Triage and Remote Assessment: Making Teledentistry Triage Software Earn Its Keep

Most practices that buy teledentistry triage software buy it for the wrong reason. The pitch is usually “see more patients”, and the deck shows a graph going up. What actually changes, in the practices where this works, is that the emergency book stops being a lottery. The 09:10 caller with a debonded bracket stops occupying a slot that the 09:40 caller with a spreading submandibular swelling needed, and your nurse stops making thirty judgement calls a day that she is not registered to make.

That is the real product: a defensible, auditable sorting layer between the patient’s problem and your chair time. Everything else, the AI radiograph reads, the automated call handling, the photo uploads, is machinery in service of that.

Three Jobs That Get Confused With Each Other

Vendors sell these as one category. Treat them as three purchases, with three different risk profiles and three different owners in your practice.

Sorting (triage proper). A patient reports a problem, structured questions and photos come back, and a clinician decides: same-day urgent, routine appointment, advice and self-care, or escalate out of primary care. Tools here include DentalMonitoring’s SmileMate, Toothpic, MouthWatch TeleDent, and, in a lot of UK practices, a well-built form in your existing patient comms module. Clinical risk: high. Owner: a named dentist.

Reading (diagnostic support on radiographs). Pearl’s Second Opinion, Overjet, VideaHealth, Diagnocat and the UK-developed AssistDent all sit in this bracket. They mark up bitewings and periapicals with suspected caries, calculus, periapical radiolucency and bone level. Clinical risk: medium, because the image was already taken and you were always going to read it. Owner: the clinical lead plus whoever runs your radiography audit.

Answering (front-desk admin). AI phone agents such as Arini, call analytics like Patient Prism, messaging automation in Weave, and note-drafting tools like Kiroku. Clinical risk: low until an AI agent starts giving clinical advice, at which point it is very high. Owner: your practice manager.

Buying all three from one supplier is rare and usually a compromise. Buying them separately means three contracts, three data processing agreements and three integration conversations, which is annoying but honest.

The Regulatory Floor You Need Before Go-Live

Do this part first, because it kills about a third of shortlists.

Medical device status. Software that makes a diagnostic or triage recommendation is a medical device under the UK Medical Devices Regulations 2002 (as amended). Under Rule 11 it is usually Class IIa, which means it needs a UKCA mark or, for now, an accepted CE mark under the ongoing recognition arrangements. Ask for the certificate number and the Approved Body, not a marketing claim. A tool that “supports clinical decisions” and has a Class I self-declaration is either miscategorised or carefully worded to avoid the claim, and you should know which.

IR(ME)R 2017. No AI tool can justify an exposure. The referrer, practitioner and operator roles remain human and named. If a workflow implies “the software decided a radiograph was needed”, your employer’s procedures are wrong. Remote assessment also cannot produce a radiograph, which constrains what you can conclude from it more than most vendors admit.

DTAC and clinical safety. If any part of the purchase touches NHS money or NHS data flows, expect the Digital Technology Assessment Criteria, plus DCB0129 from the supplier (their clinical safety case) and DCB0160 from you (your local one). You need a named Clinical Safety Officer. In a single-site practice that is usually the principal, and it takes a day of writing, not a week.

UK GDPR and ICO. Intraoral photographs are special category health data. You need a DPIA, a data processing agreement naming sub-processors, a stated retention period, and clarity on hosting region. Ask specifically whether patient images are used for model training, and whether opt-out is real or theatre. Several 2024 and 2025 contracts in this space contained a broad training licence in clause 12 of the DPA that the sales team had never read.

GDC scope of practice. A dental nurse cannot diagnose or triage clinically. She can collect, chase, format and route. The decision must be recorded against a dentist, or a therapist working within scope. Build this into the software’s role permissions on day one, not after a complaint.

Remote orthodontics. The GDC’s position is unambiguous: a clinical examination, including radiographs where indicated, is required before orthodontic treatment. Photo triage can route an ortho enquiry. It cannot substitute for the assessment.

What The Triage Questions Should Actually Ask

Bad triage forms ask “describe your problem” and get back “toothache”. Good ones ask questions where each answer changes the routing.

