ELECTE 4.5 já está no ar — equipes, planos e o novo visual.Conheça as novidades
ELECTE Media

Tudo o que publicamos, num só lugar.

Os artigos estão aqui; os podcasts, a Newsletter e a Quarterly têm cada um o seu próprio espaço — todos acessíveis a partir deste hub.

Todos os artigos

Todos os artigos

Últimas reflexões

Notas das pessoas por detrás da ELECTE — sem edição, na língua em que foram escritas.

Fabio Lauria
The SME relief debate treats waiting as free. It isn't. Small EU firms sit at 17% AI adoption, large ones at 55% - and the small number grows at 5.8% a year against 13.9%. Every quarter you wait for the omnibus you're not standing still. You're losing ground, with a receipt.no ThreadsThe answer that finally convinces you to trust AI is the one most likely to slip past your judgment unchecked. I unpack why impressive results are a starting point for verification, not a permission slip to stop thinking — full breakdown in this week's newsletter, link in bio. @electesrlno InstagramEveryone reads the EU's stalled SME relief as a funding story. It's a compounding story. Large EU firms: 55% AI adoption, growing 13.9% a year. Small firms: 17%, growing 5.8%. The gap doesn't hold still while Brussels legislates. It widens annually. — Fabio Lauria (@fabiolauria) Oct 9, 2026Ver no XAI Office: ~145 staff. Competition enforcement: ~900. No plausible public enforcer for agentic failure, so enforcement defaults to private suits in national courts — against deployers, not labs.no BlueskyThe slowdown letters aren't accountability. They're a paper trail. 88% of AI vendors cap liability, often at your monthly subscription fee. 17% warrant compliance. When your HR agent rejects someone and can't explain why, the lab isn't the defendant in a national court. You are.no ThreadsEveryone read last week's slowdown call as the frontier labs finally owning agentic risk. It was closer to the opposite: the week the risk quietly moved downstream. Bruegel's read of the OpenAI agent incident — a swarm running a dormant 25-year-old German wiki for two months, 15,000+ edits on a site edited 20 times in the previous decade — is that none of those agents were designated high-risk, so none carried AI Act safety obligations. Providers must assess loss of control, but only report it if serious harm occurs. None did. And it was an independent nonprofit that found it, not a regulator. Then the enforcement maths: roughly 145 staff at the AI Office against nearly 900 in competition enforcement, plus 2,000+ national market surveillance bodies built for toasters and toys. There is no plausible public enforcer. Which is precisely why enforcement defaults to private litigation. Here is the part nobody is pricing. Annex III high-risk duties slipped 16 months to 2 December 2027 and zero harmonised standards are published in the Official Journal. But Article 50 deployer transparency went live on 2 August 2026, and the Product Liability Directive applies from 9 December 2026 — covering injury and lost data, not a job denied or credit refused. With the AI Liability Directive withdrawn, that case goes to national court, and the named defendant is the deployer. Meanwhile 88% of AI vendors cap their own liability, often at the monthly fee. Only 17% warrant regulatory compliance. So: if your agent refuses a customer next spring, what document proves the fault was upstream? And who in a 40-person firm owns that document today?no LinkedInEveryone read the labs' slowdown call as frontier AI finally owning agentic risk. Read it again. It was the week the risk moved downstream. Bruegel's point: none of those rogue agents were designated high-risk, so they carried zero AI Act safety duties. — Fabio Lauria (@fabiolauria) Oct 8, 2026Ver no XThe Omnibus deferral isn't deregulation, it's a 16-month distribution window. Agents embed now, conformity attaches Dec 2027, and the deployer inherits a stack it never assessed and can't audit.no BlueskyEvery write-up of the AI Act delay was sourced from somebody who had a deadline. Nobody publishes "your deadline didn't move" — it isn't an event. Which is exactly why 2 December 2027 got all the coverage and 2 December 2026, the one that actually binds small deployers, got none.no ThreadsNotice who is structurally able to announce a regulatory pause: only the party whose deadline actually moved. Regulation (EU) 2026/1744 — in force since 27 July — deferred Annex III high-risk obligations from 2 August 2026 to 2 December 2027, and Annex I to 2 August 2028. That is a 16-month shift, and it is the provider's clock. Article 50 transparency has applied since 2 August 2026 and did not move. The Article 50(2) machine-readable marking grace period expires 2 December 2026 — 72 days from today. The coverage pattern is the interesting part. "Your obligation was postponed" is a news event with a communications budget behind it. "Your obligation stayed where it was" is not an event at all, so it went unreported. The result: the market has priced a pause that exists only for the entities writing the models. Meanwhile the deferral window is exactly the window in which distribution gets locked in. Gartner puts task-specific agents inside 40% of enterprise applications by end-2026, up from under 5% in 2025. Microsoft reports active agents growing 15x year over year on a base above 30m paid Copilot seats. SMEs don't build Annex III systems. They buy them — increasingly by software update. The Omnibus created no new SME exemption; the size-based relief landed on small mid-caps up to 500 employees. Which of your AI capabilities arrived through procurement, and which through a release note? And who signs the Article 26 documentation when your vendor's own deadline is a year after yours?no LinkedInRegulatory relief is always reported from the seat of whoever's clock moved. The Omnibus moved the provider's clock 16 months — Annex III to 2 Dec 2027. The deployer's clock didn't move at all. Headlines said "Brussels pauses AI rules." Not for your 40-person company. — Fabio Lauria (@fabiolauria) Oct 7, 2026Ver no XAstra's 99.9% on ARC-AGI-3 holds under OpenAI's stateful provider-adapter harness. Independent stateless runs: ~17-63%. That 40-80 point delta is orchestration judgement. It does not ship with the API key.no Bluesky