# Precision > Precision measures, diagnoses and improves the quality of healthcare AI (AI scribes, voice agents, clinical chat, coding and prior-authorisation tools) using clinician-calibrated evaluation. It is delivered today as fixed-scope services; a self-serve platform is being built. Contact: hello@precision.example Key facts: - Quality Audit (Service, Available now): Your AI, tested on your encounters and graded by your clinicians. - Vendor bake-off (Service, Available now): Run AI vendors side by side on your own cases, not their demos. - Drift monitoring (Service, Available now): Find out within days when a model update makes things worse. - Evidence pack (Service, Early access): The proof that security and clinical-safety reviews ask for. - Language evaluation (Service, Early access): Clinicians who speak the language test your product in it. - Precision Platform (Product, Coming soon): The whole loop as software, for teams that outgrow audits. - Precision Local (Product, Coming soon): Runs inside your walls. Only numbers come out. - Precision De-ID (Product, Free tool): Mask identifiers in clinical text on your own machine. - Method: a five-layer closed loop. Clinicians' reactions specify the criteria; graders are calibrated against clinicians; encounters are replayed; causes are found by swapping one component at a time; fixes are accepted only if they hold on held-out clinician-labelled cases and are recorded in an evidence ledger. - Precision never makes or changes clinical decisions, never contacts patients, and does not certify that any product is compliant. Reports describe measurements performed. ## Solutions and pricing - [Quality Audit](https://precision.example/solutions/quality-audit): Service, Available now. Your AI, tested on your encounters and graded by your clinicians. - [Vendor bake-off](https://precision.example/solutions/vendor-bake-off): Service, Available now. Run AI vendors side by side on your own cases, not their demos. - [Drift monitoring](https://precision.example/solutions/monitoring): Service, Available now. Find out within days when a model update makes things worse. - [Evidence pack](https://precision.example/solutions/evidence-pack): Service, Early access. The proof that security and clinical-safety reviews ask for. - [Language evaluation](https://precision.example/solutions/language-localisation): Service, Early access. Clinicians who speak the language test your product in it. - [Precision Platform](https://precision.example/solutions/precision-platform): Product, Coming soon. The whole loop as software, for teams that outgrow audits. - [Precision Local](https://precision.example/solutions/precision-local): Product, Coming soon. Runs inside your walls. Only numbers come out. - [Precision De-ID](https://precision.example/solutions/precision-deid): Product, Free tool. Mask identifiers in clinical text on your own machine. - [Pricing](https://precision.example/pricing): indicative USD prices for every solution. Plain-text version: https://precision.example/pricing.md ## Product and method - [How it works](https://precision.example/how-it-works): the five layers and the evidence for each - [Problems we solve](https://precision.example/problems): 8 failure patterns with sources - [Solutions](https://precision.example/solutions): services and products, and the product map ## Product types - [AI scribes](https://precision.example/solutions/ambient-scribes): Core focus. Omissions (18% of reviewed notes at UC Davis), hallucinations (11.5%), interpreter-error propagation (55-60% of errors passed into notes) - [Clinical Q&A](https://precision.example/solutions/clinical-qa): Next. Independent tests found general frontier LLMs beat OpenEvidence and UpToDate Expert AI in physician tests; stale guidance and retrieval misses (unquantified) - [Patient chat](https://precision.example/solutions/patient-chat-triage): Next. Benchmarks mislead: in an Oxford RCT (~1,300 people) users reached under 35% correct conditions and ~44% right action, no better than controls; ECRI ranked chatbot misuse its #1 2026 hazard - [Voice agents](https://precision.example/solutions/voice-agents): Core focus. Workflow pass rates vary widely (Cekura single-scenario Medicare test: 65% to 91%); ASR errors that change clinical meaning; nobody can attribute a failed call to ASR, LLM, TTS or telephony - [Coding and billing](https://precision.example/solutions/coding-prior-auth-revenue-cycle): Core focus. All tested LLMs under 50% exact-match on code assignment in 2024 (GPT-4: 45.9% ICD-9-CM, 33.9% ICD-10-CM, 49.8% CPT); GAO (July 