A working set, roughly in order of decision value:

1. Any difficulty breathing or swallowing, or is the swelling closing your eye? (Hard stop. Immediate call, same-day maxfax or A&E. This must never enter a queue.)
2. Swelling: none / gum only / face or neck. Plus: is it spreading since yesterday?
3. Temperature over 38°C, or feeling generally unwell?
4. Pain now, 0 to 10, and pain at its worst overnight.
5. Does it wake you, or stop you sleeping?
6. Trigger: cold, heat, biting, spontaneous, nothing.
7. Trauma in the last 48 hours, and was the tooth displaced or knocked out?
8. Bleeding that has not stopped after 20 minutes of pressure?
9. Anticoagulants, bisphosphonates, recent head and neck radiotherapy, immunosuppression?
10. An ulcer or lump present for more than three weeks?

That last one matters disproportionately. A persistent non-healing lesion over three weeks is a two-week-wait suspected cancer referral, and a triage system that buries it under “sore mouth, low pain score” is the single worst failure mode in this whole category. Hard-route it.

Notice what is missing: nothing asks the patient to self-diagnose. “Do you think you need a filling or an extraction?” produces noise and, worse, an expectation you then have to manage in the chair.

A Worked Monday Morning

Here is an anonymised queue from a five-surgery mixed practice in the North West, 7,200 active patients, 2.1 FTE dentists on NHS contract plus two private associates. Requests arrived Saturday 08:00 to Monday 07:30 through a photo triage form linked from their website and their voicemail message. The clinical lead worked the queue between 08:00 and 08:26 on Monday.

QUEUE EXPORT  Mon 09:40  (26 min clinician time, 14 items)

# PT     FLAGS                    PHOTOS  DECISION              SLOT
1  BK    swell-face, T38.4, spread   3    URGENT same-day       09:20 today
2  RS    ulcer 5wk, painless         2    2WW referral raised   n/a
3  MA    trauma, avulsed 21          4    URGENT same-day       09:50 today
4  JD    pain 9/10, wakes, cold      2    URGENT same-day       11:30 today
5  PC    bracket debond, no pain     3    ROUTINE ortho         Thu 15:00
6  LT    lost crown 46, no pain      2    ROUTINE               Wed 10:20
7  HN    pain 6/10, biting only      2    ROUTINE               Tue 16:40
8  OW    bleeding socket, stopped    1    ADVICE + safety net   n/a
9  SF    denture rub, ulcer 4d       3    ADVICE + review 10d   n/a
10 GR    sensitivity cold, 3/10      2    ADVICE (desens)       n/a
11 TM    swell-gum, pericoronitis    3    URGENT same-day       14:10 today
12 AE    pain 8/10, warfarin, INR?   2    URGENT same-day       15:00 today
13 CB    tmj click, no lock, 2/10    0    ROUTINE               Fri 09:00
14 YP    food packing 37, 1/10       2    ROUTINE               Tue 11:00

Five urgent same-day, one two-week-wait referral, five routine, three advice-only with safety-netting. Before the form existed, those fourteen people would have generated roughly twenty-two phone calls across Monday morning, and the two genuinely time-critical cases (BK and MA) would have been booked on a first-come basis alongside the bracket debond.

Two numbers from their first six months. Emergency slot utilisation went from 71% to 94%. Failure-to-attend on emergency slots fell from 18% to 7%, which they attribute to the fact that a patient who has already photographed their own mouth and received a personal reply is committed in a way a 08:02 phone booking is not.

One number that went the wrong way: advice-only cases earn nothing. Three NHS patients given self-care advice and a safety net generate no FP17 and no UDAs. The business case here is released chair time, not new income, and if you present it to your associates as revenue you will lose their trust in week three.

Photos: What They Can And Cannot Tell You

A retracted intraoral photograph on a 2022-or-later phone camera is diagnostically useful for a specific, limited list: fractured cusps, lost or fractured restorations, debonded brackets and appliances, the anatomical extent of swelling, soft tissue lesions, gross calculus, denture fractures, and gingival colour and contour.

It tells you nothing reliable about pulpal status, pocket depth, caries beneath an existing restoration, apical pathology, or bone level. Anyone claiming otherwise is selling something. This is why the correct mental model is routing and prioritisation, not diagnosis.

