Contents
- Two Jobs Filed Under One Word
- What Your PIMS Actually Gives You
- Building the Confirmation Sequence
- Building the Recall - The Harder, Better One
- The SMS Layer Has Rules You Cannot Design Around
- A Worked Example, With the Arithmetic Shown
- What Not to Automate in a Veterinary Practice
- A Rollout Order That Does Not Blow Up
- FAQ
Two Jobs Filed Under One Word
Ask a practice manager whether the clinic does reminders and the answer is almost always yes. Ask which of the two kinds and the conversation changes, because the word covers two automations that share nothing but a channel.
The confirmation runs off an appointment that already exists. There is a row in the practice management system with a date on it, the job is to make sure a human turns up next to it, and the whole thing expires the moment the appointment does. It is a countdown.
The recall runs off an appointment that does not exist. Somewhere in the database is a patient whose annual vaccination was due in March, whose flea and worm prescription ran out in April, or who had a dental estimate declined in January and has not been back. Nothing in the system is going to raise its hand about it. There is no deadline, no calendar entry, and no moment at which anybody is embarrassed by the omission. It is a search.
That structural difference is the whole reason recall is the neglected one. Confirmations get done because a no-show is visible - somebody stands at the front desk at 2pm looking at an empty consult room. Lapsed patients are invisible by construction. The clinic does not experience them as a loss; it experiences them as a slightly quieter Tuesday. Our no-show rate glossary entry covers the measurement side of the first problem, and it is worth reading before you try to prove any automation moved it.
Both are worth building. They are not worth building in the same sprint, and if you only ever build one, build the recall - it is the only one of the two that goes after money that has already left the building. The broader picture of what else is automatable in a practice sits on our veterinary clinic automation page.
A note on how this article is sourced. Every vendor claim below was read off the vendor's own page in August 2026, with the URL given inline. Where a page rendered as an empty application shell and we could not read it, we say that instead of filling the gap from memory. There are three such gaps in this article and they are all flagged.
What Your PIMS Actually Gives You
Everything downstream depends on this, and the answer varies enormously by vendor. It is worth establishing before anyone opens an automation tool.
ezyVet (IDEXX)
ezyVet is the friendliest of the major systems to build against, because it publishes a genuine developer portal at developers.ezyvet.com. Read in August 2026, that portal documents:
- An OAuth 2.0 model - the docs state that "ezyVet uses a oauth2.0 style authentication model and utilizes bearer tokens."
- A token endpoint at
POST /v1/oauth/access_token, which "Generates a valid access token, given client credentials." - A documented token lifetime: "ezyVet API bearer tokens have a TTL of 12 hours; noted in the response body in epoch format in the expires_in field." The docs add that "An automated solution for refreshing tokens should be put in place to ensure a valid token is always stored."
- Endpoint families grouped as Authentication, Animal Management, Appointment Management, Billing and Payments, Clinical Management, Contact Management, Inventory Management, Prescription Management, System Admin and Wellness Management.
- Guides covering API Filtering, Availability, Booking Appointments, Communication Creation and Standard of Care, with a support address of
apisupport-ezyvet@idexx.com.
Two practical consequences. First, the 12-hour TTL means your automation needs a token refresh step, not a pasted key - if you wire an API key style credential in and walk away, it dies overnight. Second, the presence of a Wellness Management family is the interesting one for recall, because that is where the "this patient is due for something" concept lives rather than in the appointment table.
Gap we could not close: the ezyVet docs list a "Request Throttling" section under API Features, but we could not get the numeric limits to render from a public fetch - the individual feature pages returned 404 to our requests and the knowledge centre served a browser-compatibility notice instead of content. Treat rate limiting as real and unquantified, back your integration off politely, and confirm the actual figures in the live docs or with API support before you schedule a large batch. Our rate limit and retry and backoff entries cover how to behave when you do not know the ceiling.
Second gap: ezyVet's own marketing page for automated reminders at ezyvet.com/features/automated-reminders returned nothing but navigation and footer when we fetched it - a client-rendered shell. We therefore make no claim about what that page promises. What we can cite is an ezyVet support article confirming the mechanism exists inside the product: to get a post-appointment callback, "You will need to create a pending automatic communication phone task to generate and run after the appointment type." That is a per-appointment-type automation configured in the PIMS itself, which for many clinics is the correct place to do confirmations and no external tooling is warranted at all.