2026) found few independent studies and a ">95% accuracy" claim that rested on one vendor's self-report - [Payer AI](https://precision.example/solutions/payer-ai): Next. Alleged error rates (a plaintiff's ~90% allegation only); state laws require human review of denials - [Imaging](https://precision.example/solutions/imaging-diagnostics): Partner pilots only. 4.8% of 903 authorised devices were recalled, median 458 days after clearance, with weak validation linked to recall; deskilling signal in colonoscopy - [Mental-health bots](https://precision.example/solutions/mental-health-bots): Not yet. Crisis detection and harmful validation; state bans (Illinois, Nevada), Pennsylvania suit over a bot posing as a licensed psychiatrist (allegation) - [EHR agents](https://precision.example/solutions/ehr-agents): Next. Retrieval from FHIR resources and reasoning both limit performance; secondary summary says task completion rose from under 70% (2025) to over 90% (2026) - [Predictive models](https://precision.example/solutions/predictive-models): Not yet. Epic sepsis AUROC ~0.65 in a systematic review; 7.6% PPV in one county-ED validation - [Pharma and trials](https://precision.example/solutions/pharma-clinical-trials): Not yet. Bain: 53% of phase 2/3 trials delayed; ~half of sponsors piloting AI - [Indian languages](https://precision.example/solutions/india-indic-languages): Next. Indic-language and code-switching speech and notes, rural screening generalisation; no verified evidence found ## Parameter library (138 parameters) Each parameter page defines the measure, how it is judged (automatable, clinician, outcomes) and the evidence grade (E1 to E4). Hub: https://precision.example/parameters - Every product type: [Clinical accuracy (rubric)](https://precision.example/parameters/clinical-accuracy-rubric), [Completeness / omission](https://precision.example/parameters/completeness-omission), [Harm severity grade](https://precision.example/parameters/harm-severity-grade), [Groundedness / faithfulness](https://precision.example/parameters/groundedness-faithfulness), [Calibration and appropriate abstention](https://precision.example/parameters/calibration-and-appropriate-abstention) - Ambient scribes and clinical documentation: [Hallucination rate](https://precision.example/parameters/hallucination-rate), [Omission rate](https://precision.example/parameters/omission-rate), [Accidental inclusion](https://precision.example/parameters/accidental-inclusion), [Error severity (major/minor)](https://precision.example/parameters/error-severity-major-minor), [PDQI-9 / modified PDQI-9](https://precision.example/parameters/pdqi-9-modified-pdqi-9) - Clinical question answering and decision support: [Rubric accuracy and completeness (clinician-facing)](https://precision.example/parameters/rubric-accuracy-and-completeness-clinician-facing), [Task-category score](https://precision.example/parameters/task-category-score), [Guideline concordance](https://precision.example/parameters/guideline-concordance), [Citation and source faithfulness](https://precision.example/parameters/citation-and-source-faithfulness), [Drug dose and interaction accuracy](https://precision.example/parameters/drug-dose-and-interaction-accuracy) - Patient-facing chat and triage: [Under-triage (missed emergency) and over-triage](https://precision.example/parameters/under-triage-missed-emergency-and-over-triage), [Emergency recognition and escalation correctness](https://precision.example/parameters/emergency-recognition-and-escalation-correctness), [Real-user task accuracy](https://precision.example/parameters/real-user-task-accuracy), [Simulator-to-real gap](https://precision.example/parameters/simulator-to-real-gap), [Unsafe-advice rate](https://precision.example/parameters/unsafe-advice-rate) - Voice agents: [Workflow pass rate (all repeats pass)](https://precision.example/parameters/workflow-pass-rate-all-repeats-pass), [Interruption recovery and naturalness](https://precision.example/parameters/interruption-recovery-and-naturalness), [ASR error and its clinical impact](https://precision.example/parameters/asr-error-and-its-clinical-impact), [Turn latency, call completion, handoff-to-human rate](https://precision.example/parameters/turn-latency-call-completion-handoff-to-human), [Clinical-safety label error rates](https://precision.example/parameters/clinical-safety-label-error-rates) - Coding, prior authorisation and revenue cycle: [Code-level exact-match accuracy](https://precision.example/parameters/code-level-exact-match-accuracy), [Coding benchmark