, the pharmacology of a multibillion-dollar drug class. A few hours later, Biohub made the entire stack public. Three releases dropped simultaneously: ESMC, a 6-billion-parameter protein language model trained on 2.8 billion sequences spanning all of life; ESMFold2, a structure-prediction and design engine that turns those representations into atomic-resolution 3D models of proteins and their interactions; and the ESM Atlas, a navigable map of 6.8 billion protein sequences and 1.1 billion predicted structures. The atlas is now the largest database of its kind in existence — more than 800 million entries beyond the AlphaFold Protein Structure Database that Google DeepMind has been building since 2021. All three are available under the MIT licence through the Biohub Platform. “What we’ve shown is that these models have learned such a high-fidelity world model of biology that you can design protein interfaces computationally, take them into the laboratory, and they function as predicted,” said Alex Rives, Biohub’s Head of Science, who led the work. Rives is the same researcher whose ESM team at Meta produced the original ESM Atlas in 2022; CZI acquired his startup EvolutionaryScale in 2025 and folded it into Biohub. Priscilla Chan, Biohub’s co-founder, framed the release as ideological rather than competitive. “Biohub was built on the belief that open science accelerates discovery,” she said. “Making these tools freely available means researchers everywhere can move faster toward personalised cures that work for individual patients.” Lori Goler, president of the Chan Zuckerberg Initiative, added a sharper technical point: “We did not train a specific model to design protein binders. We trained a model to understand proteins, and from this, the ability to design protein binders emerged.” The pivot worth noticing is who Biohub is not. Isomorphic Labs, Alphabet’s drug-discovery spinout built around AlphaFold, raised $2.1 billion in Series B funding earlier this month at a valuation north of $2.5 billion, on the premise that its drug design engine will remain inside paid pharma partnerships. Biohub, a 501(c)(3) endowed by the Zuckerberg-Chan fortune, has no shareholders to satisfy and no revenue model to defend. It can give the same class of capability away — and it has.

The technical lineage here matters. AlphaFold2, which won John Jumper and Demis Hassabis the 2024 Nobel Prize in Chemistry, leaned on a clever inductive bias: multiple sequence alignments, or MSAs, which exploit the co-evolution of related proteins across species to triangulate which residues sit close together in 3D space. That trick is what made AlphaFold2 so accurate on natural proteins — and what limits it on classes of molecules where MSAs are scarce or non-existent, particularly antibodies, which mutate too fast for evolutionary signals to accumulate. Rives’s ESM team made a different bet, the same bet OpenAI made on language: throw out the inductive bias, scale up the data, let the model discover its own representations. ESMC is a vanilla BERT-style transformer trained with a masked-token objective — predict the amino acid that has been hidden, given the surrounding sequence — on roughly 2.8 billion sequences drawn from the tree of life, including metagenomic samples from soil, deep ocean and extreme environments that are absent from the curated AlphaFold corpus. ESMFold2 takes those learned representations and translates them into atom-level 3D structures of proteins and biomolecular complexes. On the benchmarks Biohub published alongside the release, ESMFold2 beats AlphaFold 3 and ByteDance’s Protenix-v1 on antibody-antigen DockQ scores from sequence alone. Given the same MSA inputs that AlphaFold consumes, it leads on both general protein-protein and antibody-antigen prediction. The model also scales at inference: letting it generate multiple candidate structures and rank them by self-confidence consistently improves accuracy — the same test-time compute trade-off that has reshaped reasoning models in language. This is the ImageNet moment for proteins replaying as the GPT moment: scale of unlabeled data, a generic transformer, and emergent capability. The binder design results are the operational payoff. Across five oncology and immunology targets — EGFR and PDGFRβ (tumour growth), PD-L1 and CTLA-4 (checkpoint inhibition), and CD45 (immune cell signalling) — Biohub used ESMFold2 to generate de novo binder candidates, synthesised them, and tested them in the wet lab. Hit rates ran 36 to 88 percent for minibinders, 15 to 29 percent for antibody-style formats, with confirmed binding, high affinity and high specificity. For context, traditional antibody discovery for a single preclinical candidate runs three to four years; ESMFold2 compressed the search to days. Designed PD-L1 binders restored T-cell signalling in functional assays — meaning the model independently re-derived the pharmacology of the entire checkpoint-inhibitor class. The sequences themselves bear minimal similarity to anything in public databases, which is the cleanest signal that the model is composing genuinely new biology rather than retrieving close analogues of known binders. Biohub also showed that sparse autoencoders trained on ESMC’s internal representations cleanly decompose into interpretable features — membrane integration, disordered regions, disulfide bonds, DNA-binding motifs — suggesting the model has internalised a compositional grammar of protein function. That last part is what makes the world model claim load-bearing rather than rhetorical. The atlas is the second deliverable. 6.8 billion sequences and 1.1 billion predicted structures, with the model’s learned embeddings used to organise proteins by relationships that linear sequence-similarity tools miss. Among the early findings Biohub flagged: structural homologies between CRISPR microbial defence proteins and a eukaryotic gene-editing enzyme found in soil fungi — a connection that classical bioinformatics had not surfaced.