Image quality is the constraint, and it is almost never resolution. It is retraction and light. Practices that post a pack of two disposable cheek retractors, typically £1 to £2 each, to patients who submit a first-round photo that cannot be read see a sharp improvement on the second attempt. Asking patients to use their phone torch reflected off a mirror works about half the time. Standardising the shots (upper arch occlusal, lower arch occlusal, right buccal, left buccal, anterior retracted, plus one of the problem area) turns a pile of ambiguous close-ups into something you can actually compare against next month’s set.

The full setup, including consent wording, the six-shot sequence, SLA timings, who chases and how the images land in the patient record, is covered in detail at /photo-triage-workflows/. Read that before you write your own form, because the sequencing of consent relative to upload is the part people get wrong.

Radiograph AI Is A Second Read, Not A Triage Engine

Keep this firmly separate in your head from remote assessment. AI radiograph tools operate on images you have already justified and taken, in the surgery, on a patient in the chair.

The evidence base is better than most dental software but thinner than the marketing. The ADEPT study, published in the British Dental Journal in 2021, tested AssistDent on enamel-only proximal caries in bitewings and reported dentists’ sensitivity rising from roughly 44% to about 76% with AI assistance, while specificity fell from around 88% to 71%. That trade is the whole story in one sentence. You will find more early lesions and you will also be pointed at more things that are not there.

Which means the operational question is not “is it accurate” but “what do I do with a flag”. Decide in advance and write it down:

Two governance points that come up in complaints. First, an AI flag you dismissed is part of the record, so your software must log the dismissal and it must be defensible; “the tool flagged D2 caries on 26 and the dentist ignored it” reads badly without a recorded reason. Second, never show the patient the AI overlay before you have reviewed it. A red box on a screen is extraordinarily persuasive, and a patient who has seen eleven red boxes and is then told nine of them are nothing has lost confidence in you, not in the software.

Radiograph AI also has an unglamorous use that is arguably its best: retrospective audit. Running last year’s bitewings through a tool and comparing its output against what was actually diagnosed and treated is a genuinely uncomfortable, genuinely useful exercise, and it costs you nothing beyond the licence you are already trialling.

Front Desk: Where The Volume Actually Is

The same practice above logged 780 inbound calls in a representative week. Breakdown, from their phone system reporting:

• 31% appointment booking or rescheduling
• 18% “what time is my appointment”
• 14% new patient enquiries, of which about half asked whether the practice was taking NHS patients
• 12% problem or pain calls
• 9% payment, plan and insurance queries
• 8% treatment questions for a clinician
• 8% other, including suppliers and labs

Only 12% needed clinical judgement on the call. Roughly half the total was answerable from the practice management system without a human, and 41 calls that week went unanswered entirely between 08:00 and 09:00.

An AI phone agent or a decent online booking portal attacks the top three lines. The line to hold, absolutely, is that the agent must not triage. It can collect, it can book into a pre-defined emergency slot type, and it can hand off. The moment it says “that sounds like it might just be sensitivity” you have a clinical incident with an audio recording attached. Ask any vendor to demonstrate what happens when a caller says “the side of my face is swollen and I can’t open my mouth properly”. If the answer is not “immediate transfer or immediate callback with a hard flag”, walk away.

Set the call-recording and retention policy at the same time. Recordings are personal data, they are discoverable, and six years of unindexed audio is a liability rather than an asset.

Integration: The Part That Delays Go-Live By Six Weeks

Your practice management system determines how painful this is.

Dentally publishes a documented REST API, which means a triage tool can genuinely write a note, attach an image and create an appointment. Expect this to work.

SOE Exact and Carestream R4 are more closed. Integration typically runs through the vendor’s partner programme, and in practice a lot of triage tools end up with a one-way flow: the clinician reviews in the triage tool, then copies a summary into the patient record and attaches the images manually. Budget 60 to 90 seconds per case for that, which on fourteen cases is twenty minutes of nurse time daily, or about 80 hours a year.

Whatever the mechanism, insist on three things. The triage decision and its reasoning must land in the patient’s clinical record, not live only in a vendor’s cloud. Images must be attached to the record, at full resolution, not as a link to a portal you may not be subscribed to in four years. And you need an export path for everything, tested during the trial, because vendor churn in this sector is real and a triage history you cannot extract is a triage history you will lose.

