The few days ending 26 September describe a single movement: AI has spread fast enough that neither the companies building it nor the enterprises deploying it can fully account for it. The responses arriving this week — a self-regulatory governance framework, a new category of consumer AI hardware, a tool for agent inventory, and automatic routing by request complexity — are the practical consequence of that spread.

The three frontier labs propose a FINRA for AI

On 25 September, Google, OpenAI, and Anthropic announced plans to establish the Standards Authority for Frontier AI (SAFA), a self-regulatory body modelled on FINRA, the US financial industry's self-regulatory organisation. The proposal originated with Demis Hassabis in July and accelerated after a draft White House executive order was shelved. Former White House AI policy advisor Sriram Krishnan has been approached to serve as chief executive, with ex-OSTP director Arati Prabhakar also contacted for a leadership role. Launch is expected by year-end or early 2027.

SAFA's stated focus areas are pre-deployment safety testing, incident-reporting obligations, and qualification standards for independent auditors. The body is designed to create enforceable standards without waiting for formal legislation.

For an operator, this is not an abstraction. If SAFA gains traction — and with three of the four largest frontier labs behind it, it likely will — it sets de facto standards for what auditable AI deployment looks like. Enterprises that build internal governance practices aligned with SAFA's coming frameworks now will find vendor procurement, partner due diligence, and any future regulatory review considerably easier. The window to shape those internal practices proactively, rather than retrofit them to an external standard, is closing.

Meta puts a personal AI agent on your face

At Connect 2026 (23-24 September), Meta announced a set of hardware and model moves that together mark its most credible push into physical AI yet. The headlining product is Meta VR Glasses (Project Phoenix), an approximately 100-gram headset with a compute puck on an optical tether, priced at $1,299.99 for a spring 2027 launch. Ray-Ban Meta Audio, the first model in the lineup without a camera, arrives this year at $349. Over 100 styles of AI glasses across Ray-Ban, Oakley, and Meta Glasses will be available by year-end.

The model side: Meta introduced Muse Spark, a new series purpose-built for its consumer products, and confirmed that Muse, its personal AI agent, will come to AI glasses for hands-free voice activation. The Muse Charm, a pocket-sized keychain device, ships in December as a quick-access voice interface when glasses are not being worn.

The operator angle is not primarily consumer hardware. It is the installed base. Meta already has millions of Ray-Ban Meta glasses in the field, and Muse's arrival on that platform gives it immediate distribution that no other AI agent currently has. Enterprises with field workforces, customer-facing roles, or logistics operations should be tracking whether Meta opens a Muse API for enterprise use cases, and on what timeline.

Dataiku ships cross-platform agent inventory

Dataiku launched Agent Management on 24 September at its annual Succeed conference, with general availability planned for October. The product does one thing most enterprises cannot currently do: it finds every AI agent running in the organisation, regardless of which platform built it, measures both business KPIs and technical performance, and tiers agents by risk level.

The problem it addresses is real. According to IBM's "AI in Motion" research, fewer than one in five large organisations maintains a complete, current inventory of their AI systems. As agents proliferate across Salesforce, ServiceNow, Microsoft Copilot, and custom builds, the gap between "agents we approved" and "agents actually running" widens with each quarter.

  • Cobuild, Dataiku's natural-language agent-builder, is now embedded in Agent Management to recommend business value metrics and diagnose risk alerts.
  • Pricing is per instance annually, with monitoring metered per agent.
  • A risk-tiering mechanism flags the agents that warrant the closest oversight, rather than treating a low-stakes content-drafting agent the same as one with access to financial systems.

For any operator who has deployed more than a handful of AI agents across business units, this is infrastructure-layer tooling that belongs in the procurement pipeline now, before the October GA.

LLM Gateway routes requests by complexity automatically

LLM Gateway released Smart Route on 25 September, currently free in beta. The product reads each incoming API request before routing it, classifies the difficulty, and dispatches simple prompts to cheaper models while reserving frontier models for genuinely hard tasks. A one-line fix goes to a cost-efficient model. A multi-file architecture question goes to a frontier one.

The cost case is straightforward: production API traffic is not uniform, but most engineering teams hard-code a single model endpoint, which means either overpaying for easy requests or getting poor results on complex ones. Smart Route addresses both failure modes without requiring teams to rewrite their integration logic. Early guidance puts cost savings at up to 40% on typical production workloads.

This is a developer-facing tool, but the economics are immediately legible to any CxO signing an AI infrastructure budget. A 40% reduction in inference spend on an existing workload requires no new model capability, no retraining, and no change to the application layer.

The through-line

This week's cluster of announcements — a standards body, a physical agent platform, an inventory tool, and a routing layer — all point to the same inflection: the infrastructure of AI is being standardised and commoditised faster than the application layer. Governance frameworks, agent management, and cost routing are table-stakes concerns, not differentiators. The competitive advantage, for labs and for enterprises alike, is what you build on top once those foundations are in place. Operators who treat this week's tooling as optional will find themselves catching up to those who treated it as urgent.