The model race stayed busy this week with Anthropic's Fable and OpenAI's GPT-5.6. As the menu of capable models grows, so does the operating question: which one should handle which request?
At enterprise scale, that choice cannot live inside every application or depend on whichever model a team happens to prefer. Thousands of employee requests, dozens of agents, several providers, and tools connected to company systems turn the default path into a budget, quality, security, and continuity decision.
What an AI Control Plane Does - The main essay. An AI control plane is the shared set of policies and tools that governs an AI request: which models are allowed, what information and tools they receive, what actions they may take, how the company evaluates the result, and what happens when something fails.
It works as a shared front door for AI work. First, it classifies the request and decides whether it belongs with an agent, model, ordinary software service, or person. If AI is needed, it chooses an approved agent, model, and toolset based on measured quality, speed, cost, risk, and availability. It then limits information and authority to what the task requires, records the result, and applies the right recovery path when something goes wrong.
That shared layer does not replace the identity, approval, transaction, and audit controls in systems such as CRM, ERP, or payments. It can enforce policy only where it is connected, and its concentration creates risks of its own: outages and bad routing rules can affect several workflows at once. The control plane therefore needs a permanent operating team to maintain permissions, evaluations, fallbacks, success measures, and connectors as models and workflows change. That recurring work is AI Operations.
Also in this issue:
- The Wire - Frontier competition is becoming a price war, China may restrict overseas access to advanced models, Box is hiring for the work AI creates, and vendors are becoming traffic controllers for AI.
- Quick Hits - The UN is developing identity frameworks for agents, CIO checks complicate Agentforce's revenue story, multi-model systems can share the same blind spots, and token analytics are moving into developer workflows.
- Meanwhile... - Machine learning and particle-accelerator scans are helping scholars virtually unwrap and read a Herculaneum scroll carbonized by Mount Vesuvius nearly two thousand years ago.
- What I'm Consuming - Agent-led software development, recursive self-improvement, agents for everyday users, Cisco's multi-agent engineering architecture, and BCG's agentic leadership playbook.
- After Hours - Arturo Perez-Reverte's The Final Problem, a literary locked-room mystery about murder mysteries, performance, identity, and the clues hidden in silence.
Start with one workflow that has meaningful volume, cost, risk, or customer impact. Define where AI belongs, what it can access and do, how success will be judged, and what happens when it fails. Then build those controls as reusable company policies instead of instructions buried inside one agent.