Contents
The Shape of the Actual Workload
Refill requests are unusual among veterinary front-desk tasks because they arrive from everywhere at once and resolve nowhere in particular.
They come in by phone from clients. They come in by email. They come in as faxes and portal requests from online pharmacies acting on the client's behalf. They come in as a voicemail at 6.40pm. Each one is individually trivial and collectively enormous, and critically, none of them is finished when it is received - it has to sit somewhere until a veterinarian looks at the patient record and makes a decision, and then something has to happen on the other side.
That middle state is where the cost lives. A request that arrives at 9am and gets authorised at 4pm consumed almost no clinical time, but it consumed a sticky note, two client calls asking whether it had been done yet, one call to the pharmacy, and a slot in somebody's working memory for seven hours. Multiply by the volume a busy practice sees and you have the real problem: not the decisions, the tracking of the decisions.
This matters because it tells you what to build. The pages currently ranking for refill automation mostly propose an AI voice agent or a chatbot that "handles refill requests." That is the wrong shape. The bottleneck is not intake capacity, it is the absence of a single queue with a state on each item. You can fix that with an unglamorous router and see most of the benefit, and you can fix intake capacity later if it turns out to matter. Our veterinary automation overview covers where this sits among the other candidates in a practice.
Sourcing note. The regulatory quotations below are taken from primary regulation text and from the AVMA's own published summary, with URLs given inline. Where we could not retrieve a source we say so - eCFR's own site refused our automated requests and redirected to a bot-block page, so the CFR text below is quoted from Cornell's Legal Information Institute mirror of the same sections. Nothing here is legal advice and every practice should confirm its position with its own counsel and state board.
The Line You Cannot Cross, With the Regulation Quoted
This is the section the thin pages skip, and it is the reason this article exists. Before designing anything, understand that a refill authorisation is a clinical act performed inside a regulated relationship.
The federal VCPR
The AVMA publishes a summary of the federal veterinarian-client-patient relationship requirements, citing the underlying FDCA and 21 CFR Part 530. Read in August 2026, that document states:
"Establishing the federal VCPR requires a physical examination of the animal or timely and medically appropriate visits to the premises where animals are kept. The FDA does not allow the VCPR to be established through electronic means. FDA does allow the VCPR to be maintained electronically through telemedicine."
And, importantly for any multi-state group:
"Veterinarians must comply with the federal VCPR in each of the circumstances in which it applies, irrespective of whether state law defines it differently."
The same document is explicit that the federal VCPR is not an edge case. It "applies to any use of an FDA-approved human drug in animals, including over-the-counter (OTC) human drugs," to "any use of an FDA-approved animal drug in any manner that differs from its approved labeling (Extra Label Drug Use)" - different dose, different frequency, different route, different species - and to "the use of compounded drugs by veterinarians" and to a veterinary feed directive. As the AVMA puts it: "All of these are very common occurrences in the day-to-day practice of veterinary medicine."
The document also quotes an FDA letter to the AVMA of 6 April 2017 stating that "for the purposes of the federal definition, a VCPR cannot be established solely through telemedicine (e.g., photos, videos, or other electronic means that do not involve examination of the animal(s) or timely visits to the premises)."
What this means for your build. The relevant question on many refill requests is not "is this drug appropriate" but "when was this animal last examined, and is the relationship still current?" That is a data question your automation can answer and surface. It is not a decision your automation can make. The design implication is precise: compute and display the last-examination date on every request; never gate on it automatically without a veterinarian's sign-off rule that a veterinarian wrote.
Controlled substances
Here the limits are numeric and quotable, which makes them the ideal candidates for deterministic enforcement rather than model judgement.
- 21 CFR 1306.12(a): "The refilling of a prescription for a controlled substance listed in Schedule II is prohibited."
- 21 CFR 1306.22(a): "No prescription for a controlled substance listed in Schedule III or IV shall be filled or refilled more than six months after the date on which such prescription was issued," and "No prescription for a controlled substance listed in Schedule III or IV authorized to be refilled may be refilled more than five times."
Both quoted from the Cornell LII text of the CFR. Note the structure of that second rule - it is two independent ceilings, five refills and six months, and whichever is hit first ends it. A refill counter alone is an incomplete check.
These belong in your workflow as hard blocks written in code, evaluated before anything else runs, and they should be impossible to override from within the automation. Not as a line in a system prompt. A guardrail that a language model can be talked out of is not a guardrail, and this is exactly the category of rule where hallucination is unacceptable rather than merely annoying.