Covetrus (Pulse, AVImark, Impromed, Ascend, RxWorks, VisionVPM)
Covetrus is the opposite posture and you should know that before you promise anyone an integration. There is no public developer portal in the search results. What Covetrus publishes is a partner programme: its own Covetrus Connect page describes the programme as offering "Secure and permissions-based integration" and states plainly that it is "The only approved and supported way to integrate with Covetrus practice management software," listing Covetrus Ascend, RxWorks and VisionVPM among the covered systems on that page. Access is an application, not a signup.
The independent API directory Supergood grades the Covetrus Pulse API an F and characterises it as "No public API. Covetrus Connect is the sole sanctioned integration layer for Pulse, AVImark, and Impromed, gated behind a partner application, signed agreement, and NDA." That is a third-party assessment rather than a vendor statement and we present it as such - but it matches what we could and could not find directly, which was a marketing page and no docs.
The practical read
For a clinic on a gated PIMS, the realistic build is not an API integration. It is a scheduled export, or the existing client-communication product the practice already pays for, or the reporting layer. That is not a failure - it is a different, slower trigger, and for recall specifically a daily or weekly cadence is entirely sufficient because the underlying condition has been true for months. The difference between polling and webhooks matters enormously for confirmations and barely at all for recall.
Building the Confirmation Sequence
The confirmation is the simpler automation and the one most likely to already exist in some form. Build it as a state machine, not a broadcast.
Trigger. An appointment exists with a start time in the future. If your PIMS supports a webhook on appointment creation, use it to enrol the appointment; otherwise poll the appointment endpoint on a schedule and diff against what you have already enrolled. The enrolment record - appointment ID, client ID, patient, phone, appointment type, start time, current state - is the thing your automation actually operates on. Everything else is a side effect.
Cadence. Two touches beat five. A confirmation request far enough out that a cancellation is still re-fillable, and a reminder close enough in that it survives short-term memory. The exact hours are a scheduling decision, not a technical one: work backwards from how long your front desk needs to fill a released slot. If it takes you two days to fill a Thursday morning, a 24-hour reminder is a courtesy, not a revenue instrument.
State transitions. This is where most builds are thin. A reply of "yes" should mark the appointment confirmed in the PIMS and stop the sequence. A reply of "no" or "cancel" should release the slot and trigger the fill path. Anything that is not a clean yes or no is a human's problem - route it to a person rather than letting a model improvise about a sick animal. This is the human-in-the-loop boundary and in a veterinary context it should sit conservatively.
Idempotency. Every send needs a key made of appointment ID plus step number, checked before dispatch. Without it, one retried workflow run at 3am texts forty clients twice, and you will spend the morning apologising rather than iterating. See idempotency for why this is not optional in any messaging automation.
The fill path is the actual product. A confirmation that only tells you about a cancellation has converted an unknown empty slot into a known empty slot. The value appears when the release triggers an offer to somebody who wants that slot: the client whose pet is on a waitlist, the one who accepted a date three weeks out, the recall population from the next section. Build the cancellation-to-offer path in the same sprint or the confirmation sequence is a reporting tool.
Our general appointment reminders guide covers the cross-industry mechanics, and automating appointment reminders goes deeper on sequencing.
Building the Recall - The Harder, Better One
Recall is a query problem before it is a messaging problem, and the query is where the work is.
Define the cohort in writing first. Not "lapsed clients" - something a database can answer without ambiguity. For example: patients with a due-date on a wellness or vaccination item that passed more than 60 days ago, with no appointment booked or attended since that date, whose client record has a mobile number and a valid opt-in, and who are not flagged deceased, transferred or do-not-contact. Every clause in that sentence is a bug you are choosing not to ship.
Suppression before selection. Write the exclusion list before the inclusion list, because the exclusions are the ones that end projects. At minimum: deceased patients, clients who have moved practices, accounts in dispute, anybody messaged by any other automation in the last N days, and anybody flagged by staff. In a veterinary practice the deceased-patient case is not a rare edge - it is a routine one, and a cheerful "time for Bella's annual vaccination!" is a message a practice does not recover from quickly. If your PIMS does not expose a reliable deceased flag through whatever interface you have, that alone is grounds to keep a human approval step in front of every batch, permanently.
