AI scribes in 2026: Abridge, Ambience, Nabla, Microsoft DAX, and the custom-agent alternative

Most healthcare AI experiments stall in pilot. AI medical scribes do not. They are the rare healthcare-AI category running at clinical scale today, with four serious players and two clear white spaces. This is the operator read from a team that builds on top of them: where each scribe actually fits, what they do not do, and when a custom AI agent is the better answer.

Why AI scribes work when most healthcare AI does not

The pattern is unusual for healthcare AI: a generative-AI feature that clinicians keep using after the pilot. The reason is mundane. Documentation is the part of the clinician’s day that everyone (clinician, payer, patient, hospital CFO) already agrees is broken. A scribe that drafts the note while the clinician focuses on the patient does not need to convince anyone of the use case. It just needs to draft a note the clinician will accept and edit.

KLAS Research’s 2024 report on ambient documentation found that 80% of clinicians using these tools reported reduced burnout within the first 90 days. That is not a number any other healthcare AI category can produce. HIMSS 2024 survey data on ambient AI adoption tracks the same direction: ambient scribes were the single fastest-growing clinical-AI category among health systems polled.

The market sorted itself into four serious players by mid-2025. Each one owns a different segment.

Vendor comparison: Abridge vs Ambience vs Nabla vs Microsoft DAX vs custom agent

Direct answer for readers comparing options: Abridge fits multi-hospital health systems on Epic Hyperspace. Ambience fits mid-market hospitals with broad specialty coverage. Nabla fits telehealth and EU primary care. Microsoft DAX Copilot fits Epic-on-Microsoft enterprises that want one procurement contract. A custom AI agent fits specialty workflows, patient-side capture, and downstream automations the scribes do not cover.

OptionBest fitEHR integrationCompliance postureStatus check (2025)AbridgeLarge health systems (Cleveland Clinic, Emory, Christus, UPMC, Kaiser Permanente). Multi-specialty, contracted at the system level.Deep Epic Hyperspace integration. Smart Phrases, in-EHR sidebar.HIPAA BAA, SOC 2 Type II, HITRUST.Series E, $250M raised February 2025 at $2.75B valuation. The category leader in the enterprise tier.Ambience HealthcareMid-market hospitals and multi-specialty ambulatory groups. Specialty-template coverage across 38+ specialties.Epic Hyperspace and athenaOne. Inline notes, structured ICD-10 and CPT suggestions.HIPAA BAA, SOC 2 Type II.Y Combinator W20 alumni, $70M Series B in 2024. The serious challenger to Abridge.NablaTelehealth, primary care, behavioural health. Faster to deploy, EU-friendly given the team is Paris-based.Lighter EHR footprint. eClinicalWorks, NextGen, Doctolib, and a documented API for custom EHRs.HIPAA BAA, GDPR-aligned, HDS hosting available for EU customers.$24M Series B late 2023. The pragmatic choice when the integration scope is the constraint.Microsoft DAX Copilot (Nuance)Health systems already on the Microsoft stack or deep Epic shops. One procurement contract for the whole stack.Epic Hyperspace and Oracle Cerner PowerChart. Embedded in Microsoft Teams and Office.HIPAA BAA, HITRUST, ISO/IEC 27001:2022, C5 in Germany.Owned by Microsoft since the Nuance acquisition closed March 2022 ($19.7B). The default for the enterprise IT department that wants one vendor.Custom AI agent (Life Value pattern)Single-specialty depth (oncology, behavioural health, paediatrics), patient-side capture, payer-side intake, inpatient rounding, downstream automations on the scribe’s structured output.FHIR R4 native. SMART on FHIR launch. OAuth 2.0 and OIDC. Direct write-back to Epic, Oracle Cerner, MEDITECH Expanse, athenaOne, NextGen, eClinicalWorks via Redox or direct APIs.HIPAA BAA, GDPR, HL7 FHIR R4, ISO 13485, ISO/IEC 27001:2022, MDR when classed as software-as-medical-device. On-prem and on-device deployment options.Built per engagement on AWS Bedrock (HIPAA-eligible since 2024) or Azure OpenAI with a customer-owned BAA. Procurement-ready and audit-ready.

How to read the table

• If the scope is general ambulatory documentation across many specialties at a multi-hospital system, the four named scribes are the procurement decision.
• If the scope is one specialty, one workflow (prior authorization, claims, eligibility, triage, intake, follow-up), or the patient side of the encounter, a custom AI agent on the same FHIR-native plumbing usually ships faster and stays cheaper than buying a general scribe and bolting custom logic on top.

