What problem does it solve?
Health insurers and the administrators that work for them review large volumes of paper heavy cases. An inpatient health claim arrives with a discharge summary, lab reports and bills; at ICICI Lombard that meant reading 20 or more pages per claim. A prior authorization request brings the clinical notes that justify the planned treatment. In both cases an adjudicator or nurse must check the file against the policy terms and clinical criteria. Much of the time goes into reading and retyping, not into the judgment the role exists for.
The stakes are high on both sides. Slow reviews delay care and frustrate providers, while loose reviews let waste, abuse and billing errors through. Automation also carries a clear public risk: investigations and lawsuits in the United States have accused insurers of using algorithms to deny care with little human review. Any AI in this process must make reviews faster and more consistent without taking the clinical decision away from people.
- The US Centers for Medicare & Medicaid Services states that waste contributes to up to 25% of health care spending in the United States, and targets its WISeR prior authorization model at services with a history of waste, fraud and abuse.WISeR (Wasteful and Inappropriate Service Reduction) Model (2025)
How does it work?
- Take in the request or claim. Documents arrive through portals, electronic submissions, email or scans; the AI classifies them and links them to the member and the case.
- Extract and structure. It extracts diagnosis, clinical presentation, history, treatment, procedures, codes and billed amounts, with confidence levels.
- Check against policy and criteria. It compares the case with the member's cover and with the published clinical guidelines the insurer uses, and lists what matches, what is missing and what conflicts.
- Prepare the file. The reviewer receives a summary with the evidence behind each point and a suggested outcome; clean, low risk cases that meet all criteria can be approved automatically.
- Decide and communicate. A clinician or adjudicator makes every denial or reduction, and the AI drafts the plain language decision or appeal letter for approval.
- Audience
- Employee facing
- Autonomy
- Copilot
- Adoption
- Early adopters
- Channels
- Internal tools, API and system to system
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
| KPI | Median | Reported range | Data points | Claimed by |
|---|---|---|---|---|
| Handling time reduction | Too few to pool | 50% to 50% | 2 | 2 vendor |
| Accuracy | Too few to pool | 98% | 1 | 1 vendor |
| Hours saved | Not pooled | 11,000 hours | 1 | 1 vendor |
| Interactions handled | Not pooled | at least 1 million | 1 | 1 vendor |
Value drivers: Speed and cycle time, Employee productivity, Lower cost to serve, Compliance quality, Customer experience.
Indicative value
A health insurer that reviews 400,000 claims and authorization requests by hand each year
USD 1.1 million to USD 6.4 million
Reviewer time released per year
How this is calculated
Formula: cases * minutesSaved / 60 * costPerHour. The low scenario uses every low input, the high scenario every high input.
| Input | Low | High | Basis |
|---|---|---|---|
| Claims and authorization requests reviewed by a person per year cases, cases per year | 400,000 | 400,000 | The reference insurer. |
| Reviewer minutes saved per case minutesSaved, minutes per case | 4 | 12 | Editorial assumption. The evidence on this page reports percentages, not minutes, for full case reviews (Microsoft reports that ICICI Lombard cut the time to process a single health claim by over 50%), and Acentra Health saved three minutes on the letter step alone (from six to three minutes per appeal letter). Replace with a time study. |
| Fully loaded cost per reviewer hour costPerHour, USD per hour | 40 | 80 | Editorial assumption; nurses and clinical reviewers cost more than claims processors. Replace with your own. |
What it leaves out: Review effort only. It leaves out faster decisions for members and providers, the effect of more consistent reviews on waste and appeals, the cost of clinical content and validation, and the cost of the platform.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Centers for Medicare & Medicaid Services
United States · Government and public sector · 2026
In the WISeR model, the US Centers for Medicare & Medicaid Services works with technology companies that use AI and machine learning, together with clinical review, to review prior authorization requests and claims before payment for a selected set of Original Medicare services with a history of waste, fraud and abuse. It runs from January 1, 2026 to December 31, 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington. Every recommendation for non payment is decided by an appropriately licensed clinician. Six technology companies take part, one per state, and are paid a share of the averted spending, adjusted for performance measures that include provider experience. The model page reports no results yet.
No outcome disclosed.
Manulife
Canada · Insurance · 2026
In its first quarter 2026 report to shareholders, Manulife says it enhanced online claims processing for its Affinity health and dental business in Canada with AI driven document processing for the majority of claims that were processed manually, which improved processing speed and paid customers faster. No figures were published.
No outcome disclosed.
Acentra Health
United States · Healthcare · 2024
Acentra Health, which reviews Medicare appeals, built MedScribe on Azure OpenAI Service to turn a physician's clinical rationale into a plain language, empathetic appeal determination letter for the beneficiary and the provider, a task specially trained nurses used to do by hand in an appeals process where a decision can be due within 24 hours. It was tested with 10 nurses who rated every draft, then rolled out to all nurses who write these letters. Microsoft reports that time per letter fell by about 50%, from six to three minutes, saving 11,000 nursing hours and nearly USD 800,000 since deployment, and that nurses gave the generated letters a 99% approval rating.
- Handling time reduction: about 50%, Nurse time per appeal determination letter
"With MedScribe, Acentra Health reduced the time that its specially trained nursing staff spent on each appeal determination letter by approximately 50%."
Claimed by: vendor - Hours saved: 11,000 hours, Nursing hours saved by mid 2024
"By mid-2024, the company had saved 11,000 nursing hours and begun preparations to expand MedScribe to other areas to drive organizational efficiency."
Claimed by: vendor - Cost savings: about USD 800,000, Saved since deployment
"This adds up to 11,000 nursing hours and nearly $800,000 that have been saved since deploying MedScribe."
Claimed by: vendor
ICICI Lombard
India · Insurance · 2024
ICICI Lombard built a claims copilot for its health claim adjudicators with Azure OCR, Azure AI Document Intelligence, Azure OpenAI and an in house model. It structures discharge summaries, lab reports and bills into diagnosis, clinical presentation, history, treatment and investigations, and compares the treatment with National Health Authority and disease treatment guidelines, so the adjudicator reads a summary instead of 20 or more pages. Microsoft reports that the time to process a single health claim fell by over 50%.
- Handling time reduction: at least 50%, Time for an adjudicator to process a single health claim
"This solution has reduced the time for claims adjudicators to process a single health claim by over 50%."
Claimed by: vendor
AdvanceCare
Portugal · Insurance · 2023
AdvanceCare, a health insurance group in Portugal that is part of the Generali Group and acts as a third party administrator for 1.7 million members, has used Sprout.ai's platform since January 2023 to extract, structure and validate data from hospital, dental, pharmacy and other healthcare invoices, map claim descriptions to codes with confidence levels, and automate settlement of routine claims. Sprout.ai reports that some routine claims are now settled in 60 seconds, that over a million claims went through the platform in the past year, that automation levels rose by more than 10%, and that a pilot on a sample of invoices before the partnership was 98% accurate.
- Interactions handled: at least 1 million, Claims processed in the past year
"The milestone comes after the health insurer TPA (Third-Party Administrator) processed over a million claims in the past year using Sprout.ai’s patented AI platform."
Claimed by: vendor - Accuracy: 98%, Pilot on a sample of healthcare invoices across categories, before the partnership
"The results were 98% accurate across all categories."
Claimed by: vendor
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- Policy wordings, benefit rules and the clinical criteria or guidelines the insurer applies
- Historical cases with documents, decisions and appeal outcomes
- Code sets and provider data for billing checks
- Approved letter templates for approvals, denials and appeal outcomes
Systems to integrate
- Claims adjudication and utilization management systems
- Provider portals and electronic prior authorization interfaces
- Document capture and storage for clinical records
- Member and provider communication channels for decisions and letters
Complexity: High
Extraction from medical documents is proven, but the process handles sensitive health data, works under strict decision deadlines and appeal rules, and needs clinical criteria in a form the system can apply. Clinical, legal and compliance teams must own the design, and every adverse decision stays with a licensed reviewer.
- 1
Start with summaries for reviewers
Let the AI structure documents and write a case summary that the reviewer checks. It saves time immediately and changes no decision rights.
- 2
Encode the criteria with clinicians
Work with medical directors to turn the clinical criteria and benefit rules into checks the system can apply, each linked to its source document and version.
- 3
Automate approvals only
If automation goes further, let it approve clean cases that meet every criterion. Denials, reductions and partial approvals always go to a licensed reviewer. In the EU, an automatic approval is still a solely automated decision based on health data: GDPR Article 22(4) allows it only with the member's explicit consent or on grounds of substantial public interest in law, so check that basis before you switch it on.
- 4
Draft decision letters in plain language
Use the AI to turn the reviewer's rationale into a clear letter for the member and provider, approved by the reviewer before it is sent.
- 5
Monitor outcomes by group and by reviewer
Track approval, denial and overturn rates on appeal across member groups, conditions and reviewers, and investigate any pattern the AI may have introduced.
Guardrails
- No denial, reduction or termination of care without a licensed clinician or adjudicator reviewing the case
- Every recommendation cites the clinical criterion or policy clause it relies on
- Automatic decisions limited to approvals of cases that meet every criterion
- Health data processed under HIPAA or GDPR special category rules, masked in logs and prompts
- Overturn rates on appeal monitored for AI supported decisions
KPIs to instrument
- Reviewer minutes per case, before and after
- Time from request to decision, against regulatory deadlines
- Share of cases approved automatically and share sent to review
- Overturn rate on appeal for AI supported decisions
- Agreement between AI summaries and reviewer findings on a weekly sample
Human in the loop
Licensed clinicians and adjudicators make every adverse decision and approve every letter before it is sent. Medical directors own the criteria encoded in the system, and a quality team samples AI supported approvals and summaries every week.