score](https://precision.example/parameters/coding-benchmark-score), [Claim-level versus code-level accuracy](https://precision.example/parameters/claim-level-versus-code-level-accuracy), [Clean-claim rate and first-pass yield](https://precision.example/parameters/clean-claim-rate-and-first-pass-yield), [First-submission denial rate](https://precision.example/parameters/first-submission-denial-rate) - Payer-side AI: [Human-review compliance](https://precision.example/parameters/human-review-compliance), [Denial error rate](https://precision.example/parameters/denial-error-rate), [Subgroup denial disparity](https://precision.example/parameters/subgroup-denial-disparity), [Denial-reason quality](https://precision.example/parameters/denial-reason-quality), [Coverage-policy consistency](https://precision.example/parameters/coverage-policy-consistency) - Imaging and diagnostics: [Sensitivity, specificity, AUROC](https://precision.example/parameters/sensitivity-specificity-auroc), [PPV and NPV at deployment prevalence](https://precision.example/parameters/ppv-and-npv-at-deployment-prevalence), [Calibration](https://precision.example/parameters/calibration), [Subgroup performance tables and model cards](https://precision.example/parameters/subgroup-performance-tables-and-model-cards), [Out-of-distribution scanner and site failure](https://precision.example/parameters/out-of-distribution-scanner-and-site-failure) - Mental-health bots: [Suicide-risk detection and response](https://precision.example/parameters/suicide-risk-detection-and-response), [Risk exploration](https://precision.example/parameters/risk-exploration), [Validation of delusions and harmful content](https://precision.example/parameters/validation-of-delusions-and-harmful-content), [Crisis-resource provision and handoff](https://precision.example/parameters/crisis-resource-provision-and-handoff), [Therapeutic alliance and modality fidelity](https://precision.example/parameters/therapeutic-alliance-and-modality-fidelity) - Agentic EHR workflows: [Task completion](https://precision.example/parameters/task-completion), [Retrieval versus reasoning failure share](https://precision.example/parameters/retrieval-versus-reasoning-failure-share), [Action safety](https://precision.example/parameters/action-safety), [Order and medication accuracy](https://precision.example/parameters/order-and-medication-accuracy), [Unit and aggregation correctness](https://precision.example/parameters/unit-and-aggregation-correctness) - Predictive models: [Discrimination (AUROC)](https://precision.example/parameters/discrimination-auroc), [Calibration](https://precision.example/parameters/calibration-predictive-models), [PPV and alert burden](https://precision.example/parameters/ppv-and-alert-burden), [Lead time](https://precision.example/parameters/lead-time), [Subgroup performance and drift](https://precision.example/parameters/subgroup-performance-and-drift) - Pharma and clinical trials: [Trial cycle-time and delay](https://precision.example/parameters/trial-cycle-time-and-delay), [Digital-twin prediction accuracy and variance reduction](https://precision.example/parameters/digital-twin-prediction-accuracy-and-variance-reduction), [Model credibility assessment](https://precision.example/parameters/model-credibility-assessment), [Electronic-record audit trail (21 CFR Part 11)](https://precision.example/parameters/electronic-record-audit-trail-21-cfr-part), [Pharmacovigilance case-extraction accuracy](https://precision.example/parameters/pharmacovigilance-case-extraction-accuracy) - India and Indic languages: [Indic-language and code-switching speech and note quality](https://precision.example/parameters/indic-language-and-code-switching-speech-and), [ABDM interoperability conformance](https://precision.example/parameters/abdm-interoperability-conformance), [CDSCO software-as-medical-device classification](https://precision.example/parameters/cdsco-software-as-medical-device-classification), [Public-programme screening performance (for example TB, breast)](https://precision.example/parameters/public-programme-screening-performance-for-example-tb), [BODH benchmark performance](https://precision.example/parameters/bodh-benchmark-performance) ## Agents (18) - [Clinician reaction elicitor](https://precision.example/agents/clinician-reaction-elicitor) (In audits): Turn clinicians' pass/fail and comments on real outputs into a failure taxonomy and candidate criteria - [Gold-set curator and labelling triage](https://precision.example/agents/gold-set-curator-and-labelling-triage) (In audits): Choose which cases clinicians label (uncertain, high-severity, new-specialty); minimise clinician minutes - [Clinical reviewer (rubric grader)](https://precision.example/agents/clinical-reviewer-rubric-grader) (In audits): Grade outputs against clinician-authored rubrics - [Omission auditor](https://precision.example/agents/omission-auditor) (In audits): Enumerate source facts, check each against output, flag missing material - [Hallucination and commission auditor](https://precision.example/agents/hallucination-and-commission-auditor) (In audits): Verify each output claim against source - [Medication, dose, negation and laterality checker](https://precision.example/agents/medication-dose-negation-and-laterality-checker) (In audits): Deterministic and targeted checks for high-harm error classes - [Severity grader](https://precision.example/agents/severity-grader) (In audits): Classify errors by type, potential harm and reached-patient status - [Judge calibration and abstention agent](https://precision.example/agents/judge-calibration-and-abstention-agent) (In audits): Maintain judge error rates, intervals, drift in judge; route uncertain cases to clinicians - [Patient simulator](https://precision.example/agents/patient-simulator) (Building): Generate realistic, wording-varied patient and caregiver conversations - [Clinician and caller simulator for voice](https://precision.example/agents/clinician-and-caller-simulator-for-voice) (Building): Simulate callers with accents, noise, interruptions - [Replay and counterfactual diagnosis agent](https://precision.example/agents/replay-and-counterfactual-diagnosis-agent) (Building): Swap ASR, prompt, retriever, model or guideline version on fixed inputs; attribute failures - [Red-team agent](https://precision.example/agents/red-team-agent) (Building): Probe emergencies, jailbreaks, injection in chart text, scope violations, minors - [Escalation and safety-gate agent](https://precision.example/agents/escalation-and-safety-gate-agent) (Roadmap): Deterministic detection of red-flag and crisis content and forced handoff - [Policy-compiler agent](https://precision.example/agents/policy-compiler-agent) (Roadmap): Compile written rules (scope of practice, payer policy, state disclosure, authority limits) into checkable gates - [Repair agent (evaluator-bounded)](https://precision.example/agents/repair-agent-evaluator-bounded) (Roadmap): Propose prompt, retrieval or routing fixes; accept only if held-out clinician-labelled results hold within judge error bounds - [Drift monitor](https://precision.example/agents/drift-monitor) (In audits): Scheduled regression runs with repeated trials after vendor, model, EHR or ASR updates - [Regulatory evidence-pack agent](https://precision.example/agents/regulatory-evidence-pack-agent) (Building): Assemble ledger entries into Joint Commission, CHAI model-card, NHS DCB0129/0160, AB 3030 and state artefacts - [Bias and equity auditor](https://precision.example/agents/bias-and-equity-auditor) (In audits): Subgroup error and denial analysis; language and dialect parity ## Learn: healthcare AI library (regulation first) Sourced, dated reference pages. Plain Markdown: add .md to any /learn URL. - [EU AI Act dates for medical device AI after the Omnibus](https://precision.example/learn/eu-ai-act-digital-omnibus-medical-devices) (Regulation, EU, reviewed 2026-11-15): The Digital Omnibus on AI, in force 27 July 2026, moves standalone high-risk AI to 2 December 2027 and AI in regulated products such as devices to 2 August 2028. - [What changed in FDA's 2026 clinical decision support guidance?](https://precision.example/learn/fda-clinical-decision-support-guidance-2026) (Regulation, US, reviewed 2026-11-15): FDA's revised CDS guidance (6 January 2026) widens enforcement discretion for single-recommendation tools. What it covers, what it leaves out, and its silence on AI. - [FDA's PCCP guidance for AI devices, and the lifecycle draft](https://precision.example/learn/fda-pccp-ai-enabled-devices) (Regulation, US, reviewed 2026-11-15): FDA's final PCCP guidance for AI-enabled device software was posted on 3 December 2024. The broader AI lifecycle guidance is still a January 2025 draft. - [India's CDSCO software guidance, BODH and DPDP for health AI](https://precision.example/learn/india-cdsco-software-guidance-bodh-dpdp) (Regulation, India, reviewed 2026-11-15): India's July 2026 device software guidance, the BODH benchmarking platform and the DPDP Rules: what each asks of healthcare AI builders, with our source limits. - [Joint Commission and CHAI on AI: what health systems must