The four named scribes, in plain words

Suki, Augmedix, and DeepScribe round out the field. Suki ships voice-first and is the right pick for specialty practices that already use voice commands. Augmedix was acquired by Commure in 2024 and now lives inside their broader clinical-AI platform. DeepScribe occupies the smaller primary-care end of the market.

“AI scribes are the only healthcare-AI category right now where the question is which one rather than whether. That tells you something about the maturity of the underlying problem.”

Alex Szilagyi, CEO, Life Value

What every general AI scribe still does not do

• Specialty depth. Oncology, behavioural health, paediatric subspecialties, and complex inpatient documentation are partially covered, not great. The templates are general; the language is specialised.
• Inpatient rounding. The scribe was designed for the ambulatory visit. Rounding, ICU sign-out, and inpatient continuity are different problems.
• Patient-side capture. The clinician records the visit and gets the note. The patient gets no equivalent. No recording, no AI-drafted summary in plain language, no continuity record they can carry to the next provider.
• Cross-encounter intelligence. The note this scribe writes does not learn from the note the same patient’s prior scribe wrote at a different system. The records remain siloed. Brendan Keeler’s PHR-tarpit critique applies. Patient-mediated exchange under TEFCA is the architecturally honest answer, not vendor-locked patient portals.

Two white spaces still open in 2026

The first is a patient-side AI scribe. The HTN community has been asking for it since at least February 2026: who is building the AI scribe for patients? The use case is direct. The patient records the visit with the doctor’s consent and a visible LED. The AI drafts a summary in plain language, surfaces what to remember, queues follow-up questions, and stores the artefact on the patient’s device. The data plumbing already exists. FHIR R4, the Cures Act patient-access rule, TEFCA, Apple Health, Google Health Connect, Apple Wallet, Google Wallet. The product does not. This is the HealthWallet.me lane.

The second is specialty-specific scribes that go deeper than the general players. An oncology scribe that understands NCCN guidelines, parses tumour-board notation, and pre-fills the molecular-tumour-board template. A behavioural-health scribe that respects the 42 CFR Part 2 substance-use disorder restrictions and produces SOAP notes that hold up to insurance review. These are narrow markets, but each one has 50,000 to 200,000 clinicians and clear willingness to pay.

The custom-agent alternative

For a hospital, ISV, or payer with a scoped workflow, a custom AI agent on FHIR-native infrastructure is the better answer than a general scribe contract. The pattern Life Value runs:

• Inference layer. AWS Bedrock (HIPAA-eligible since 2024) or Azure OpenAI under the customer’s own BAA. Model choice per workflow, not per vendor lock.
• Data layer. FHIR R4 resources read and written through SMART on FHIR with OAuth 2.0 and OIDC. Integration with Epic Hyperspace, Oracle Cerner PowerChart, MEDITECH Expanse, athenaOne, NextGen, eClinicalWorks via Redox, Particle Health, Health Gorilla, or direct APIs.
• Compliance layer. HIPAA BAA, GDPR, HL7 FHIR R4, ISO 13485 when the agent is a regulated medical device, ISO/IEC 27001:2022, SOC 2 Type II, HITRUST, and C5 or HDS when the deployment is in Germany or France.
• Identity layer. SAML or OIDC single sign-on for clinicians, biometric unlock for patients on HealthWallet.me, full audit log per FHIR resource read or write.
• Orchestration. This is the orchestration gap Jan-Felix Schneider names. Hospitals are not buyers of microservices. Life Value runs the orchestration layer that ties Redox, Particle Health, Health Gorilla, Candid, MedPlum, Aptible, Verifiable, and Eligible together so the customer gets a single contract for a single working workflow.

This pattern fits prior authorization, claims, revenue cycle, eligibility, triage, intake, follow-up, and downstream consumers of a scribe’s structured output. It does not fit the general ambulatory-scribe procurement decision, where Abridge, Ambience, Nabla, and Microsoft DAX have already won the right to that line item.

