Amazon's cloud division has released a free tool called Dogwood. It is a policy language, meaning a way of writing down rules that software must follow. What makes it notable is that it can look at an AI agent's history of actions, not just the single request in front of it. So a rule can say things like "this needs approval", "only so many times per hour", or "not once the running total passes this point".
For almost every small business in North Phoenix, this changes nothing today. It is infrastructure plumbing, released under an open licence, and Amazon itself says the reference version is not ready for production use. You will not install it, and your web developer probably will not either. It gets attention because it is a signal about where the wider industry thinks the risk sits with AI.
That signal is the useful part. The problem being solved is not "can this AI agent do this action" but "has it already done this action forty times". Any single request can look reasonable while the sequence is a disaster. That is worth holding onto, because it describes a real risk that arrives the moment you let any automated tool act on your behalf rather than just talk.
Where that shows up for a small business is in the small automations that creep in quietly. A chatbot that can issue refunds. A booking tool that can move appointments. An email assistant that can send on your behalf. Something that posts to your accounting software. Each one is fine in isolation. The question nobody asks at setup is what happens if it misfires in a loop, or if someone works out how to keep asking.
What to do: nothing about Dogwood. But if you have anything on your site or in your systems that can take an action rather than just answer a question, ask whoever maintains it two things. Is there a ceiling on how often it can act, and does anything anywhere alert you when that ceiling is hit? If the answer to either is no, that is a small job worth booking, whatever the tooling underneath ends up being called.
Reported by:
