Industries
AI for healthcare
Administrative AI for providers: prior authorisation packets, denial appeals and patient access, built under a BAA with staff approving every submission. Clinical decisions stay with clinicians.
The sector
What's different here
Clinicians and revenue-cycle staff spend a large part of the day on documentation: notes, in-basket replies, prior authorisation requests, appeal letters. These are drafting and evidence-gathering tasks, and the evidence already sits in the EHR.
The evidence on returns is mixed. AI scribes appear to reduce burnout, but the Peterson Health Technology Institute found no proven financial return yet, and a study of AI-drafted patient replies found no change in reply time. We set targets from measured baselines, not vendor claims.
Anything that touches PHI needs a BAA and minimum-necessary access, and software that influences clinical decisions can fall under FDA or EU device rules. For a mid-size provider we recommend buying an EHR-integrated scribe and building custom only where the workflow is specific to you, which is usually in administration.
66%
of US physicians reported using AI in 2024, up from 38% in 2023. AMA survey of 1,183 physicians, fielded Nov 2024.[1]
15,791 hours
of documentation time saved, by estimate, across more than 2.5 million uses of ambient AI scribes by Kaiser Permanente (TPMG) physicians in the first year. About 1,794 eight-hour workdays.[2]
40%
reduction in reported burnout in a six-week AI scribe pilot at Mass General Brigham (63% at MultiCare). The Peterson Health Technology Institute found scribes have not yet proven a financial return.[3]
No change
in reply time when physicians at UC San Diego Health used AI-drafted patient-message replies. Read times and reply length went up. Study of 122 physicians.[4]
Workflows
Where AI does useful work
Each of these runs inside systems the sector already uses. Steps that carry real risk wait for a person.
Prior authorisation packets
An agent gathers chart evidence against the payer's criteria and drafts the submission for staff to check.
- 1A prior authorisation request is created in the EHR or practice management system
- 2The agent retrieves the payer's criteria for the service
- 3It pulls matching evidence from the chart: diagnoses, prior treatments, results
- 4It drafts the submission with each criterion linked to its evidence
- 5Staff review, complete and submit
Authorisation staff review and submit every request. The agent never submits on its own.
EHR (Epic, Oracle Health) · Practice management and RCM · Payer portals · FHIR APIs
Denial appeal letters
Denied claims are matched to the reason code and the chart, and an appeal letter is drafted for sign-off.
- 1Pull the denial and reason code from the remittance
- 2Retrieve the relevant chart documentation
- 3Draft an appeal letter that addresses the stated reason
- 4A biller or coder reviews and signs off
- 5Track the outcome to improve later drafts
A biller or coder signs off every appeal before it is sent.
RCM platform · Clearinghouse · EHR · Document management
Ambient documentation
Visit audio becomes a structured draft note that the clinician edits and signs. We recommend buying this from an EHR-integrated vendor.
- 1Record the visit with the patient's consent
- 2Transcribe and draft the note in SOAP format
- 3Suggest orders and codes for review
- 4The clinician edits and signs in the EHR
The clinician reviews, edits and signs every note.
EHR (Epic, MEDITECH, athenahealth) · Scribe vendor · Mobile or room device
Patient message drafting
Portal messages are classified and a reply is drafted from the chart for a clinician or nurse to edit.
- 1Classify the message as clinical, administrative or urgent
- 2Route urgent messages straight to the care team
- 3Draft a reply grounded in the chart
- 4A clinician or nurse edits and sends
A clinician or nurse sends every reply.
Patient portal (MyChart) · EHR in-basket · Nurse triage protocols
Patient access and scheduling
Non-clinical questions, intake forms and scheduling handled by an agent, with clinical questions routed to staff.
- 1Answer non-clinical questions from approved content
- 2Book, move or cancel appointments through the scheduling system
- 3Collect intake forms before the visit
- 4Route anything clinical to staff
Clinical questions always go to staff, and the agent gives no medical advice.
Scheduling system · Patient portal · EHR · Contact-centre platform
Systems and rules
What the design has to respect
The systems we expect to connect to, and the regulation that decides what an AI system may do with the data in them.