Segment by reason, not by recency. A lapsed vaccination, an unfilled repeat prescription, and a declined dental estimate are three different conversations. The vaccination is a due-date nudge. The prescription is a convenience problem - the client did not decide against it, they just never got round to reordering. The declined estimate is a price and trust conversation and should usually go to a human, or at most to a message that offers a call rather than a booking link. Treating all three as one campaign is what makes recall feel like spam to the recipient and produce nothing for the clinic. The same pattern - unaccepted work that was never actually declined, just deferred - is covered from the trades angle in our unsold estimate follow-up guide, and the mechanics transfer almost directly.
Batch, throttle, and give the front desk a fighting chance. A recall run that texts 900 clients on a Monday morning generates inbound volume the practice cannot absorb, and the phone system becomes the bottleneck instead of the schedule. Release in daily tranches sized to spare capacity. This is the single most common way recall automations fail: not because the message was wrong, but because the clinic could not answer the replies. If inbound capacity is the constraint, that is a different automation - see what an AI receptionist costs a veterinary clinic before assuming you need one.
Close the loop on outcome, not on delivery. The metric is appointments booked and attended from the cohort, compared against a held-back control group from the same cohort that received nothing. Without a control you cannot distinguish the automation from seasonality, and you will either over-credit or kill something that worked. Hold back 10 percent for the first three months. It costs you a tenth of the upside and it is the only way the number means anything.
For a worked example of the same shape against a different system, our Open Dental API recall automation guide walks the equivalent build in a dental PMS, including the query design.
The SMS Layer Has Rules You Cannot Design Around
If the practice's existing client-communication product sends the messages, this section is its problem. If you are sending them yourself from a US 10-digit number, it is yours, and it constrains the build.
Twilio's compliance documentation defines A2P 10DLC as "the standard that United States telecom carriers have put in place to ensure that SMS traffic to US end users through long code phone numbers is verified and consensual." Registration has two parts: you "Create a Brand" describing the sender, then "Create a Campaign" detailing opt-in and opt-out procedures and message purpose. Unregistered senders, the docs state, "will receive additional carrier fees for sending unregistered traffic."
The throughput tiers matter directly for batch recall. From the same page:
- Sole Proprietor brand: "1,000 SMS segments and MMS per day to T-Mobile"
- Low-Volume Standard brand: "Up to 2,000 SMS segments and MMS per day to T-Mobile"
- Standard brand: "From 2,000 and up to unlimited SMS segments and MMS per day to T-Mobile, depending on your Trust Score"
Twilio also notes that Standard and Low-Volume Standard brands require a Tax ID, with each tax ID supporting up to five brands.
Note the unit: segments, not messages. A message long enough to split into two segments consumes two. A recall message with a booking link, a clinic name and a compliant opt-out line is very easily a two-segment message, which halves your effective daily ceiling. Draft the copy against the segment count, not against a word count. Our A2P 10DLC entry and the fuller Twilio A2P 10DLC guide cover registration in detail.
The consent requirement is the part people try to route around and should not. A client who gave a phone number to book a consultation has not necessarily consented to a marketing recall about a product. Where the message is genuinely a care reminder for an animal under the practice's care, most clinics treat that as within the treatment relationship; where it shades into promotion, treat it as marketing and hold a real opt-in. If that distinction is unclear for a given campaign, it is a question for the practice's own compliance position, not for an automation builder to decide unilaterally. And STOP handling must be wired into the suppression list itself, not just handled by the carrier - a client who opts out of one campaign should disappear from all of them.
A Worked Example, With the Arithmetic Shown
This is a model, not a case study. We have no client results to report and would not report them if we did. The point of showing the arithmetic is that you can substitute your own numbers and see immediately whether the build is worth doing - and for some clinics it will not be.
Inputs you supply from your own system. A two-vet practice pulls a lapsed cohort and finds 640 patients meeting the definition in the previous section. Average value of the appointment type being recalled - say an annual health check with vaccinations - is a figure the practice reads off its own invoicing, not one we can supply. For this model we will call it $130.
