The gist: Chief-of-IT topics for week 23: The EU Commission is consulting on high-risk AI guidelines — IT leaders must now assess their own AI landscape against Annex III for the first time. Regulatory sandboxes are established in all 27 member states. Anthropic documents its sandbox architecture. And a single customer consumes 500 million dollars in Claude tokens without usage limits — the scaling question becomes concrete.
What IT leadership must address this week.
1. High-risk classification: now operationalizable
The EU Commission’s draft guidelines on the classification of high-risk AI systems have been publicly accessible since 1 June. This means Annex III is now underpinned by examples and decision logic for the first time. For IT leaders, this means: The arguments “we don’t know if that’s high-risk” and “the guidelines are still missing” are no longer tenable from this week on.
Operational recommendation: Create a lean AI system inventory this week. Columns: System, use case, affected persons, data flow, presumed Annex III category. Half a page per system is sufficient to reach a reliable compliance status in 90 days.
2. Regulatory sandboxes: use them, don’t ignore them
All 27 EU member states have established their AI regulatory sandboxes. For IT leaders who want to deploy AI in regulated domains (health, education, human resources, finance), this is a practical vehicle: sandboxes offer a legally safe testing environment, often with support from the national authority. Anyone with an AI pilot project in one of these domains should examine the sandbox option before launching regularly.
In Austria, the RTR/Federal Government sandbox is the first point of contact. In Germany, the Federal Network Agency coordinates in cooperation with the BMWK.
3. Anthropic sandbox architecture as a discussion basis
Anthropic has documented the sandbox architecture across all Claude products — from Claude.ai through Claude Code, Cowork, and API connectors to Computer Use. For IT leaders, this is the long-overdue basis for creating a shared conceptual framework with data protection, compliance, and CISO teams: Which data leave which sandbox when? What is trained, what is not? How long retention?
Practical application: Use the document as a template for your own AI data processing atlas — and map it against your GDPR record-keeping structures.
4. Effective structures for long-running AI agents
Multiple sources this week address the topic of long-running, autonomous AI agents — architecture, evaluation, workflow design. What was academic six months ago is now an operational IT question: when an agent runs 24/7 in a ticket system, an order workflow, or a monitoring stack, you need a new mental model for availability, observability, and failure scenarios — AI agents do not behave like classical microservices.
Recommendation: Begin this week to sketch a first internal AI agent operations guideline. Three pages suffice. Topics: SLA definition, escalation paths, audit logging, emergency stop.
5. Scaling reality: 500 million tokens without limit
A single Anthropic customer consumes 500 million dollars in token volume without usage limits. This is not merely a commercial curiosity — it signals that the token economy in the enterprise sector is making the same leap that cloud economics made ten years ago. For IT leaders, this means: AI consumption becomes its own cost center position. Whoever does not plan this as a budget line item in Q3 will communicate it as an oversight in Q4.
What should be decided this week
- Start AI system inventory (target: first version by end of June)
- Sandbox check for regulated pilot projects
- AI data processing atlas outline based on Anthropic sandbox documentation
- AI agent operations guideline prepare as discussion paper
- AI cost center include as budget position in Q3 planning
Week 23 is the week in which AI compliance, AI operations, and AI budget move from the strategy paper into IT organization charts.
Lumi AI News Chief-of-IT Brief — curated from 14 IT leadership-relevant sources, classified through Lumi News Pipeline v1.2.8. Marked in accordance with Art. 50 EU AI Act: AI-assisted editorial.