State law adds further requirements on top of all of this and varies considerably. Your build should have a per-jurisdiction configuration block even if you only operate in one state today, because the alternative is a rewrite the day the practice opens a second location.
What Is Actually Automatable
Strip out the decision and a great deal remains. Here is the honest inventory.
Intake normalisation. Phone, email, portal, fax and pharmacy request all become one record with the same fields: client, patient, drug, strength, quantity, requesting channel, requesting party, timestamp. This is the single highest-value step and it involves no intelligence at all for most channels. Where it does need intelligence - reading a faxed or emailed request that is unstructured prose - that is a well-bounded extraction task, and the correct output is a draft record flagged for confirmation, not a committed one.
Record matching. Attach the patient, the prescription history, the last examination date, the remaining authorised refills and the drug's schedule. Every one of these is a lookup. If your PIMS exposes them, this is the step that turns a five-minute decision into a fifteen-second one, because the vet stops hunting.
Deterministic blocking. Schedule II requests never enter the normal queue. Requests where the prescription is past six months or past five refills for a Schedule III or IV drug are flagged with the reason. Requests where the last exam is older than the practice's own threshold are flagged with the date shown. Flagged, routed, and explained - not silently declined, because a client whose request vanishes will phone, and you have re-created the problem.
Chasing and status. The request has a state and an age. Anything unresolved past a threshold escalates. The client gets an automatic acknowledgement on receipt and an automatic notification on outcome. This alone removes the "has it been done yet?" call, which in most practices is a larger volume than the original requests.
Dispatch. Once a veterinarian authorises, sending the authorisation onward to the pharmacy, the in-house dispensary or the client is mechanical.
Not automatable, at all: the authorisation, any judgement about whether the VCPR is current, any substitution or dose change, and any conversation with a client about why something was declined. The last one is worth stating explicitly because it is tempting - a declined refill is usually a conversation about a check-up that is overdue, and that conversation converts far better from a human.
If you want the equivalent decomposition for a different regulated setting, our AI agents in healthcare piece walks the same reasoning, and what AI agents can do reliably covers the general boundary.
Getting the Data Out: The PIMS Question
Everything in the matching step depends on what your practice management system will tell you, and that varies sharply by vendor.
ezyVet publishes a developer portal at developers.ezyvet.com. Read in August 2026, it documents an OAuth 2.0 bearer-token model - "ezyVet uses a oauth2.0 style authentication model and utilizes bearer tokens" - with tokens issued from POST /v1/oauth/access_token and a stated TTL of 12 hours, which the docs note requires "An automated solution for refreshing tokens." The endpoint families listed include Prescription Management, Clinical Management, Contact Management and Animal Management, and the guides index includes one specifically titled External Pharmacy - which is precisely the surface a refill workflow needs. We could not retrieve the contents of that individual guide (the per-guide URLs returned 404 to our fetches), so confirm the exact endpoints and fields in the live docs rather than taking a shape from this article.
We also could not read ezyVet's documented "Request Throttling" limits from any public page, and the ezyVet knowledge centre served a browser-compatibility notice rather than content when fetched. Assume throttling exists, back off politely, and ask apisupport-ezyvet@idexx.com for the real numbers before you build anything that polls hard. See rate limit and polling vs webhook for how that choice shapes the design.
Covetrus systems - Pulse, AVImark, Impromed, Ascend and others - take the opposite approach. We found no public developer documentation. Covetrus's own partner page describes Covetrus Connect as offering "Secure and permissions-based integration" and states it is "The only approved and supported way to integrate with Covetrus practice management software." That makes integration a commercial conversation rather than a technical one, and you should establish whether it is available to you before scoping a build.
If you have no API at all, the router still works. The queue lives outside the PIMS, the vet opens the patient record manually to decide, and the automation handles intake, state, chasing and notification. You lose the matching step, which is the biggest single time saver, but you keep the tracking - and the tracking was the original problem. Do not let the absence of an integration stop the project; let it change the scope.
A Worked Time Model, With the Arithmetic Shown
An illustrative model, not a result we have measured at a client. Substitute your own counts - the structure is the point.
