Designing trust into a pet health AI.
A mobile companion for the moment you look at your dog's mouth, sense something's off, and don't know what to do with the feeling.
The problem
Owners know oral health matters. They just don't know what to act on. The brief said education gap. Research said action gap.
Who it's for
Pet parents in a moment of doubt. Vets who'd rather receive a specific question than a vague worry. Mars teams scaling the pattern to other scan products.
The outcome
A worry becomes a structured next step. Non-diagnostic signal at population scale. The pattern Mars now reuses for PoopScan, PawScan and beyond.
A 30-second visual signal
A guided photo, four owner-reported symptoms, a tooth-by-tooth map, and a one-tap route to a vet.
A diagnosis
It never names a disease, never replaces a vet, and never functions as permission to skip a check-up.
The problem
Awareness is high. Action isn't.
Owners already know oral health matters. Three-quarters of them. But the signs feel ambiguous, the advice feels generic, and the next step feels optional. What's missing isn't information. It's a moment of personal, visible, actionable signal.
80% of dogs over 3 have periodontal disease · 72% of owners rate their dog's teeth as good or perfect · 4% brush regularly.
Core principles
Three principles that shaped the design.
Each one started as a question against the research, and ended as a constraint I'd defend in a portfolio review.
Refusal as design
How might we make every model failure feel like a path forward rather than a dead end, by designing five refusal states with the same care as the success states?
Owner instinct as signal
How might we treat the pet parent's own observations (foul breath, discomfort, bleeding) as data worth capturing alongside model output, not as decoration?
Activation, not verdict
How might we ensure every terminal state (alert, clear, or refused) ends on a real next step toward a vet conversation, never an editorial conclusion?
See it work
Six surfaces, thirty seconds, one decision.
Scroll through the flow. The device stays. The screen changes as each step takes its turn. Every step earns the design choice behind it.
Set the contract first.
Before any data is collected, the product names what it is, and what it isn't. An informational tool, not a diagnosis. The disclaimer link sits on the very first screen, not buried in settings.
Trust is offered, not assumed.
The disclaimer is content, not friction.
Tapped, the disclaimer opens as a full modal: plain language, real legal weight, no dark patterns to bypass it. The model methodology lives one tap from the home screen.
If the product can't be honest on the first screen, the rest is wasted.
Teach the photo, not just the disease.
A good model needs a good photo. Three illustrated steps show owners exactly how to frame the shot: back teeth included, lips pulled back, fur out of the way. Self-correction happens here, before the model has to refuse anyone.
Friction up front saves frustration later.
The owner's instinct is data too.
Four questions ride alongside the model: name, foul breath, blood, discomfort. Owner-reported answers surface in the result and in the vet handoff. Each "yes" reveals a short, evidence-backed explainer. Education happens in the flow.
A consumer health tool that ignores the consumer's own observations is wasting its richest input.
Show what the model saw.
Results are structured, not editorialised. Teeth analysed. Teeth with visible signs of tartar. Areas with possible gum irritation. Each number is a count, not a verdict. The owner's questionnaire reads beside the model's findings, so both signals can be weighed together.
We flag, suggest, share. We never diagnose.
Hand the result to someone who can help.
Results don't stop at the screen. Owners get two structured next steps: chat live with a credentialed vet technician, or email a copy of their results to take into their own clinic. Both routes carry Scout's full context: questionnaire answers, model findings, the photo itself.
A vague worry becomes a specific question, and that's where this product actually does its job.
The hard call
How do you say something useful without saying anything diagnostic?
The result page had to satisfy three constituencies, each pulling in a different direction. Pick any two and you ship a product that fails the third.
"Just tell me if my dog has gum disease."
Wants a clear answer they can act on.
"Don't pretend you can replace me in the consult."
Wants the model to defer to clinical judgment.
"Don't make a clinical claim you can't defend."
Wants zero diagnostic language anywhere in the UI.
The result page shows counts, not verdicts. Per-tooth tartar count. Per-area gum signal. Every owner-reported symptom carried into the handoff. The language is structural throughout: flag, suggest, share with your vet. We never use "has" as a conclusion, and the empty state ends on a vet CTA so a clean scan doesn't become permission to skip the consult.
The owner gets clarity. The vet stays the authority. Regulatory keeps its line. Nobody got the product they wanted; everybody got the product they could defend.
Under the hood
Nine stages, eight models, one photo.
Owners experience a single, calm flow. Underneath, the image passes through nine model stages: three that decide whether to proceed, four that extract signal, two that quality-check before anything reaches the screen. Each stage has a UI consequence.
Receive image
Capture or upload
Dog finder
Is this a dog?
Body part
Is this a mouth?
Tooth ID
Locate each tooth
Tartar?
Per tooth
Gum ID
Locate gingiva
Irritation?
Per gum line
Quality
Blur, noise, light
Return
API → UI
Vet-supervised training
Every training image labelled by a licensed vet. Outputs map to clinically meaningful signal, not classes the model invented.
3rd-party scientific governance
External review of methodology and performance. Mars publishes its own model methodology, a designed surface, not a hidden disclaimer.
Designed refusal
Stages 02, 03, 08 can stop the flow. Each refusal has its own UI state, written plainly. Better to say "we can't" than to guess.
The finding we didn't expect
A consumer tool became a research instrument.
"How do we transform awareness into action? By making oral health visible."
Mars Petcare · 2024 product conference
That framing wasn't just marketing. Submissions started revealing patterns no clinic data could show, and one of them validated the entire design hypothesis.
where owners reported foul breath
where owners reported no symptoms
That's the design unlock. ToothScan didn't override owner instinct. It validated it, structured it, and gave it somewhere to go.
17 markets · methodology published on mars.com · reopening the question of true periodontal prevalence in dogs
Impact
From awareness to action, at population scale.
Markets live
EU + Asia + Americas. Same flow, localised: same disclaimer hierarchy, same handoff, same refusal states.
Photo → PDF
Median end-to-end time, including the questionnaire. Coffee-table fast.
Diagnostic claims
Across every result state in every locale. Audited with vets, then regulatory.
Of scans in 6 markets surfaced signal
Reopening the question of true periodontal prevalence in the general dog population.
Reflection
I came in thinking the design problem was the model. I left thinking it was the disclaimer.
The model is the easy part of an AI product. What's hard is the thirty seconds before any data is collected: the disclaimer, the photo instructions, the language on the result page. That's where the product gets earned, or lost.
Every AI product in a high-stakes domain has a moment where it has to tell the user what it isn't. If you skip that moment, no amount of clever model output will save you. If you nail it, the rest of the design becomes possible.
For the next product I build, I'd start with the disclaimer and work backwards.
I'd embed a veterinarian on the design pod from week one, not bring them in as content reviewers later. We rewrote the result-page language three times: once for clinical accuracy, once for owner accessibility, once for regulatory defensibility. With a vet co-designing from day one, those constraints become inputs to the first sketch, not corrections to the third.
A longitudinal loop. Today ToothScan is a snapshot. The same infrastructure could give owners a six-month view of their dog's mouth, and turn a one-shot tool into a habit.