How to choose an AI assistant for your veterinary clinic: eight questions to ask

Every practice-management vendor now advertises an AI feature, which makes it hard to tell a clinical-grade assistant from a slick demo. These eight questions cut through the marketing. Put them to any vendor - including us - and the answers will tell you most of what you need to know.

1. Does it cite its sources, or does it just sound confident?

Ask to see a clinical answer with the citation attached. A grounded assistant shows you where each claim came from and lets you open it. If the answer is fluent but sourceless, you are being asked to trust a guess.

2. What does it do when it does not know?

Push it on an obscure question. The right behaviour is to say it does not have a grounded source, not to invent one. An assistant that always has an answer is the one to worry about.

3. How does it handle doses and numbers?

Doses should be quoted from a source, never calculated or estimated by the model. Ask specifically how it treats drug doses - this is where confident fabrication does the most harm.

4. Where is our data processed and stored?

You should get named regions - ideally UK or EEA - not "the cloud." Data residency is a question you are accountable for, so the vendor needs a real answer.

5. Is client and patient data de-identified before it reaches a model?

The strongest tools strip identifiers before any model call, so the model never sees who the record is about. Ask whether that happens, and when.

6. Who are the sub-processors, and can we see the list?

A serious vendor can name every processor that touches your data and where each operates, and give you a data processing agreement. Vagueness here is a red flag.

7. Does it fit how we already work?

The best assistant changes your day as little as possible. Does the scribe match your note structure? Does the answer tool live where you already work? A tool that demands a new workflow tends to get abandoned.

8. Who stays in control of the record?

The clinician should remain the author. Drafts are reviewed and signed by a person; nothing is filed automatically. If the tool positions itself as the decision-maker rather than the assistant, that is the wrong shape for clinical work.

Scoring the answers

You are listening for specifics. "It cites every claim and quotes doses from the source," "data stays in the EEA," "we de-identify before any model call," "here is our sub-processor list" - these are the answers of a tool built for clinical use. Hedging, hand-waving and "trust us" are the answers of a demo.

If you would like to run these eight questions against Arlo, book a demo - the honest answers are the point.