Veterinary care
When the regulator, not the architect, drew the system boundary
A practice group wanted an AI assistant to advise owners out of hours. Professional rules on what constitutes an animal being “under care” meant the assistant could never do the valuable-sounding part, so the design changed to the part it could legitimately do.
A regulatory rule turned a broad product idea into a narrow, useful one.
Problem
Out-of-hours calls are answered by a third-party service at significant cost. Many calls need no clinical intervention. The group wanted a conversational assistant to triage, advise, and where appropriate arrange a prescription.
Context
Eleven sites, one shared practice management system, an out-of-hours contract renewing in the same year.
Current architecture
Practice management system on a vendor cloud, telephony diverted after hours to the third-party provider, no integration between the two.
Constraints
- Regulatory: RCVS guidance on “under care” requires a physical examination before prescribing antimicrobials or controlled drugs, other than in exceptional circumstances.
- Clinical safety: the failure mode of a wrong “no need to come in” is an animal that deteriorates overnight.
- Data: owners describe symptoms in lay terms; records are clinical.
Evidence
Each statement placed on the ladder before it was used.
RCVS Council approved new ‘under care’ guidance in January 2023; the parts covering prescribing POM-Vs, limited service providers and advice-only services took effect on 1 September 2023, requiring a physical examination for antimicrobials and controlled drugs save in exceptional circumstances.
Whether to prescribe remotely remains the veterinary surgeon's professional decision, so the system can never be the decision-maker, only the preparer.
A substantial share of out-of-hours calls are advice-only. The share was believed by the clinical lead and never counted.
Structured triage before the call could shorten the call, or remove the need for it.
Questions that changed the answer
- What proportion of out-of-hours calls end with no medicine and no visit? Nobody had counted.
- Which of those calls could have been answered by a structured questionnaire?
- Who carries professional liability for advice the assistant gives?
- What does the out-of-hours contract charge for: calls taken, or cases seen?
Options
Including the one nobody wanted to discuss.
Conversational advice assistant
A model answers owner questions directly and recommends a course of action.
What it costs: Offers advice that professional guidance reserves for a veterinary surgeon, on the highest-consequence failure mode in the business.
Structured triage that prepares the call
Owner completes an adaptive questionnaire; the system produces a structured summary and a red-flag list for the on-call vet, who makes every decision.
What it costs: Less impressive to demonstrate, and only pays if it genuinely shortens calls.
Renegotiate the out-of-hours contract
Change the commercial terms rather than the technology.
What it costs: One-off saving, no capability, and it was pursued in parallel.
Economics
Four horizons, not one estimate.
- Build
- Triage flow plus practice-management integration. Modest, because no model training is involved.
- Run
- Per-conversation inference cost is small next to the per-call fee it is meant to reduce; the real run cost is clinical review of the question set.
- Change
- The question set changes as species mix and seasonality change, so an owner is needed inside the practice, not at the vendor.
- Exit
- Low. The questionnaire and its logic are the asset, and they are plain text.
Decision
Build structured pre-call triage that produces a summary and red-flag list for the on-call vet. The system never advises the owner and never prescribes.
Why
The valuable, defensible part of the workflow is preparing a clinician to act quickly, not replacing the clinician. That conclusion came from professional guidance rather than a technology preference.
Why not the alternatives
- Conversational advice assistant: It would occupy the space professional guidance reserves for a veterinary surgeon's judgement, with an irreversible failure mode.
- Renegotiate the out-of-hours contract: Not rejected. It was pursued in parallel by the commercial team. It lowers the bill and creates no capability, so it was never an alternative to this work.
Trade-offs accepted
- A much smaller product than the one originally asked for.
- Clinical time is required permanently to keep the question set safe. That run cost never ends.
Reversibility
A questionnaire and a summariser can be withdrawn overnight without affecting the ability to treat animals. That is precisely why it was a safe place to start.
Non-functional requirements
- Every red flag must route to a human within seconds
- The full owner transcript must be attached to the clinical record
- The system must degrade to ‘call us’ when uncertain, not guess
Decision gate
PAUSE, not GO. The central economic assumption (the share of advice-only calls) has never been counted. Four weeks of call coding is what moves it to GO, and it runs before any build.
Implementation
An adaptive questionnaire, a summariser that writes into the practice management system, and a hard rule that any red flag dials a human immediately.
What the decision was expected to achieve
Calls should get shorter, and should start with better information. Call duration and the proportion of calls arriving with a complete structured summary would show it; a satisfaction score would not.
No outcome is claimed. This is an illustrative example, so there is nothing measured to report, and a real engagement would state what happened and how it was verified.
Lessons
- Read the professional rules before the architecture. They are constraints in the engineering sense, because they delete options.
- The feature that demos best is often the one you are least permitted to ship.
- ‘Assist the professional’ and ‘replace the professional’ are different systems with different liability, not different settings.