- EHR (Epic, Oracle Health/Cerner, MEDITECH, athenahealth)
- PACS/RIS
- LIS
- Practice management and RCM
- Clearinghouses
- Patient portal (MyChart)
- Scheduling
- HL7v2 and FHIR interfaces
- Payer portals
- HIPAA Privacy and Security Rules
- Any AI vendor handling PHI needs a Business Associate Agreement, and the minimum-necessary standard limits what each workflow may read.
- Proposed HIPAA Security Rule update
- Proposed on 6 Jan 2025, with mandatory MFA and encryption among other changes. It was still not final as of mid-2026. We design to the proposed controls now.
- FDA device and CDS rules
- Software that meets the definition of a medical device, or clinical decision support outside the non-device criteria, is regulated. The administrative workflows we recommend first are designed to stay outside both.
- EU MDR and EU AI Act
- AI that is a safety component of a medical device is high-risk under Annex I. Under the Digital Omnibus those obligations are scheduled to apply from 2 Aug 2028.
- GDPR and 42 CFR Part 2
- Health data is special-category data under GDPR, and US substance-use records carry extra restrictions under Part 2.
Risks
What goes wrong, and how we design for it
- Errors in drafts that get signed anyway
- Every claim in a draft links to its source in the chart, and we sample signed outputs to measure how often errors get through.
- Automation bias
- Drafts are labelled as drafts, low-confidence items are highlighted, and reviewers see the evidence next to the text.
- PHI sent to tools without a BAA
- Models run in your cloud tenant or with a provider under a BAA, and no PHI is logged outside that boundary.
- Algorithmic denial of care
- We don't build tools that deny or limit care. Pending US lawsuits over insurers' use of algorithms in post-acute care decisions show the legal and reputational exposure.
- Savings that don't appear
- Baseline turnaround and first-pass approval before launch, then compare. Published scribe and in-basket results show gains are not guaranteed.
Where to start
Prior authorisation or denial appeal drafting
Both run under a BAA, keep staff approval before anything is submitted, and have clear metrics. They avoid clinical decision support, and they don't compete with the scribe products already integrated into the major EHRs.
- Pick one service line or one payer with high prior authorisation volume
- Connect read access to the EHR through FHIR or the vendor's API, under a BAA
- Draft submissions with each payer criterion linked to chart evidence
- Staff review, edit and submit every packet from their existing work queue
- Measure turnaround time, first-pass approval rate and staff edits per packet against a pre-launch baseline
In the sector
Publicly reported examples
What others in the industry have announced. These are not our clients.
Kaiser Permanente ambient scribes[5]
TPMG physicians used ambient AI scribes more than 2.5 million times in the first year, with an estimated 15,791 hours of documentation saved.
Scribe pilots at Mass General Brigham and MultiCare[6]
Reported burnout fell 40% in a six-week pilot at Mass General Brigham and 63% at MultiCare. The Peterson Health Technology Institute concluded scribes have not yet shown a financial return.
Use cases
Related use cases
- Document processingInvoices, forms and contracts are read, checked against your own records and posted to the right system, with low-confidence fields and failed checks sent to a person.
- Back-office workflowsReconciliations, approvals and onboarding are prepared by agents with the evidence attached, then routed to the right approver. Nothing is posted or paid without a person signing off.
- Company knowledge searchStaff ask a question in plain language and get an answer drawn from your own documents, with citations, limited to what each person is already allowed to see.
- Customer support automationAn agent answers routine questions from your own policies and account data, takes a small set of allowed actions, and hands everything else to a person with the case already summarised.
Sources
- 1.AMA, Physician sentiment report on AI (Feb 2025)
- 2.AMA, AI scribes save 15,000 hours and restore the human side of medicine (2025)
- 3.Axios, AI scribes reduce burnout but not yet costs (Mar 2025)
- 4.JAMA Network Open, AI-generated draft replies integrated into health records and physicians' electronic communication (Apr 2024)
- 5.Kaiser Permanente Division of Research, AI-assisted notetaking gains steady support from Kaiser Permanente physicians
- 6.Healthcare Dive, AI scribes, provider burnout and financial impact: Peterson Health Technology Institute (Mar 2025)
Figures are as reported by each source on its publication date. Examples are public industry cases, not AI Sewa clients.
Next step