Current state. A three-vet practice logs, say, 35 refill requests a day across all channels. Assume each currently costs:
- 2 minutes of front-desk time on intake and writing it down
- 3 minutes of veterinarian time, of which most is finding the record rather than deciding
- 2 minutes of front-desk time on the follow-up call from the client asking whether it is done, which happens on perhaps 40 percent of requests
- 1.5 minutes on dispatch to pharmacy or dispensary
Per request: 2 + 3 + (0.4 x 2) + 1.5 = 7.3 minutes. Across 35 requests: 35 x 7.3 = 255.5 minutes, or 4.26 hours a day spread across the team. Over a five-day week that is 21.3 hours.
After a router. The realistic post-build figures, being deliberately conservative:
- Intake drops to 0.5 minutes averaged - portal and pharmacy requests self-populate, phone requests still need a human but the fields are prompted
- Veterinarian time drops to 1.2 minutes, because the record, the history and the exam date are already attached and the decision is batched rather than interrupting
- The status-chasing call largely disappears, to perhaps 5 percent of requests: 0.05 x 2 = 0.1 minutes
- Dispatch drops to 0.3 minutes
Per request: 0.5 + 1.2 + 0.1 + 0.3 = 2.1 minutes. Across 35: 73.5 minutes, or 1.23 hours a day. Weekly: 6.1 hours.
The delta is 21.3 - 6.1 = 15.2 hours a week. Now the honest part. That 15.2 hours is not a redundancy - it is 15.2 hours returned to a team that was already behind on other things, and it is split across four or five people in fragments. If you price it at a blended loaded rate of, say, $28/hour, it is 15.2 x $28 = $425.60 a week, or roughly $22,100 a year. If you would not actually reduce hours, then the number is a capacity figure and not a saving, and you should say so in the business case rather than presenting it as cash. Our ROI of AI agents piece goes into why that distinction is where most automation business cases quietly cheat.
Against build cost. If the router costs $3,000 to build and $60/month to run - replace with your own quotes; automation build cost explains what drives the range - then first-year cost is $3,000 + $720 = $3,720. Against $22,100 of recovered capacity, payback on the capacity figure is $3,720 / $425.60 = 8.7 weeks. Against a cash saving of zero, payback is never, and the decision becomes "is 15 hours a week of front-desk and vet capacity worth $3,720?" - which most practices will answer yes to, but for a reason they can defend. See payback period for how to frame it.
The single number most sensitive to error here is the veterinarian's 3 minutes. Time it before you build. If it is really 90 seconds because your team is already efficient, the whole case shrinks by more than a third and you may be better off spending the budget on phone coverage instead.
Designing the Queue
Concrete design decisions, in the order they matter.
One queue, not one per channel. The whole point is a single place with a single state model. If pharmacy faxes live somewhere different from portal requests, you have automated two problems into two problems.
States, explicitly. Received, blocked (with reason), awaiting vet, authorised, declined, dispatched, closed. Every item is in exactly one. Age is tracked from Received, not from the last transition, or long-stuck items hide behind recent activity.
Batch the vet's attention. Two review windows a day beats a notification per request. The single largest cost in the current state is context-switching, and an automation that pings a veterinarian 35 times a day has made that worse while looking busy. Configure the batch, and let the escalation path handle the genuinely urgent one.
Make the decision one tap with the reason attached. Authorise, decline, or request an appointment. That third option is the one that matters commercially, because a refill request from a patient who has not been examined recently is an appointment the practice should be booking - and it should generate a booking offer to the client automatically once the vet selects it. That is the same fill mechanic described in our veterinary no-show and recall automation piece, and it is where a refill router stops being a cost saving and starts being revenue.
Idempotency on dispatch. An authorisation sent twice to an online pharmacy can become a duplicate dispense. Key every outbound on request ID and check before sending. See idempotency.
Log everything, immutably. Who requested, what was attached, who decided, when, and what was sent. In a regulated setting the audit trail is not a nice-to-have and it is far easier to build in on day one than to retrofit. Related: observability.
Measure decision latency, not automation rate. The metric that matters to a client is how long from request to answer. The metric that matters to the practice is how many minutes it cost. "Percentage automated" is a vanity metric here and optimising it pushes you straight at the decision boundary you must not cross.
Where a Voice or AI Layer Genuinely Helps - And Where It Does Not
Given how much of this article argues for an unglamorous router, it is worth being specific about where the AI layer earns its place.
It helps at unstructured intake. A faxed refill request from a pharmacy, an email written in prose, a voicemail - turning these into structured fields is genuinely a language task and genuinely hard to do with rules. Output a draft for confirmation, never a committed record.
