Two months after the EU AI Act's high-risk provisions took effect, the first enforcement actions are landing — and they are not aimed at the labs people expected. Meanwhile, agentic workflows quietly crossed a load-bearing threshold inside three Fortune 100 finance teams, and the open-weights community released a 70B model that genuinely changes the on-prem economics. This week is less about splashy launches and more about the boring middle layer finally catching up.
Italian Regulator Fines a Mid-Size Bank €18M for an Unaudited Credit-Scoring Model
The Garante's first AI Act enforcement action did not target OpenAI, Anthropic, or any frontier lab — it targeted a regional bank that deployed a gradient-boosted credit model in production without the conformity assessment the high-risk clause now requires. The fine is a rounding error for the bank; the precedent is not.
JPMorgan, Citi, and Visa Each Quietly Crossed 40% of Back-Office Tasks Handled by Multi-Step Agents
None of the three announced it. Each disclosed the figure inside an earnings footnote or a vendor onboarding blog post, and each qualified it the same way: human review of every output, escalation queue intact, error rate still above the human baseline on edge cases. The race is now from 40% to 60%, not from 0% to 40%.
Mistral releases Mixtral-8x22B-Omni, a 70B-Active MoE That Runs on a Single H200 Node
The release makes on-prem deployment of a frontier-tier multilingual model operationally possible for mid-size enterprises that cannot send customer data to a hosted API. Independent benchmarks within 48 hours matched or beat Llama 4 Scout on four of seven reasoning evaluations. The open-weights-versus-API decision matrix just got simpler — and more political.
Anthropic Publishes a Tool-Use Specification That Every Other Lab Quietly Adopts Within a Week
The new spec codifies how agents describe tools, handle multi-step errors, and surface confidence to the calling model. It is open, royalty-free, and explicitly designed to be forkable. Within seven days, OpenAI, Google, and Mistral shipped updates that align with it. Standards wars rarely end this cleanly; this one apparently did.
The First "Agent Reliability Engineer" Job Listings Appear — at $280K–$410K Base
The postings, from a mix of insurance carriers, logistics platforms, and two large banks, describe a role that is roughly 40% SRE, 30% prompt-and-tool-set author, and 30% compliance liaison. The title is new; the work is what senior platform engineers have been quietly doing for six months. Compensation has finally caught up to the on-call burden.
Know someone who would benefit from a weekly briefing like this? Forward this issue or share your referral link from the account page — every new member keeps The AI Edge independent. See you next Tuesday.
Growth Partner for founders stuck between $1M and $20M. I find the money you’re sitting on and build the machine that collects it. $1B+ in client results.