Appointment Scheduling Redesign
One shared language for what an appointment is, readable by the clinics using it and the business running them.
- Company
- IVC Evidensia
- Role
- Senior Designer
- Date
- July 2026
- Timeline
- 1 Month
A note on this one: it's live, ongoing work under NDA, so the product itself stays out of view, and figures are kept directional. What you'll find here is the reasoning, the trade-offs, and why they mattered.
I was asked to restructure the appointment types in a veterinary practice management system (PMS). Classifying 109,202 real bookings showed the categories weren't just messy, they were making it impossible to measure how clinics were actually performing.
Problem
One list nobody could use, one box doing eight jobs
Every appointment started the same way: pick a type from a flat list, then type what was actually happening into a free-text Subject box. Both halves were failing together.
The list had no search, so it got slower as it grew and staff fell back on the handful of types they knew. The types overlapped: a telemedicine chronic-illness follow-up genuinely matched three at once, so there was no right answer to train anyone toward. One booking in seven ended up in a generic "Consult - Other".
Everything the list couldn't express went in the Subject box: reason for visit, vaccines due, add-on services, risk flags, drop-off logistics, booking origin, pricing, operational notes. Eight jobs, one field, two languages, no structure.
The easy explanation was that staff needed training. Type usage five months apart was almost identical. That is not what a training problem looks like.
109202
Appointments classified across 302 Canadian clinic databases
85%
User interviewed flag appointment types as a top complaint
123
Distinct appointment types
1 in 7
Bookings in one generic "Consult - Other" catch-all
8
Identified unrelated jobs done by a single free-text field
Business goal
The part nobody was talking about
A list that is hard to use is an annoyance. A list that is used incorrectly is a measurement problem, and that is the more expensive failure.
Appointment data is how the business understands whether its clinics are running well: how vet time is used, which services are delivered, where capacity is lost. When staff pick whichever type seems closest and put the real content in a free-text box, none of that can be measured reliably. Vet utilisation, the most important operational metric a clinic has, was being calculated from categories that did not mean what they appeared to mean.
Automated reminders, calendar filtering, online-booking policies and cross-region reporting were all blocked on the same thing: a shared definition of what an appointment is. That definition now exists.
Taxonomy agreed, mapping in progress. Form prototyped and validated in comparison testing. No live adoption data yet.
Eight jobs jammed into one field
8 jobs crammed into one field
Business levers
× 3
Reason for visit
→ Service utilisation & revenue
Add-on services
→ Revenue per visit
Booking origin
→ Channel performance
Clinical & operational
× 3
Vaccines due
→ Reminders & medical compliance
Behavioural & medical flags
→ Clinical safety for animals & staff
Patient admit
→ Calendar automation & record integrity
Deliberately unstructured
× 2
Operational notes
→ Stays genuinely free-form
Pricing notes
→ Discounts & promotions live in invoices
How AI sat in the process
AI didn't replace a step in my process. It changed what was affordable at three points. The meaning and the decisions were never delegated. They stayed with me.
The honest alternative at stage two was never a slower analysis. One designer doesn't get weeks to hand-build a bilingual classifier. The alternative was no analysis and a redesign argued from anecdote, against stakeholders who each had a different view.
The tools widened what one designer could look at. Every column on the right stayed a decision.
Structure without flattening
The receptionist captures the owner's world. The system routes it to the vet.
The narrative categories had the highest uncategorised rates in the data: 68% for sick pets, 56% for emergencies. It wasn't the classification breaking down, it was an actual finding. Those weren't miscellaneous entries, but owner-reported symptoms in each receptionist's own words. "Not doing well", "Back end seems painful".
This was the opposite problem to the appointment types. Types were an overgrown list, and the fix was reduction. Symptoms were rich and varied, and any short list would have thrown that richness away.