The send. 640 patients, released in tranches of 60 per weekday over about two weeks, keeping the front desk inside its inbound capacity. Assume the copy runs to two SMS segments. That is 1,280 segments total, roughly 120 per day - comfortably inside even a Sole Proprietor brand's 1,000-per-day ceiling to T-Mobile, so throughput is not the binding constraint here. Note the phrasing: this is a capacity calculation, not a prediction that anything gets booked.
The response, expressed as a break-even rather than a forecast. We are not going to tell you a response rate, because the honest answer is that it depends on how long the cohort has been lapsed, what the message offers, and whether the practice has any relationship left with those clients. What we can do is invert the question. If the build costs, say, $2,400 once and $70 a month to run - figures you should replace with real quotes, and our AI automation cost page explains what drives them - then the first-year cost is $2,400 + (12 x $70) = $3,240.
At $130 per recovered appointment, break-even is $3,240 / $130 = 24.9, so 25 attended appointments in the first year. Against a cohort of 640, that is 25 / 640 = 3.9 percent. The question you are actually deciding is not "will this work" but "over a full year, is a 3.9 percent recovery rate on a list of lapsed patients plausible for this practice?" That is a question a practice manager can answer from experience, and it is a far more honest framing than any percentage we could quote at you.
Two adjustments that change the answer materially:
- Recurrence. A recovered client who returns to an annual cycle is worth the visit again next year, and the year after. If you believe half of the recovered appointments become returning clients, the effective break-even roughly halves in year two onward, because the build cost is already paid.
- The cohort is not renewable at the same size. The first run works a backlog that has been accumulating for years. The second run, six months later, works only what has lapsed since. Do not model year two on year one's list size - model it on your monthly lapse rate. This is the single most common error in recall business cases and it makes ongoing automations look worse than the one-off sweep that preceded them, which is backwards.
Run the same arithmetic with your own average appointment value and your own cohort size before committing. If break-even needs a recovery rate you would not bet on, the correct decision is to not build it - or to build the far cheaper front-desk capture first, since a missed-call text back works on demand that is already trying to reach you.
What Not to Automate in a Veterinary Practice
Some of this is judgement and some of it is not negotiable.
Anything clinical. A client replying "she's been vomiting since last night, should I still come Thursday?" is a triage question. It goes to a person, immediately, with the whole thread attached. Do not let a language model decide urgency about an animal. Write this as a hard routing rule on keyword and on any inbound that is not a clean confirm or cancel, not as an instruction in a prompt - prompts are probabilistic and this rule should not be.
Euthanasia and end-of-life follow-up. These sequences exist in some practices and they should be human-composed and human-sent, every time. The automation's only job here is to make sure the client is suppressed from everything else.
Price negotiation on a declined estimate. An automated discount offer trains clients to decline first. Route the conversation to a person and let them decide what is appropriate for that client and that case.
The first batch of anything. Every recall cohort gets a human eyeball on the list before the first send, and ideally on every send for the first month. What you are checking for is not message quality, it is the record that should not be there. You will find one. Everybody does.
The general principle is in when not to use an AI agent, and it applies with unusual force in a setting where the customer is emotionally attached to the subject of the appointment.
A Rollout Order That Does Not Blow Up
In sequence, with a gate at each step:
- Read-only first. Connect, authenticate, pull the cohort, and send the list to a spreadsheet and nowhere else. Run it for a week. Have the practice manager mark every record that should not be on it. Fix the query. Repeat until the list is clean. No messages have been sent yet, and this is where the project is actually won.
- Manual send, automated list. The automation produces the list and the copy; a human presses send. This is a real, valuable, shippable stage and some practices should stop here permanently.
- Automated send, small tranche, human approval per batch. Add idempotency keys and suppression checks. Cap the daily volume well under whatever your A2P tier allows.
- Reply handling. Clean confirmations and cancellations write back to the PIMS. Everything else routes to a human. Measure how much lands in "everything else" - if it is above a fifth, your copy is ambiguous.