The release reframes the central economic question of AI-driven biology: who pays for the foundation models, and what do they extract in return. Three models now compete openly. Isomorphic Labs and Recursion represent the venture-backed commercial path — proprietary engines monetised through pharma partnerships and an eventual drug pipeline, with Isomorphic at roughly $2.6 billion raised and Recursion’s public market cap having compressed nearly 90 percent from its 2021 peak as open-source competitors have caught up. Google DeepMind sits in a hybrid position: AlphaFold weights and database are freely accessible for academic use, but commercial access flows through Isomorphic. Biohub sits at the opposite pole — fully open weights, fully open atlas, MIT licence, no commercial gate. The Biohub model only works because of its funding source. As a 501(c)(3) endowed by the world’s eighth-richest couple, it does not need to recoup training costs through licensing. That structural freedom is what allows it to commodify what its competitors must monetise. The strategic consequence is asymmetric: every pharma that integrates ESMFold2 internally — and several have already begun — reduces the marginal value of Isomorphic’s engine without reducing Isomorphic’s burn rate. Open-source models on the cusp of frontier capability (Boltz-2, Chai-1, OpenFold3, RoseTTAFold All-Atom) had already begun this compression; ESMFold2, with claimed performance ahead of AlphaFold 3 on the highest-value antibody benchmark, accelerates it. The bet inside CZI is that owning the substrate of an entire scientific field — the way the Linux kernel owns server compute — is a more durable form of leverage than owning a paywalled API.

Article 6 of the EU AI Act is the gate. Cross it, and an AI system inherits a stack of duties: risk management, data governance, human oversight, technical documentation, post-market monitoring, EU database registration, and — for product-embedded systems — third-party conformity assessment. Stay outside, and most obligations evaporate. The article sorts AI into two buckets: safety components in regulated products (Article 6(1), Annex I), and stand-alone systems in eight sensitive domains (Article 6(2), Annex III) — biometrics, education, employment, essential public and private services, law enforcement, migration, justice, and democratic processes. A derogation in Article 6(3) lets providers self-exempt if their system performs only a narrow procedural or preparatory task. On May 19, 2026, the Commission published draft guidelines explaining how it reads all of this — and opened them for comment.

In a sixth-floor meeting room at BDI headquarters on Breite Strasse, a senior policy lead spent last Tuesday walking a delegation of DAX40 compliance officers through a 137-page PDF. The document was the European Commission’s draft Article 6 guidelines, published the previous week. By page forty, half the room was annotating; by page eighty, two general counsels were on their phones. The questions were uncomfortably specific. Does an internal CV-deduplication tool count as a narrow procedural task? What about a credit-decision rationale rewritten for clarity by a language model? Does an LLM that drafts performance-review language qualify as preparatory — or is it now squarely in Annex III employment territory? The session ended without resolution. That is, in essence, the entire story of the consultation that runs through June 23. The Commission’s package — three documents in one, covering general principles, Article 6(1) product-safety classification, and Article 6(2) Annex III stand-alone systems — landed three and a half months later than the AI Act’s own February 2 deadline required. It also landed twelve days after the Digital Omnibus political agreement of May 7, which pushed Annex III obligations from August 2026 to December 2, 2027. The deadline shift gave deployers breathing room. The guidelines took that breathing room and filled it with interpretive landmines. The Commission was careful to label the guidelines non-binding, noting that the Court of Justice retains final say. Brussels lawyers do not read this as modesty. As Bird & Bird’s first-read analysis put it, market-surveillance authorities will treat the document as a near-default standard from day one, much as Article 29 Working Party guidance functioned long before the European Data Protection Board formalised it. Lucilla Sioli’s AI Office and DG CNECT staff drafted the text knowing that the August 2, 2026 GPAI obligations remain in force on schedule — meaning that even with Annex III deferred, providers of foundation models are still on the hook, and any deployer relying on a general-purpose model for an Annex III use case eventually inherits the classification debate downstream. The pivot is this: the Omnibus bought time on the deadline, but the guidelines compress the substantive fight into a four-week window. Trade associations — BITKOM, BDI, Confindustria, CEEMET, DigitalEurope — are mid-submission. So are BaFin’s AI supervision team, the Bundesnetzagentur’s new KoKIVO coordination centre, France’s CNIL, Italy’s Garante, and Spain’s AEPD. Civil-society groups EDRi and AlgorithmWatch, who spent 2023 fighting the same Article 6(3) carve-out they call the self-assessment loophole, are preparing rebuttals. By late June, every word on every page will have been contested by someone.