do](https://precision.example/learn/joint-commission-chai-responsible-use-of-ai) (Regulation, US, reviewed 2026-11-15): Joint Commission and CHAI guidance (17 September 2025) and the voluntary 2026 certification: what they cover, what they do not certify, and CHAI's assurance labs. - [What does ONC's HTI-5 rule remove for AI decision support?](https://precision.example/learn/onc-hti-5-final-rule-decision-support) (Regulation, US, reviewed 2026-11-15): The HTI-5 final rule (28 October 2026) removes source-attribute and risk-management requirements for predictive decision support. Dates, what stays, what buyers ask. - [Which open-source tools and benchmarks evaluate healthcare AI?](https://precision.example/learn/open-source-healthcare-ai-evaluation-tools) (Open source, Global, reviewed 2026-11-15): HealthBench, MedHELM, MedAgentBench and Epic's seismometer compared: what each measures, licence, GitHub traction on 15 November 2026 and where each falls short. - [Section 1557 and AI decision tools: what 92.210 requires](https://precision.example/learn/section-1557-patient-care-decision-support-tools) (Regulation, US, reviewed 2026-11-15): Section 1557's rule makes covered entities identify decision tools that use protected characteristics and mitigate discrimination risk. Dates and limits. - [NHS England rules for AI scribes (ambient voice technology)](https://precision.example/learn/uk-ambient-voice-technology-guidance) (Regulation, UK, reviewed 2026-11-15): NHS England's guidance and supplier registry for ambient voice technology: DTAC, DCB0129 and DCB0160 clinical safety standards, and where MHRA device rules apply. - [US state healthcare AI laws: payer review, chatbots, therapy](https://precision.example/learn/us-state-healthcare-ai-laws-payer-chatbot) (Regulation, US, reviewed 2026-11-15): US states have passed laws on insurer AI in prior authorisation, chatbot disclosure and AI therapy. A dated overview with Colorado's replacement law and our gaps. ## Insights - [Who evaluates healthcare AI in 2026, and what do they miss?](https://precision.example/blog/who-evaluates-healthcare-ai-2026): Qualified Health, Vals AI, Alterra Labs, voice-testing vendors, CHAI and Epic: who evaluates healthcare AI today, what each does, and the gaps that remain. - [Why do AI graders miss omissions in clinical notes?](https://precision.example/blog/why-llm-judges-miss-omissions-in-clinical-notes): Two 2026 preprints find LLM judges near chance at spotting what a clinical note leaves out, yet good at catching invented content. What the numbers say and do not. - [Why can't you compare AI scribe error rates between studies?](https://precision.example/blog/scribe-error-rates-cannot-be-compared): Published scribe error rates run from 1.47% of sentences to 31% of notes. Unit, sampling and review instructions change the answer: what that means for buyers. - [Abridge, GAO and Ontario's auditor on evaluating AI scribes](https://precision.example/blog/what-abridge-gao-and-ontario-say-about-scribe-evaluation): Abridge's published method excludes omissions. GAO found few independent accuracy studies. Ontario's auditor found 11 of 20 scribe vendors had no third-party audit. - [Voice AI testing for healthcare: Coval, Hamming, Cekura, Bluejay](https://precision.example/blog/voice-ai-testing-for-healthcare-compared): Four funded voice-agent testing vendors, what each does, what they charge where public, and why healthcare call failures still lack independent benchmarks. - [Why chatbot benchmark scores miss real-user medical outcomes](https://precision.example/blog/benchmarks-do-not-predict-real-user-outcomes-oxford-trial): A 2026 Oxford randomised trial of 1,298 people found chatbot users no better than controls at naming conditions or the right action. What it means for evaluation. - [Five gaps in healthcare AI evaluation that nobody fills well](https://precision.example/blog/healthcare-ai-evaluation-gaps-nobody-fills): Calibrated graders, causal diagnosis, buyer-run testing, monitoring and language-specific evaluation: the gaps we found in healthcare AI evaluation today. - [What do health-system AI governance committees need from vendors?](https://precision.example/blog/what-health-system-ai-governance-committees-need): Local validation, bias testing, PHI safeguards, monitoring plans and audit trails: what buyers say they require, and how little testing capacity they have. ## Optional - [Topics](https://precision.example/blog/topics): Evaluation science, Landscape and gaps, Building with evidence - [Full text for LLMs](https://precision.example/llms-full.txt) - [RSS](https://precision.example/blog/rss.xml)