A specialty pharmacy gets a letter it has been preparing for: a reviewer from URAC, the accrediting body that launched the first national Health Care AI Accreditation, is coming, and as part of the review she wants to pull a single prior authorization at random and ask the pharmacy to reconstruct exactly what happened. She points at one: a biologic for a patient with rheumatoid arthritis, submitted four months ago, approved on first pass. Her questions are precise and they are the questions the whole program has been building toward. What did the AI produce for this request? What clinical facts did it assert, and from where? Which payer criterion did it cite? Who verified it, what did they check, and what, if anything, did they change before they signed? And can you show me all of that, for this specific request, without reconstructing it from memory? If the pharmacy can answer, calmly and from a record, the accreditation is within reach. If the pharmacy shrugs and says the workflow is fast and the pharmacist is careful, it has failed, because careful-but-undocumented is, to an accreditor, indistinguishable from careless. This lesson is about building the record that lets a pharmacy answer that reviewer with confidence: the governance and audit trail that make an AI-assisted prior authorization not just fast and sound, but provable.

Why the Record Is the Deliverable

It is tempting to think of governance and documentation as overhead, the paperwork tax you pay after the real work of building a fast, accurate prior authorization workflow. That framing is backwards, and getting it right reframes this whole lesson. In an AI-assisted clinical process, the record is not a description of the work; for an accreditor, a payer auditor, or a patient-safety investigation, the record is the work, because it is the only thing they can actually examine. The reviewer cannot watch the pharmacist think; she can only read what the pharmacist documented. So a verification that happened but was not recorded did not happen, as far as anyone outside the pharmacist's own head is concerned. The graduate's credential, established back at the start of the goldmine, has three parts: a turnaround that collapsed, a verification standard that got tighter, and documentation that proves competent, governed use. The first two lessons of this chapter built the speed and the analytics. This lesson builds the third part, the proof, and it is the part that turns a good workflow into a defensible one.

The forcing function is concrete and current. URAC's Health Care AI Accreditation has two tracks, one for the organizations that develop AI tools and a separate one for the organizations that use them, and a pharmacy is squarely on the user track. The user track does not ask whether your vendor's model is well-built; it asks whether your pharmacy uses the tool competently and under governance, and the evidence it accepts is documentation. A pharmacy that cannot produce, for a given prior authorization, what the AI did and what the human verified has no way to demonstrate governed use, no matter how careful its people actually are. The record is therefore not overhead; it is the single deliverable the accreditation evaluates, which is why a workflow designed for accreditation generates the record as a byproduct of doing the work, rather than reconstructing it under audit pressure.

For an accreditor, an auditor, or a safety investigation, the record is the work, because it is the only thing they can examine. A verification that happened but was not recorded did not happen. Careful-but-undocumented is, to a reviewer, indistinguishable from careless.

What a Defensible PA Record Contains

A defensible prior authorization record has a specific job: it must let someone who was not there reconstruct exactly what the AI produced and exactly what the human verified, for this specific request, after the fact. Working backward from that job gives a clear list of what the record must capture, and each element maps to a stage of the workflow built in the first lesson of this chapter. What the AI produced. The record captures the AI's output as it was generated: the extracted clinical facts, the cited payer criterion, and the drafted justification. This matters because the reviewer's first question is what the tool asserted, and you cannot answer it if the output was overwritten without a trace the moment the human edited it.

The source grounding. For the extracted facts, the record shows where each came from in the chart, the source trace, and for the cited criterion, which payer policy and version it was matched against. This is what proves the workflow was grounded, that it used retrieval-augmented generation, abbreviated RAG, over the actual chart and the actual payer rules rather than the model's memory. What the human verified. The record captures the verification itself: which facts were traced and confirmed, that the cited criterion was checked against the actual current policy, and that the clinical assertions were checked against the record. What the human changed. Any override or correction the pharmacist made, the fabricated fact they removed, the mismatched criterion they fixed, is recorded, because the corrections are some of the most important evidence in the file: they prove the verification was real and not a rubber stamp. Who signed, and when. The record attributes the final clinical sign-off to a named, credentialed pharmacist with a timestamp, because accountability that is not attributable to a person is not accountability. A record with all five elements lets the URAC reviewer reconstruct the rheumatoid arthritis prior authorization completely, four months later, without anyone reaching for their memory.

