A Tuesday morning at a four-branch independent optometry practice in Pretoria East. The owner is in clinic, three optometrists are seeing patients across the other branches, and the front desk has been running on a single receptionist since the second one resigned in March. Between 09:00 and 11:30 she has handled forty-one phone calls and roughly the same number of WhatsApps. Seventeen of those are "is my Discovery still going to cover this", twelve are "do you have the new ProDesign frame in black", four are "when am I due for my next eye test", and the rest are appointment moves. Three patients walked out of the waiting room because the till queue had stalled while she was dealing with a Bonitas claim rejection.
That morning, in a practice that turns over a respectable number per year, the owner is paying a clinical optometrist's hourly rate to sit between consultations and tell the receptionist whether a frame is in stock at Brooklyn Mall.
This is the practical starting point for an AI conversation in a South African optometry practice. Not robots replacing optometrists. Not autonomous refraction. Just taking the queue of small, structured, predictable patient traffic off the front desk so the clinic floor can run.
Where SA optometry practices actually lose time
Across the independents, the small chains, and the franchise groups — Spec-Savers, Torga Optical, Mellins iStyle members, the family practices of two-to-five branches scattered across Pretoria, Joburg North, the Western Cape and the KZN coast — the same handful of items dominate the non-clinical workload:
- Two-year recall reminders for the standard eye-exam cycle, often handled with whichever marketing email blast goes out on a Friday and whichever patients happen to open it
- Medical-aid pre-authorisation and benefit checks across Discovery, Bonitas, Bestmed, GEMS, Polmed, Medshield, Profmed and KeyHealth, each with their own rules on frame allowance and lens cover
- Frame and lens stock queries on WhatsApp — close to half the inbound traffic at a busy branch, and nearly all of it is a database question
- Appointment booking and the inevitable reschedules when a patient's child gets sick or the parent forgets the time
- Script copies for sunglasses bought online, which has grown into a meaningful chunk of work since Sunglass Hut and the direct-to-consumer brands started asking for valid prescriptions
- Follow-up on lab orders — Essilor, Hoya, ZEISS Vision Care, the in-house edging room — keeping the patient informed during the seven-to-twelve working days between order and collection
Some of this work is clinical. Most of it is not. The job is sorting the two properly.
Recall reminders: where the recurring revenue actually lives
Most independent practices run on a two-year recall cycle. The patient came in, the optometrist tested, a script was issued, frames were dispensed, the file was closed. Two years later, the patient is due. In practice, this only happens reliably for the patients who walk in unprompted. For the rest, recall depends on whoever remembers to look at the spreadsheet on a Friday afternoon.
This is where AI repays itself fastest. The data is already there. Whichever practice management system you run — Optomate, FocusEyes, OPC, Optidrive — already holds the last-exam date for every patient on file. The gap is in outbound contact at the right moment, on the right channel, with the right tone.
A WhatsApp-led recall flow reads the due list at the start of each week, sends a polite nudge five weeks before the two-year mark ("Hi Mr Pillay, your last test with us in Greenstone was October 2024 — you're due for a check. Want me to find you a slot the week of the 16th?"), reads the reply, and if it's positive, books an appointment against the optometrist's calendar. If the patient ignores the first nudge, a softer second one goes two weeks later. If still nothing, the file flags for a human phone call — usually the receptionist, sometimes the practice owner if the patient is long-standing enough to deserve the personal touch.
In the four-branch practices we have watched run this honestly, recall conversion lifts by single-digit percentage points within a quarter, with the biggest gains in the over-50 cohort where lapses are most common and most clinically meaningful. Glaucoma surveillance, in particular, depends on recall holding.
A note on tone. Patients can tell the difference between a thoughtful nudge and a marketing blast. The thoughtful nudge respects their last exam date, their last frame purchase, and their preferred branch. The blast goes to the spam folder, or worse, ends the relationship. If you cannot personalise the recall to the patient's actual file, do not send it.
Medical-aid admin: the quiet drain on every branch
There is no medical scheme in South Africa that handles eye care the same way as the next one. Day-to-day benefit checks, single-vision-versus-multifocal cover, frame-spend limits, dependant codes, the exhausted-benefits patient who shows up still expecting the practice to cover the gap: every branch handles a dozen of these conversations a day.
A structured AI layer sitting between the receptionist and the BHF switch does not make clinical decisions. It does fetch the patient's remaining benefit on a known scheme, drafts the polite "here's where you stand" message in plain language, flags cases where the available benefit will not cover the planned dispense, and queues the gap conversation for the receptionist to have in person.
