In 30 seconds
- Every authorization decision draws on six inputs; if one is missing or stale at decision time, automation can't decide safely.
- The usual breaks: eligibility overwritten by retroactive changes, criteria written as prose, documents stored as images, and decisions kept without reasons.
- Federal timeframes and the need to explain every denial make speed and traceability requirements, not nice-to-haves.
- Start with one service category, rebuild 50 past decisions as of their decision date, and fix what blocks it.
Prior authorization is one of the most promising places for AI in a health plan. Most requests are routine, reviewers are stretched, and providers and members feel every day of delay. Automation can approve the clear cases, route the rest to the right reviewer, and draft the rationale for each decision.
But an authorization decision draws on more data than almost any other process in a plan. If one of those inputs is missing, stale or unreadable at the moment of the decision, the automation either can't decide or decides wrongly. Neither is acceptable for a decision about someone's care.
The six inputs every decision depends on
| Input | What goes wrong | What ready looks like |
|---|---|---|
| The request: services and diagnoses | Codes missing, or only in a faxed form | Codes captured as structured fields, whatever the channel |
| Member eligibility and benefits | Today's coverage, not coverage for the date of service | Coverage retrievable as of any date of service |
| Provider and network status | One clinician under several NPIs and TINs; stale network flags | One identity per clinician, with current network status |
| Medical policy criteria | Criteria written as prose, not rules a system can apply | Criteria versioned and expressed as rules |
| Clinical documentation | Scanned images with no extracted text | Text extracted and linked to the request |
| Decision history and rationale | Outcomes kept, reasons not; past decisions overwritten | Each decision stored with its criteria, documents and reason |
The request: services and diagnoses
What goes wrong: Codes missing, or only in a faxed form
What ready looks like: Codes captured as structured fields, whatever the channel
Member eligibility and benefits
What goes wrong: Today's coverage, not coverage for the date of service
What ready looks like: Coverage retrievable as of any date of service
Provider and network status
What goes wrong: One clinician under several NPIs and TINs; stale network flags
What ready looks like: One identity per clinician, with current network status
Medical policy criteria
What goes wrong: Criteria written as prose, not rules a system can apply
What ready looks like: Criteria versioned and expressed as rules
Clinical documentation
What goes wrong: Scanned images with no extracted text
What ready looks like: Text extracted and linked to the request
Decision history and rationale
What goes wrong: Outcomes kept, reasons not; past decisions overwritten
What ready looks like: Each decision stored with its criteria, documents and reason
Why the rationale matters as much as the decision
Authorization decisions get appealed, audited and reviewed by regulators. For many health plans, federal rules now put a clock on every request and require a reason for every denial.
72 hrs
to decide an expedited request
7 days
to decide a standard request
Every
denial needs a specific reason
AI consumability: the inputs have to be readable fast enough to meet the clock. Lineage: every decision has to be traceable to the request, the criteria and the documentation it was based on, as they stood at the time.
An automated approval nobody can explain is a faster way to create an audit finding.
Where the data usually breaks
Eligibility as of the date of service
Retroactive changes overwrite the past, so the system knows who is covered today, not on the day of the procedure. A model trained on today's view learns from information it won't have.
Medical policy as prose
Criteria live in documents written for human reviewers. Until they're expressed as structured rules, automation can only suggest, never decide.
Clinical documents as images
The evidence that a criterion is met is usually in an attachment. If text isn't extracted and linked to the request, a model can't read it at decision time.
Decisions without reasons
Many systems keep the outcome of past reviews but not the criteria applied or the documents relied on. That leaves nothing to train on, and nothing to show an auditor.
How to check your readiness
- 1
Pick one service category.
Start where volume is high and criteria are clear, such as advanced imaging, and list every system its decisions draw on.
- 2
Rebuild 50 past decisions as of their decision date.
For each, try to retrieve the request, eligibility, provider status, criteria version and documents as they stood that day. Every gap you hit is a lineage or consumability finding.
- 3
Time the inputs.
Measure how long each input takes to become available after a request arrives, against the timeframes you have to meet.
- 4
Check that criteria can be applied by a system.
If a reviewer has to interpret the policy text, the first project is turning that category's criteria into rules.
The result is a short, ranked list: what blocks automation for that category, what would reduce its accuracy, and what creates audit risk. Fix the blockers for one category, prove it, then widen.

