Write the Standard Before the Tool

OpenAI, Anthropic, and Google are talking about shared AI safety standards.

Fine. Useful, even.

But do not make that a reason to wait.

A lab standard will not tell your sales agent whether it can email a prospect. It will not decide whether an HR summary belongs in a performance conversation. It will not know which finance workflow can touch a board report, or who gets called when the agent quietly does the wrong thing.

That is company work.

The missing layer is not a grand governance program. It is an operating standard: who owns the workflow, what the AI can touch, what it can change, what needs approval, what gets logged, and what counts as an incident.

Safety without that layer is mostly theater.

The FINRA comparison in the standards talk matters because finance learned this the hard way. Trust is not a vibe. It is records, supervision, permissions, audits, and consequences.

AI inside a business needs the same bones.

Pick one live workflow.

Write the standard before you scale the tool.