- The fill path. Only now connect cancellations to offers. This is the step that turns a cost centre into a revenue one, and it needs the previous four to be boring before it is safe.
Two systems worth checking before you build any of it. First, whether the practice already pays for a client-communication product that does most of this - many do, and the honest answer is often "turn on what you already own." Second, whether the PIMS's own automatic communication tasks cover the confirmation case, as ezyVet's support documentation indicates they can. Building an external integration to replicate a checkbox is a common and expensive mistake, and our automation services page is explicit that we would rather tell you that than bill you for it.
If you want a second opinion on which of these is worth doing at your practice, the free automation audit is a structured walk through your current stack. If you are still working out what any of this should cost, start at AI automation cost and the automation build cost page.
FAQ
- Does ezyVet have a public API I can build against?
- Yes. ezyVet publishes a developer portal at developers.ezyvet.com documenting an OAuth 2.0 bearer token model, a token endpoint at POST /v1/oauth/access_token, and endpoint families including Appointment Management, Contact Management and Wellness Management. Tokens have a documented TTL of 12 hours, so your integration needs an automatic refresh rather than a static credential. The docs list a Request Throttling section, but we could not get the numeric limits to render from a public fetch - confirm those in the live docs or with apisupport-ezyvet@idexx.com before scheduling large batches.
- Can I integrate with Covetrus Pulse, AVImark or Impromed the same way?
- Not on the same terms. We found no public developer portal for those systems. Covetrus publishes a partner programme instead - its own page describes Covetrus Connect as "The only approved and supported way to integrate with Covetrus practice management software" and lists Ascend, RxWorks and VisionVPM among covered systems. Practically that means an application and a commercial relationship rather than a signup, so plan for a scheduled export or the practice's existing communication product as the near-term path.
- How many recall texts can I send per day?
- If you send over a US 10-digit long code, Twilio's A2P 10DLC documentation puts a Sole Proprietor brand at 1,000 SMS segments and MMS per day to T-Mobile, a Low-Volume Standard brand at up to 2,000, and a Standard brand from 2,000 upward depending on Trust Score. The unit is segments, not messages, so a two-segment recall message halves your effective ceiling. In practice the binding constraint is usually not the carrier - it is how many replies your front desk can answer in a morning.
- Should confirmations and recall use the same phone number?
- There is a real tension here. One number is simpler for clients and keeps the conversation history in one thread. Separate numbers let you keep transactional confirmations away from anything that could be classed as marketing, which matters for both deliverability and consent hygiene. Most single-site practices are better served by one number plus rigorous suppression, because two numbers means two registrations, two inboxes and two places for a reply to get lost.
- What is the biggest cause of recall automations failing?
- Inbound capacity, not message quality. A batch that goes out too fast produces more calls and texts than the front desk can absorb, replies go unanswered, and clients conclude the practice does not follow up - which is worse than not having messaged them. Release in daily tranches sized to spare capacity. The second biggest cause is a suppression list that was written after the first send rather than before it.
- How do I prove the automation worked rather than the season?
- Hold back a control group. Take 10 percent of each cohort, message them with nothing, and compare booked-and-attended rates against the treated group over the same window. Without a control you are comparing against a period, not against a counterfactual, and vaccination and parasite recall demand is strongly seasonal. Ten percent of the upside is a cheap price for a number you can actually defend.
- Can an AI agent handle the replies to a recall campaign?
- It can handle the clean ones - a yes, a request for a different day, an unsubscribe. It should not handle anything clinical, anything about a deceased animal, or any conversation about price on a declined treatment. Write those as deterministic routing rules on keywords and on any non-matching reply, rather than as instructions in a prompt, because a routing rule fails the same way every time and a prompt does not.
- Is it cheaper to use the practice's existing client-communication product?
- Very often, yes, and it is the first thing to check. Many practices already pay for a product that sends confirmations and reminders and has a supported connection to their PIMS. Building a bespoke integration to replicate a feature already included in a subscription is a common and avoidable expense. The case for a custom build is strongest where the cohort logic is specific to the practice - the recall queries in this article - rather than where it is a standard confirmation sequence.
If this resonated
Module 02 is the long version of this post.
Eval set, refusal pattern, kill switch. Build it in a weekend.