Common failure modes
- Rubber stamp review
- Reviewers approve the AI's suggested denial without reading the file. Measure review time per case, audit samples and never let the system propose a denial without the evidence attached.
- Criteria that drift from medicine
- Encoded criteria fall behind updated guidelines. Give every criterion an owner, a version and a review date.
- Extraction errors in clinical detail
- A missed comorbidity or wrong code changes the outcome. Show confidence per field and require the reviewer to confirm key facts.
- Unequal outcomes for vulnerable members
- The system denies or delays care more often for elderly, disabled or chronically ill members, for example because their files are longer and less standard. Monitor outcomes by group and fix the cause.
What are the risks and rules?
EU AI Act
Depends on design
Annex III point 5(a) makes AI high risk when it is used by or on behalf of public authorities to evaluate eligibility for essential public assistance benefits and services, including healthcare services, or to grant, reduce or revoke them, which can cover statutory health schemes run by or for public bodies. Point 5(c) covers risk assessment and pricing in life and health insurance, not claim review. A copilot for a private insurer's claim review, where people decide, is usually outside Annex III; for public schemes, Article 6(3) may exempt a system that only performs a preparatory task, unless it profiles natural persons. GDPR rules on health data (Article 9) and on solely automated decisions (Article 22) apply in every case.
Guidance
- CMS Interoperability and Prior Authorization Final Rule (CMS-0057-F) (Centers for Medicare & Medicaid Services, North America). Requires impacted payers, such as Medicare Advantage organizations and state Medicaid programs, to send prior authorization decisions within 72 hours for urgent and seven calendar days for standard requests, to give a specific reason for denials from 2026, and to offer a Prior Authorization API.
- WISeR (Wasteful and Inappropriate Service Reduction) Model (Centers for Medicare & Medicaid Services, North America). Shows the US government's design for AI assisted prior authorization, in which every recommendation for non payment is decided by a licensed clinician.
- Annex III, high risk AI systems referred to in Article 6(2) (European Union, Europe). Point 5(a) on public assistance benefits and services, including healthcare services, and point 5(c) on life and health insurance determine when this use is high risk in the EU.
- Opinion on Artificial Intelligence governance and risk management (European Insurance and Occupational Pensions Authority, Europe). Addressed to national supervisors, it clarifies how existing insurance legislation applies to the governance and risk management of AI systems used by insurers, following a risk based and proportionate approach.
Controls to put in place
- Documented decision rights, where AI may recommend and approve within criteria and only people deny or reduce
- Versioned clinical criteria and benefit rules with clinical owners
- Audit trail of documents, extracted facts, criteria applied, reviewer and decision per case
- Outcome and appeal overturn monitoring by member group and condition
- Privacy and security controls for health data, including access logging
When it went wrong elsewhere
- How Cigna Saves Millions by Having Its Doctors Reject Claims Without Reading Them. ProPublica reported that Cigna doctors signed off payment denials flagged by its PXDX review process in batches, spending an average of 1.2 seconds per case according to company documents. Cigna said it was incorrect that the process lets its doctors reject claims without examining them. A warning about human review that becomes a formality.
- UnitedHealth sued over use of algorithm in Medicare Advantage plans. A class action alleged that UnitedHealth and its subsidiary NaviHealth used an algorithm, nH Predict, to deny rehabilitation care to seriously ill Medicare Advantage patients, and claimed a 90% error rate based on the share of denials reversed on appeal. UnitedHealth said the tool is not used to make coverage determinations.
Frequently asked questions
- Can AI deny prior authorization requests or claims?
- It should not. The US government's own WISeR model uses AI to assist prior authorization reviews, but every recommendation for non payment is decided by an appropriately licensed clinician. The safe design lets AI summarise, check criteria and approve clean cases, and leaves every denial or reduction to a person.
- What time savings do health insurers report?
- Microsoft reports that ICICI Lombard's claims copilot cut the time to process a single health claim by over 50%, and that Acentra Health halved nurse time per Medicare appeal letter, saving 11,000 nursing hours. Sprout.ai reports that AdvanceCare settles some routine claims in 60 seconds.
- What went wrong in the publicised cases of algorithmic denials?
- ProPublica reported that Cigna doctors signed off denials in batches, at an average of 1.2 seconds per case, and a class action alleged that UnitedHealth used an algorithm to deny rehabilitation care and that most appealed denials were reversed. Both companies disputed these accounts. The lesson is to measure real human review (time per case, overturn rates on appeal) and to keep denials out of automation.
How to cite this page
Blits.ai AI Use Case Library, "AI for health insurance prior authorization and claims adjudication support", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/health-prior-authorization-and-claims-adjudication. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 27 September 2026: First published