Measure These Six Things Or Do Not Bother

Baseline for four weeks before go-live. Without a baseline, any vendor’s dashboard will tell you the tool is working.

That last metric is the one nobody tracks and the one that tells you the truth. Pull twenty triaged cases a month and ask: did the clinical findings in the chair match the triage category? If a case triaged as routine turned out to need same-day intervention, that is a near miss, and it belongs in your significant event log with the triage record attached. Three of those in a quarter means your question set is wrong, not that your clinicians are.

Costing It Honestly

Indicative UK figures, which you should treat as a starting point for negotiation and verify in writing per site rather than per group:

• Photo triage platforms: roughly £80 to £250 per site per month, sometimes per-clinician
• Radiograph AI: roughly £200 to £500 per site per month, occasionally with a per-image element
• AI phone agents: roughly £300 to £900 per month, scaling with call volume
• Your own time: 1 to 2 days for DPIA, DCB0160 and SOP writing, plus 2 hours per clinician on calibration

Against that, the NHS side. Urgent treatment is 1.2 UDAs. Band 1 is 1 UDA, and since 1 January 2025 Band 2 splits into 2a at 3 UDAs, 2b at 5 and 2c at 7. Patient charges for 2025/26 are £27.40, £75.30 and £326.70. If your UDA value is £32, six additional filled emergency slots a week across 44 weeks at 1.2 UDAs is about £10,100 of activity you were previously leaving on the table through unfilled and failed appointments. Private emergency assessments at £65 to £95 shift the same arithmetic considerably further.

Set against that: every advice-only case is £0 of NHS activity, and if your contract is already delivering above target, releasing capacity has no financial value at all and you should buy this purely for risk management and patient experience. Be clear which of those two arguments you are making, because they lead to different vendors and different budgets.

Run It As A 90-Day Pilot With A Kill Criterion

Agree the kill criterion before you sign. Something like: if triage-to-appointment concordance is under 80% at week 10, or if clinician time per case exceeds three minutes, we stop.

Weeks 1 and 2, no patients. Write the SOP, complete the DPIA and DCB0160, define the question set, define the three or four routing outcomes, and set the named clinician rota for working the queue. Weeks 3 and 4, calibrate: every clinician triages the same fifteen historical cases independently and you compare. Disagreement on more than three of fifteen means your categories are not defined tightly enough, and that is a cheap thing to discover before real patients are involved.

Weeks 5 to 8, live but narrow. One entry point only, typically the website form, and existing patients only. New patient triage adds identity verification and payment questions that will muddy your data.

Weeks 9 to 12, widen and audit. Add the voicemail route and the “we’re closed” out-of-hours message. Pull your twenty-case concordance audit. Ask reception, separately from the clinicians, whether their day got better, because if it did not, the tool will quietly stop being used by March regardless of what the dashboard says.

The Cases That Will Catch You

Every practice running remote assessment eventually meets these. Write the handling into your SOP now.

The patient who submits photos at 22:40 on a Friday and receives a reply at 09:15 on Monday, having spent the weekend in pain. Your out-of-hours message must state the response window explicitly and direct urgent problems to NHS 111 or your out-of-hours provider. A form that accepts submissions without stating when a human will look is an unsafe form.

The patient whose photos are unusable and who does not respond to a request for better ones. After two attempts, that becomes an appointment offer, not an open loop. Close it and record the closure.

The anxious patient who prefers remote contact and keeps submitting photos instead of attending. Remote assessment can accommodate anxiety for a while, and then it becomes avoidance you are enabling. Cap it, and name the cap in the notes.

The patient who photographs something incidental and alarming that they did not ask about. You have seen it and it is in your record, so you own it. This is an argument for a slightly wider consent wording and for a reviewing clinician who looks at the whole image, not just the arrow.

Pull ten triage records from your last month and read them as though you are opening a complaint file. Is the decision recorded, with a reason, against a named dentist? Are the images attached to the clinical record rather than sitting in someone’s cloud? Is the safety-net advice in writing, with what to do and by when? If any of those three is missing, that is the thing to fix this week, ahead of any new purchase.