The landscape, and which tool for what
Free or paid, small or large, general or built for veterinary work: this module is practice at choosing the right tool for the job, and you leave with a card that lists the tools you have. One note first: the course author leads a vet-specific AI product at VIN and is AVI's president. So this module names no products, and the same questions apply to every tool, that one included.
Small or large: what you just saw¶
Larger models tend to do better on hard reasoning. They take more computing power for each answer, so they cost more and run slower. On an easy job, such as rewording an email, a small, fast model does as well. On a hard clinical question, the small model can still sound sure of itself while it misses the point.
So match the model to the task. Many free plans offer a smaller model by default and a larger one within limits. Paid plans raise those limits.
Try it yourself (if you have time)¶
Take the clinical question you brought, the one where you already know the answer. Ask it on two models you can reach for free: two providers, or a small and a large model in the same app. Then note where the answers differ, and which one you would act on. The card at the end of this page has a line for it.
Which tool for which job¶
For each task, decide the lane before you open it:
- Small and fast is fine.
- The best model you have.
- A grounded tool, or check the source yourself.
- Not an AI job.
Reword an overdue-bill reminder so it sounds friendlier
Small and fast is fine. It is a language job with low stakes, and any model does it well.
Fix the spelling and grammar in a job advert
Small and fast is fine. Easy, and you read the result before it goes out.
Turn your bullet points into a paragraph for the clinic newsletter
Small and fast is fine. A short writing job that you will edit anyway.
Summarise a long research paper for a team meeting
The best model you have. It is long, and a larger model keeps track of more of it. Check the summary against the abstract and the results.
Build a staff rota around five people's leave and shift limits
The best model you have. Many rules at once is hard reasoning, where small models slip. Check every rule in the result.
Work through a differential list for a complicated case, with no client details
The best model you have. It is hard reasoning. Treat the list as prompts for your own thinking, not as the answer.
Is it safe to keep a dog on meloxicam now that its kidney values are up?
A grounded tool, or check the source yourself. The source matters, and so do this patient's values. The taste test showed how a fluent answer can miss them.
What is the dose of a drug that was approved last month?
A grounded tool, or check the source yourself. A model's training may stop before the approval, so the answer needs a current source you can open.
Draft a letter to a client, pasting in their name, address and full history
Not an AI job, as asked. Client details do not go into a general chat tool. Remove them, or use a tool your practice has approved for records. Module 6 covers this.
Send an owner a treatment plan the tool wrote, without reading it first
Not an AI job. The tool drafts; you check and sign. Nothing clinical goes out unread.
The four questions behind the lanes: Is it hard, or long? Does the source matter? Does it hold client or patient details? Does it replace your own judgement? Your answers go on the card as your minimums.
General chatbots and vet-specific tools¶
A general chatbot answers from what it absorbed in training. A vet-specific tool is usually built to search a curated veterinary library first, then answer from what it found, with citations you can open. Use general tools for writing, summarising and making things. Use grounded tools for clinical questions where the source matters.
Here is the meloxicam question from the taste test, asked about the same dog, with its chart in reach. Answer 1 comes from a model that sees only the question. Answer 2 comes from a tool that searched the chart first.
Answer 1. Meloxicam is commonly used long term for osteoarthritis in dogs. At 0.1 mg/kg once daily it is a standard maintenance dose, so continuing is generally reasonable with periodic monitoring.
Answer 2. No, not as things stand. The chart shows azotemia with elevated SDMA [3] and poorly concentrated urine with proteinuria [4] while the dog is on daily meloxicam [5]. NSAIDs are a concern with reduced kidney function, and the plan already stops it [6].
Which answer can you check, and why does it matter?
Answer 2. Every claim points to a line in the chart:
- [3] Chem: BUN 58, CREA 3.1, SDMA 28, phos 7.2.
- [4] UA: USG 1.012, protein 2+, no casts.
- [5] Meds: meloxicam 0.1 mg/kg q24h for OA since June.
- [6] Plan: stop NSAID, start renal diet, recheck 2 wk.
Answer 1 is fluent and plausible, but blind to this patient's kidney values. A citation only helps if you open it.
Other kinds of veterinary AI work differently again: scribes that write up notes from a consult, imaging and lab AI that flags results, and client communication tools that answer owners.
Six questions to ask of any tool¶
- What sources does it draw on?
- Can you open the citations?
- How current is it?
- How was it tested, and by whom?
- What happens to your data?
- What does it do when it does not know?
The Judging AI tools course uses these in full. For now, notice that the same questions land differently on each kind of tool.
Which question bites hardest for an AI scribe?
What happens to your data. A scribe records the consult, so the client's voice and details go to the tool. Ask where the recording is stored, for how long, and who can use it.
Try it yourself: ground an answer (if you have time)¶
Find a short passage from a guideline or protocol you can open freely, such as your practice's own protocol for a common procedure. Paste it into any chat tool and ask: "Answer only from this text. Quote the sentence you used. If the text does not cover it, say so." Then ask the same question in a new chat, without the text. Which answer can you check? Which one tells you when it does not know?
Free or paid, and what happens to your data¶
People pay for higher usage limits, access to the larger models, and features such as file uploads, longer documents, search and saved projects. Business and team plans usually come with stronger data terms than personal plans.
Does a free plan use your chats to train its models by default?
Often, yes. On many free and personal plans, chats can be used for training unless you turn it off. Business and team plans usually exclude them. The setting differs by provider and changes over time, and turning it off stops future training only. So check your own account.
Try it yourself: find it in your own account¶
Open the settings of the chat tool you use, such as Claude, ChatGPT or Gemini. Look for four things, and note each one on your card with today's date:
- Which model am I on, and can I switch?
- Are my chats used to train the model by default, and where is the switch?
- How long are my chats kept, and how do I delete them?
- Is there a temporary chat that is not saved?
For each one, note whether you found it and how many clicks it took. A setting that takes eight clicks to find tells you something about the provider. The screens change often, so look for the idea, not a particular button.
No account yet? Read two providers' public privacy pages, which need no login, and answer question 2 for each. Then choose the one you will sign up with.
Your toolkit card¶
Before you list chatbots, look for the AI you already use without having chosen it:
- your practice management software's add-ons
- flags from your lab analysers
- imaging reads from your reference lab
- client messaging and reminder features
- your employer's list of approved tools
- your memberships and library access
Put everything you find on the card below.
My toolkit card
Keep this card. Modules 4, 6 and 7 come back to it. Fill in what you found in this module, then download your copy.
What you type here stays in this browser on this device only. Nothing is sent to AVI or anyone else. Do not include client names or anything that could identify a client or patient.
From the sort above. For example: rewording to any tool, clinical questions to a grounded tool or I check the source, anything with client details to no general tool.
What differed when you asked your own clinical question on two models. If you skipped that step, note what you noticed in the taste test.
Could you check the answer against the text? Did it say when the text did not cover the question? If you skipped that step, note what you saw in the example.
Which model, whether chats train the model by default, how long chats are kept, and whether there is a temporary chat, with how many clicks each took. No account yet? Note which provider you chose and why.
Include AI you did not choose, such as features in your practice software or lab equipment. For each one, note whether it comes through you, your employer or a membership.
From your employer's policy, if there is one. If there is none, write that down and ask.
Where next
Using AI in practice: Module 3: How to prompt, and how not to (coming soon)
A shared question channel is on the way. When it opens, each answer will be written once and shared with everyone taking the course.