Every metric we track — published. Real accuracy, latency, cost, and provider numbers from production.
No other dental SaaS — Dentrix, Eaglesoft, Practo Ray, Drlogy, Cliniify, Carestream, or any of the new “AI dental” entrants — publishes their AI accuracy, cost, latency, or provider numbers. You either trust their marketing copy, or you don't. There is nothing in between.
Dentospire is built by a practising dentist for clinics that handle real patient data under DPDP Act 2023 and UK GDPR. We think you deserve to see the numbers. So we publish them — every AI call is logged, every cost is summed, every provider fallback is tracked, and this page reads straight from that table. If our numbers ever get worse, you'll see it here before we tell you.
Aggregated from ai_usage_events (per-call telemetry table). 36,016 calls logged in the last 30 days.
If any single AI provider is unavailable — rate-limited, quota-exhausted, returning errors, or globally degraded — Dentospire automatically routes to the next provider. We track and publish per-provider success rates.
| Provider / model | Calls (30d) | Share of traffic |
|---|---|---|
| gemini-2.5-flash | 34,566 | 51.9% |
| claude-haiku-4-5 | 18,004 | 27.0% |
| groq-gpt-oss-120b | 6,961 | 10.4% |
| groq-llama-3.3-70b | 6,457 | 9.7% |
| unknown | 514 | 0.8% |
| gemini-3.1-pro-preview | 61 | 0.1% |
| llama-3.2-90b-vision | 46 | 0.1% |
| gemini-3-flash-preview | 5 | 0.0% |
Read from ai_corrections. Accuracy = 1 − (dentist edits / total runs). Whenever a dentist accepts an AI suggestion as-is, it's logged as correct; any change, deletion, or supplied missing answer is counted as an edit. Window: last 30 days for the trend; lifetime for per-entity totals.
| Feature | Total runs | Dentist edits | Accuracy |
|---|---|---|---|
| differential | 6 | 0 | 100.0% |
| clinical_advice | 1 | 0 | 100.0% |
Aggregated by prompt_hash (first 16 chars shown — the full hash never leaves the server). Lower accuracy here = candidates for prompt refinement or model upgrade.
| Prompt hash | Runs | Edits | Accuracy |
|---|---|---|---|
| da53b57de8e7079e… | 6 | 0 | 100.0% |
| d2555186308513cb… | 1 | 0 | 100.0% |
Every AI call writes to ai_usage_events. The numbers above read directly from that table.
Model name, input/output tokens, cache tokens, latency, cost, success flag, error code.
All PHI is redacted via redactPII() and verified by assertNoPhiLeak before any external AI provider call.
Full details in our Data Processing Agreement and Sub-processors list.