So the fix was the reverse: a large controlled vocabulary of 300+ veterinary reasons for visit, surfaced through typeahead and best ordering in a new reason field. The receptionist starts typing what the owner said, the matching reason appears, one tap files it. Their experience "just type it" gets faster and what lands in the database is a value the system can count and act on.
Structuring the symptom is what turns it from a free text into a system input. Once "limping" resolves to a known clinical concept instead of a string of characters, the system has something to work with: it can suggest how to triage the visit, surface the right workflow and assist the first vet's diagnosis, before the patient is even in the room. This one decision is what makes the automation and clinical assistance on the roadmap buildable.
Capturing the symptom as structured data turned a receptionist's note into something the system can act on. It makes the difference between just capturing a pet is unwell and routing it toward the right assisted care.
A 300-term vocabulary needs an owner: new terms, retired terms, all of it in two languages. And if the matching fails, staff give up and type free text, which is the problem I started with. A short picklist would have been cheaper and faster, but it would have thrown away the owner's own words. I took the maintenance over the data loss.
Where the analysis was wrong
Every one of these took domain knowledge to catch, and each would have produced a worse taxonomy if I'd trusted the output.
It couldn't see the rest of the product
Hospitalization looked like a legitimate low-volume workflow. It isn't one: the software has a separate hospitalization module, and no appointment is ever booked as one. It becomes one during a consult, and staff were editing appointments afterwards to record it.
It mistook under-specification for noise
Grooming came back half uncategorised, apparently a junk field. In fact most entries said only "grooming" with no service type, which was genuinely missing structure. The rest was initials recording which groomer was assigned, and the form already had a field for that. Staff were writing it into the Subject line anyway, which is a flow that doesn't fit how they work rather than a field that doesn't exist.
The same number meant opposite things
A high uncategorised rate meant the boarding data was unreliable. It also meant the sick-pet content was real and had to be preserved. The numbers are nearly identical and nothing in the output separates them.
Data got me to seventeen, people to thirteen
The workaround that had become a category
The first taxonomy I could defend from data alone had seventeen categories. Thirteen is what survived contact with the people who run clinics. Five dropped through those conversations, and one, dental prophylaxis, deliberately added back.
The sharpest example was "Online booking", which appeared in the data as an appointment type. It never should have been one. Clinics only wanted online bookings to stand out visually in the calendar so front-of-house could review them. Because the software could only differentiate appointments by type, a type is what they were given. A workaround that had hardened into a category, and every online booking was consequently filed with no record of whether it was a wellness visit or a consultation. The fix was to make booking channel a property of the appointment rather than its identity.
It worked the other way too. One category was added back against the data: a dental prophylaxis has different prep and duration from a surgery and is a metric the business is pushing. Low volume today, deliberately separate tomorrow.
Takeaways
What I don't know yet
Comparison testing has run with receptionists across four clinics, and the design holds up: one flow won clearly, the workflow names validated, and time-to-book held.
That validates the design. It doesn't yet validate the decision. A receptionist filling a form in a scheduled session is not the same person filling it at 9am with two lines ringing, and this design adds structure to exactly that moment. The failure I'm watching for isn't rejection, it's speed damage: half-filled forms, or the fastest plausible option selected rather than the right one. For a system whose entire purpose is telling the business how its clinics actually run, a fast wrong value is worth no more than the free text it replaced.
What's missing is daily use over time: completion rates and value accuracy from live clinics, not from interviews.
What I would do differently
I would connect the AI tools to each other, not just use each one well. Analysis, synthesis and design each lived in a different tool with no shared context, so I became the integration layer, carrying findings between them by hand and re-explaining the same background every time. It works, but the thread depends entirely on me remembering it. Next time I would invest early in keeping that context connected across tools, so the reasoning survives the gaps between them rather than living only in my head.
The taxonomy is agreed and the design tests well. The part I'm still waiting on is whether it survives a Monday morning, because a form filled fast and wrong tells the business exactly as little as the free text it replaced.