Connect
PetConnect
Cortex
When evaluating a veterinary AI scribe, do not stop the timer when the first draft appears. The practical question is how much work a veterinarian needs to check that draft against the right source, make corrections and complete the record. Fast text generation is only one part of that process.
This guide gives practice managers and veterinarians a method for evaluating an AI scribe or voice documentation tool. It addresses source consistency and human review, rather than diagnosis or treatment decisions.
Separate draft generation from record approval#
Treat the output as a draft awaiting review. Make the reviewing clinician, corrections and final status visible. Test whether the product's save action creates a draft, a shareable document or a completed record; those states should be understood before routine use.
The RCVS guidance on AI in practice emphasises verification and correction of AI-generated records by a veterinary professional. This is UK professional guidance, not a certification that a product meets requirements in every country.
If your team does not yet share a record structure, start with the SOAP documentation basics. Do not assume a tool will automatically resolve inconsistent documentation habits.
Select useful pilot examples#
Begin with synthetic scenarios or data your organisation has confirmed is appropriate for the test. Before using live audio, client information or clinical documents, assess recording, processing, retention and access arrangements. Avoid adding unnecessary personal information to evaluation files.
Instead of one uninterrupted dictation, prepare examples involving an interrupted sentence, references to two pets, a negative statement, uncertainty in an owner's account and a later correction. These are proposed test cases, not claims that a particular tool has failed them.
Give each case a source text and a clear boundary for the expected information. The output need not match a reference word for word, but it should preserve the information, its source and any uncertainty.
Six checks for an AI-generated veterinary note#
Check | Review question | Result to record |
|---|---|---|
Patient and visit | Does the draft belong to this patient and encounter? | Confirmed match or corrected field |
Information source | Are owner reports distinct from clinician findings? | Example of any source confusion |
Negative statements | Are meanings such as not seen or absent preserved? | Meaning that required correction |
Uncertainty | Has a tentative statement become a definite one? | Certainty level corrected |
Missing or added information | Has anything unsupported been introduced? | Omissions and additions separately |
Final status | Are review, correction and approval complete? | Reviewer and completion time |
Have an authorised clinician separately check fields that may affect clinical meaning, including numbers, dates, units and medication references. This table is a proposed review aid, not a complete clinical safety standard. Incorporate it into your practice's own documentation review procedure.
Make a small change in meaning visible#
Consider a wholly fictional example. The source says, “The owner reports not seeing any vomiting.” A draft that simply says, “No vomiting” has lost the source of the observation. Record the correction as “owner attribution restored” rather than just “edited.” This makes recurring categories of correction easier to compare.
Measure the whole documentation task#
Use the same measurement boundary for the starting workflow and the AI-assisted workflow: input, waiting for the draft, reading, correction and completion. Measuring typing alone leaves review work out of the comparison.
As a practical pilot design, select a comparable group of record types for each method. Record total time per note, correction categories and the number of examples assessed. Use the results to understand where work occurs, rather than ranking clinicians against each other by speed.
If the team considers a type of error clinically unacceptable, a favourable average time does not make it acceptable. Define pause and reassessment conditions before the pilot begins. Include the cost of evaluating and operating the tool in your software total cost worksheet.
Ask the supplier to demonstrate the workflow#
- Where are audio and text processed, how long are they kept and how can they be deleted?
- Is information used for model training, and where is that documented contractually?
- How are roles and record access restricted?
- How does a draft differ from a completed record?
- How can a corrected note and an earlier version be inspected?
- How does a veterinarian continue normal documentation if the service is unavailable?
Keep the answer, demonstration evidence and relevant document together. A public feature description should not be read as a contractual guarantee about data processing or compliance.
The Vetigen SOAP feature page describes creating structured draft notes from speech and a workflow in which the veterinarian reviews and approves the result. Evaluate that workflow using your own source examples and review table.
Frequently asked questions#
Does a fluent note still need review?#
Yes. Fluency does not establish consistency with the source. Check patient matching, attribution and uncertainty independently of how well the text reads.
Is one accuracy percentage enough for a pilot?#
A single rate can conceal different types of correction. Define what counted as an error, the number of examples reviewed and the categories found. Do not assume corrections with different clinical implications have equal significance.
Does intake belong in a documentation review?#
The origin of information and the encounter it belongs to should be clear from the start. The patient intake template can help organise those matching and handover points.
Prepare your pilot record#
Download the AI note review worksheet. Record a synthetic example reference, check outcome, correction type, review time and responsible reviewer together. Keep identifiable clinical content in the authorised clinical system, rather than duplicating it in an evaluation sheet.
A blank tracking table is also included. Use record references; do not duplicate patient or client information in this working file.
Review Vetigen’s AI note workflow with your own checklist
Assess how a draft is created and how clinician review leads to a completed record.
