Agents belong in the back office, not just the chat box

Most "AI for vets" right now is a chat box bolted onto a website. It can answer questions. What it can't do is the thing that would actually help a short-staffed clinic: the work. Surface the patients overdue for vaccines. Draft the recall list. Look up a weight-based dose. Find the open slot and book the visit. That work lives inside the PIMS, and an agent can only do it if the PIMS has a real, write-capable API.

Why the closed PIMS blocks this

An agent is only as useful as the tools it can call. If the system of record won't let software write back (create the appointment, record the note, update the record), then the agent is stuck reading. That's the wall most veterinary AI hits today. It's not a model problem. It's an access problem.

What we're building

OpenVPM includes an optional agent that requires a configured model key and explicit clinic opt-in. Its typed tools can find records, retrieve supported clinical context, list overdue vaccines, reference supported dosing information, find open slots, and book appointments. Staff remain responsible for reviewing outputs and approving enabled writes.

• Typed tools narrow the agent's scope, but AI output can still be incomplete or wrong.
• Writes require explicit clinic opt-in and should remain under staff review.
• Bring your own model key. The implementation is public and the enabled workflow is configurable.

The useful near-term role for this agent is supervised assistance on a small set of explicit tools. Voice agents, transcript scribing, diagnostic AI, and autonomous outreach are not supported OpenVPM capabilities today.