How to evaluate before signing (any of the five)

• BAA and HIPAA-eligible inference endpoint. Confirm. Ask for the deployment architecture diagram and the model-vendor BAA chain.
• EHR integration. Specifically the EHR the buyer actually runs. Not compatible with Epic. The named module (Epic Hyperdrive, Athena Marketplace) with a reference customer running it in production.
• Specialty template coverage. Ask for the template list, not a marketing slide.
• Voice-to-note accuracy. Run a 10-minute pilot on a clinician who routinely sees the messy cases. Note error rate, time saved per visit, and the clinician’s read on the tone of the output.
• Audit and retention policy. Where do the recordings live, for how long, who can access them, is the LLM provider’s BAA visible.
• Pricing model. Per-clinician monthly subscriptions are the norm. Watch for hidden charges on specialty templates, additional integrations, or transcription minutes.
• CMS-0057-F readiness. The prior-authorization API rule was effective January 2026. Confirm the vendor’s plan for the downstream interoperability obligations.

Where Life Value sits

Life Value does not build a general-purpose AI scribe. The four players above are well past the stage where a new entrant adds value in that market. Life Value builds the products that sit alongside or downstream from the scribe. Patient-side capture for HealthWallet.me, the open-source patient-held EHR that is FHIR-native, encrypted on device, biometric-unlocked, and TEFCA-certified via the FastenHealth on-prem partnership. Specialty scribes for clients with deep clinical-vocabulary needs. AI agents that consume the scribe’s structured output for billing, prior auth, claims, eligibility, or care coordination.

Credentials: HIPAA BAA, GDPR, HL7 FHIR R4, ISO 13485, ISO/IEC 27001:2022, SOC 2 Type II, HITRUST, Cyber Essentials Plus. The team behind FastenHealth on-prem and HealthWallet.me.

FAQ

Which AI scribe is the best for a large health system in 2026?

For a multi-hospital system on Epic Hyperspace, Abridge is the category leader after the February 2025 Series E at $2.75B valuation, with named references at Cleveland Clinic, Emory, Christus, UPMC, and Kaiser Permanente. Microsoft DAX Copilot is the alternative when the IT department wants a single procurement contract under the existing Microsoft enterprise agreement.

What is the difference between Abridge and Ambience?

Abridge contracts at the health-system level and is concentrated in the largest Epic-based enterprises. Ambience covers more specialty templates (38+) and ships into mid-market hospitals and multi-specialty ambulatory groups on both Epic and athenaOne. Ambience raised a $70M Series B in 2024 against Abridge’s $250M Series E in February 2025.

Is Nabla a good fit for telehealth and EU primary care?

Yes. Nabla is the pragmatic choice when the EHR integration scope is the constraint. Lighter footprint, faster deployment, GDPR-aligned, with HDS hosting available for French customers. Paris-based team, $24M Series B late 2023. Strong in telehealth, primary care, and behavioural health.

Does Microsoft DAX Copilot work outside Epic?

Yes. Microsoft DAX Copilot integrates with Oracle Cerner PowerChart in addition to Epic Hyperspace, and lives inside Microsoft Teams and Office. The strongest commercial argument is the single procurement contract across the Microsoft stack for buyers already on Azure and Microsoft 365 E5.

Where is the white space in AI scribes in 2026?

Two spaces remain. First, a patient-side AI scribe that records the visit on the patient’s device with the clinician’s consent and drafts a plain-language summary, stored on the patient’s phone via Apple Health or Google Health Connect. Second, specialty-specific scribes that go deeper than the general players, particularly oncology (NCCN guidelines, molecular tumour boards) and behavioural health (42 CFR Part 2-compliant SOAP notes).

When is a custom AI agent the better choice than a general AI scribe?

When the scope is a single specialty, a single regulated workflow (prior authorization under CMS-0057-F, claims, eligibility, intake, triage, follow-up), or the patient side of the encounter. A custom agent built on FHIR R4 with SMART on FHIR launch, the customer’s own HIPAA BAA on AWS Bedrock or Azure OpenAI, and an audit log per FHIR resource ships faster and stays cheaper than buying a general scribe and bolting custom logic on top.

Sources

KLAS Research, Ambient Speech Tech 2024. HIMSS 2024 clinical-AI adoption survey. Abridge Series E announcement, February 2025. Ambience Healthcare Series B announcement, 2024. Nabla Series B announcement, 2023. Microsoft Nuance acquisition closing, March 2022 ($19.7B). HHS OCR HIPAA guidance. ONC Cures Act patient-access rule. CMS-0057-F prior authorization API rule, effective January 2026.