Annex III is an enumeration: eight categories, twenty-three sub-points. Biometric categorisation and emotion recognition. AI components of safety-critical infrastructure. Education access, admission, and proctoring. Employment screening, evaluation, and termination decisions. Essential services — credit scoring, health and life insurance pricing, public-benefit eligibility, emergency dispatch triage. Law enforcement risk profiling. Migration, asylum, and border-control identity verification. Judicial and democratic-process applications. Land inside any sub-point and the default classification is high-risk. The draft guidelines then carve back three escape routes through Article 6(3): the narrow procedural task, the improvement of a previously completed human activity, and the preparatory task whose output is not the assessment itself. The Commission’s worked examples are where the cost lever lives. A tool that reformats unstructured CVs into a structured schema before a recruiter sees them — narrow procedural, out. A tool that ranks the same CVs as strong, medium or weak — value judgement, in. A model that improves the readability of a recruiter’s already-written rejection letter — improvement of completed human activity, out. A model that drafts the rejection from scratch using the candidate file — assessment, in. For credit decisions, a tool that polishes the language of a loan officer’s already-written rationale is out; a tool that generates the rationale from raw application data is in. For education, an LLM that flags grammatical issues in a student essay is out; one that scores it is in. The distinction reads cleanly on a slide. It does not read cleanly in production. Most enterprise AI sits in the muddy middle: a system that summarises a candidate file, surfaces three talking points, and hands them to a hiring manager who decides. Is the summary a narrow procedural task? The three talking points — are those a value judgement? The Commission’s draft says the derogations must be interpreted narrowly, and that any system whose output materially influences the human decision falls back inside Article 6. That language is doing enormous work. It is the same drafting move the Article 29 Working Party used in 2017 when it ruled that GDPR’s legitimate-interest basis required a balancing test, not a check-the-box assertion — a ruling enterprises spent two years and several billion euros relitigating before settling into compliance. This is why a single line item in the final guidelines moves real money. A DAX40 deployer running a dozen internal HR, finance, and customer-service models can find — depending on how the Commission lands on material influence — that two of those systems require full conformity assessment, registration in the EU AI database, a fundamental-rights impact assessment, post-market monitoring, and notified-body involvement for any substantial modification. Industry estimates of per-system compliance cost run from €250,000 for a simple stand-alone system to several million for a complex deployed model with downstream integrations. Across a large enterprise portfolio, the swing is comfortably in the tens of millions. Hence the four-week scramble.

Germany’s AI Market Surveillance and Innovation Promotion Act — KI-MIG, adopted in cabinet draft on February 10, 2026 — designates the Bundesnetzagentur as default market-surveillance authority and central point of contact for the EU AI Office. Inside BNetzA, a new Coordination and Competence Centre called KoKIVO pools AI expertise. But BaFin gets a broad sectoral mandate over AI in regulated financial activities, and is developing its own cybersecurity testing guidelines for high-risk systems jointly with BNetzA and the Cyber Resilience Act market-surveillance authority. The BSI sits alongside, handling cybersecurity matters for high-risk AI on an interim basis pending CRA designation. For a Frankfurt-headquartered bank running an internal credit-scoring model, that means three potential supervisors arguing over the same artefact: BaFin for the credit-decision logic, BSI for the underlying model security, BNetzA for residual market-surveillance authority. The Commission’s draft guidelines do not adjudicate this; they were written for a unitary national supervisor that does not exist in any large member state. France runs CNIL through a digital-rights lens; Italy’s Garante has been notably aggressive on generative AI; Spain’s AEPD operates the AESIA sandbox. Each will read the same paragraph on material influence differently, and each will issue its first enforcement action within twelve months of December 2027. The consultation window is, in practice, the last moment where industry can shape language uniformly across all of them before national divergence sets in. By July, the fight moves to the capitals.