It helps at phone deflection for the status question. "Is my refill ready?" is a lookup with a fixed answer set, it is high volume, and getting it wrong is low-consequence. This is close to an ideal voice AI agent task. What that costs for a practice is broken down on our AI receptionist cost for veterinary clinics page.
It helps at drafting the decline explanation for a human to review and send - not at sending it.
It does not help at the decision, and any vendor pitching an agent that "approves routine refills automatically" is selling you the one part of this workflow that a veterinarian's licence is attached to. Ask them directly which regulation they believe permits that and who carries the liability. Our guide to evaluating AI agent vendors covers the general pattern of claims that dissolve under a direct question.
It does not help at triage of clinical concern. A refill request that arrives with "she's been off her food since Tuesday" attached is not a refill request. Route any inbound containing clinical content to a human on a keyword rule, deterministically, and accept the false positives.
If you are weighing whether to build this at all, the free automation audit is a structured way to compare it against the other candidates in the practice, and automation services explains how we scope work like this. Budget context lives on AI automation cost.
FAQ
- Can an AI agent approve a routine refill on its own?
- No, and you should treat any vendor claiming otherwise with suspicion. The authorisation is a clinical act performed inside a veterinarian-client-patient relationship. The AVMA's summary of the federal requirements states that establishing the federal VCPR "requires a physical examination of the animal or timely and medically appropriate visits to the premises" and that "The FDA does not allow the VCPR to be established through electronic means." An automation can assemble everything the veterinarian needs to decide in seconds. It cannot be the one deciding.
- Does state law override the federal VCPR if my state is more permissive?
- The other way round, where the federal standard applies. The AVMA document states that veterinarians "must comply with the federal VCPR in each of the circumstances in which it applies, irrespective of whether state law defines it differently," and notes that federal applicability covers extralabel drug use, any use of human drugs in animals including OTC, compounded drugs and veterinary feed directives - which it describes as "very common occurrences in the day-to-day practice of veterinary medicine." Confirm your own position with counsel; this is not legal advice.
- How should the workflow handle controlled substances?
- As deterministic hard blocks evaluated before anything else. 21 CFR 1306.12(a) states "The refilling of a prescription for a controlled substance listed in Schedule II is prohibited." 21 CFR 1306.22(a) states that Schedule III and IV prescriptions may not be filled or refilled "more than six months after the date on which such prescription was issued" and may not be "refilled more than five times." Note those are two independent ceilings and whichever is reached first ends it, so a refill counter by itself is not a sufficient check. Write these in code, not in a prompt.
- Do I need a PIMS integration to make this worthwhile?
- No, though it is where the biggest single time saving comes from. Without an integration you still get one queue, one state model, automatic acknowledgements, automatic status notifications and escalation on age - which addresses the tracking problem that generates most of the wasted time. What you lose is auto-attaching the patient record, prescription history and last examination date, so the veterinarian still hunts. Scope accordingly rather than cancelling.
- Which PIMS is easiest to build against?
- Of the systems we could verify, ezyVet is the most open: it publishes a developer portal documenting OAuth 2.0 bearer tokens, a 12-hour token TTL, and endpoint families including Prescription Management and Clinical Management, plus a guide specifically on External Pharmacy. Covetrus systems have no public developer documentation we could find; Covetrus's own page describes Covetrus Connect as "The only approved and supported way to integrate with Covetrus practice management software," which makes access a partner application rather than a signup.
- Should the automation decline a request when the last exam is too old?
- It should flag it, show the exam date, and route it - not decline it. Two reasons. First, whether the relationship is current is a judgement a veterinarian makes, not a date comparison. Second, an overdue examination is a booking opportunity, and the best outcome of that request is an appointment rather than a refusal. Build the decline path to offer a booking, and have a human make the call.
- What metric should I hold the build to?
- Decision latency - median and 90th-percentile time from request received to client notified of the outcome - plus total staff minutes per request, timed before and after. Avoid "percentage of requests automated." It is a vanity metric in this workflow and optimising for it pushes the design directly at the regulatory boundary, which is the one place you cannot afford to be creative.
- Will this reduce headcount?
- Almost certainly not, and a business case that claims it will is usually wrong. The recovered time comes in fragments spread across several people, which converts into capacity rather than cash. That is still worth paying for in a practice that is behind on callbacks, recall and reception cover - but present it as capacity and let the practice decide what to spend the capacity on, rather than dressing it up as a salary saving that never materialises.
If this resonated
Module 02 is the long version of this post.
Eval set, refusal pattern, kill switch. Build it in a weekend.