What the layer must not do: quote a private-pay top-up price, commit the practice to a specific frame discount, or rephrase a scheme rejection in a way that creates a contractual expectation. The receptionist still has those conversations. The system gives her the relevant numbers, drafted, on the right WhatsApp thread, in time for her to have the chat before the patient leaves.
For practices running across multiple branches, the second-order benefit is consistency. The Gateway branch and the Brooklyn Mall branch end up giving the same answer to the same scheme question, which is more than most independents currently manage.
Frame and lens stock queries: the simple win
About a third of the WhatsApp traffic at a four-branch practice is "do you have X frame in Y colour". The receptionist opens the stock system, checks, replies. Two minutes per query. Forty queries a day. That is more than an hour gone, every day, on a question a properly designed system can answer in a second.
The pattern is straightforward. An incoming WhatsApp arrives. The system parses the frame brand and model — handling the messy reality of how patients actually ask for frames (the customer who calls Ray-Ban Wayfarers "the Ray-Bans like Brad Pitt's", the one who refers to ProDesign as "those Danish ones I liked"). It checks live stock across all branches, replies with the answer, and if the frame is at another branch, offers to have it sent for fitting at the patient's preferred location.
Two boundaries to bolt in from day one. The system answers stock-on-hand only — never lens-and-coating recommendations, which is a dispensing optician's conversation. And anything that smells clinical ("my eyes have been blurry for three weeks, what frame do you recommend") routes immediately to a real human, with the full transcript, never resolved by the bot.
Independents that ship this end up with one to two receptionist hours back on the front desk every day. Translated into clinic output, that is more attentive walk-in handling, faster collections processing, and noticeably fewer patients leaving because no one greeted them within ninety seconds of arrival.
Where AI does not belong in an SA optometry practice
This is the part the HPCSA's Professional Board for Optometry and Dispensing Opticians will care about most, and it is the part most vendors skip past quickly. In my experience the line is drawn in roughly five places.
Refraction and clinical assessment. A bot does not refract. It does not interpret a fundus image. It does not decide whether a visual field defect is glaucomatous or anything else. The optometrist does. If a vendor pitches you "AI-led refraction" for an SA practice, read the regulatory framework carefully before signing anything — and ask them to walk you through the HPCSA scope of practice on the call.
Script issuing and modifications. The script is the optometrist's clinical decision, signed by them, stored on file. A bot cannot issue a script copy for a patient buying sunglasses online without the optometrist's confirmation that the script is still valid and the patient is the right one. The copy-script request from a Sunglass Hut purchase feels like a customer-service ticket. It is also a piece of clinical documentation, and treating it as the latter at intake saves complaints later.
Specific lens recommendations. Single-vision versus office lens versus progressive, Crizal versus a basic AR, transitions versus polarised — these are dispensing-optician conversations. The system can show options on a screen. The recommendation belongs to a qualified person.
Children's data. Section 32 of POPIA covers personal information of children, and optometry practices handle a lot of it during the January back-to-school rush. Any automated flow touching a child's record needs explicit parental consent, captured in writing and tied to the patient file. Skipping this because it is awkward is how complaints end up at the Information Regulator.
Margin-led product nudging. Do not let the bot push patients toward higher-margin frames or coatings. The audit trail will exist. The HPCSA can ask for it. The trust a practice has built across years dies quickly when patients work out that the recommendation was algorithmic.
A sensible first project for an SA optometry practice
If you run a single branch or a small group, the project that pays back inside a quarter is almost always the recall layer, sitting in front of your existing practice management system. The data is there. The patients are on WhatsApp anyway. The compliance lift is measurable per age bracket. No clinical decision is being delegated.
Run it for ninety days. Track recall conversion by age cohort against the previous comparable quarter — under-30, 30-50, 50-65, 65-plus. Track the no-show rate on booked recalls (the AI nudge tends to halve it, because the patient consciously chose the slot). Track the glaucoma-surveillance subgroup separately; that one matters clinically as well as commercially.
If the recall layer works, the next two extensions in order are the WhatsApp stock-query handler and the medical-aid benefit-check drafter. Each one builds on a base the team already trusts, rather than another platform to learn.
What to leave for later, or for never: anything that touches refraction, clinical recommendation, or script issuance without an optometrist in the loop. The line is not arbitrary. It is the line the PBOC drew, and crossing it costs more than any efficiency it could ever pay back.
What an honest AI layer does in an optometry practice is take the queue of small, structured, repeatable traffic that quietly eats the front desk's day and route it through a system the practice actually controls. The chair stays the optometrist's. The dispensing conversations stay the dispensing optician's. What changes is the practice's ability to show up consistently for the patient between consultations.
For a four-branch independent fighting for the same patient against the Spec-Savers down the road, that consistency is the only sustainable edge.