Walk the reviewer through that record and watch the five elements answer her five questions in order. She asks what the AI produced: the record shows the extracted diagnosis of rheumatoid arthritis with its code, the documented inadequate response to a conventional disease-modifying agent, the relevant labs, and the drafted justification, exactly as the tool generated them. She asks where the facts came from: each is tied to the chart note it was pulled from, and the cited step-therapy criterion names the payer policy and the version in force four months ago. She asks who verified and what they checked: the log shows the pharmacist traced the inadequate-response note to the chart, opened the payer policy and confirmed the step-therapy requirement was stated as cited and satisfied, and checked each clinical assertion against the record. She asks what changed: the override log shows the pharmacist corrected one lab date the extraction had transposed before signing. She asks who signed: a named, credentialed pharmacist, with a timestamp, after the gate was passed. Five questions, five answers, none from memory. That is what a defensible record does, and it is the difference between a review that goes well and one that does not.

Capture It as a Byproduct, Not an Afterthought

The single most important design principle for the audit trail is that it must be generated by the workflow as it runs, not assembled afterward, and the reason is both practical and ethical. Practically, a pharmacy processing dozens of prior authorizations a week cannot stop to write a narrative reconstruction of each one; if documentation is a separate manual step, it will be skipped under exactly the queue pressure that the AI workflow created, and the audit trail will be thin precisely where volume was highest. Ethically, a record assembled after the fact, from memory, is a reconstruction, and a reconstruction is exactly the thing an auditor distrusts, because it is written knowing the outcome. The trustworthy record is the contemporaneous one, captured at the moment the work happened, which is only possible if the workflow itself logs each stage as it executes.

In practice this means the workflow is designed so that capturing the trail is not an extra task but a structural consequence of doing the work. When the AI extracts the facts, the system stores that output and its source traces. When the pharmacist runs the verification gate, the act of confirming each item is itself logged, including any correction. When the pharmacist signs off, the system records the named attribution and timestamp and locks the prior version so the AI's original output and the human's changes both survive. The pharmacist does not write a report; the pharmacist does the verification, and the verification produces the record. This is the difference between a workflow that is auditable and one where audit is a painful, error-prone archaeology project run months later. The first lesson's four-stage workflow, assemble, match, verify, submit, was already the right shape for this, because each stage is a natural logging point; governance is largely the discipline of making sure each stage actually writes its log and that nothing can skip it.

Governance Beyond the Single Record

The audit trail answers questions about one prior authorization. Governance is the layer above it that answers questions about the whole program, and an accreditor asks both kinds. Where the record proves a single request was handled correctly, governance proves the pharmacy has a system that makes correct handling the default rather than the lucky outcome. A few elements carry most of the weight. Written policy on AI use in prior authorization: a documented statement of which steps the AI performs, which steps are human-only, that the pharmacist owns the clinical sign-off, and that verification is mandatory before submission. This is the cardinal rule, that AI supports the pharmacist's judgment and never replaces it, written down as policy rather than left as folklore.

Staff competency documentation: evidence that the people using the tool are trained to use it under the verification standard, which is exactly what the URAC user track expects and exactly what a program like this one produces. The metrics review: the cadence from the previous lesson, where turnaround, denial rate, and the verification catch rate are reviewed by someone with authority, which is governance functioning as an early-warning system; a catch rate sliding toward zero is a governance signal that verification discipline is eroding, caught before it becomes an incident. An incident pathway: a defined process for what happens when a fabricated fact does get through, because a governance program that assumes nothing will ever go wrong is not credible to an accreditor who has seen things go wrong. Together these turn a set of individual good prior authorizations into a governed program, and the distinction matters enormously to a reviewer: one good record could be luck, but a policy, trained staff, a review cadence, and an incident pathway are a system, and a system is what gets accredited.

It also matters who owns these elements, because governance that belongs to no one belongs to no one. In a well-run pharmacy the written policy is approved by leadership and revisited when the tool or the payer landscape changes, the competency documentation is maintained as staff are onboarded and as the workflow evolves, the metrics review has a named owner with the authority to act, and the incident pathway names who is notified and who decides when a fabricated fact gets through. None of this requires a large bureaucracy; a small specialty pharmacy can hold all of it in a few living documents and a recurring review. What it requires is intentionality: the decision to treat governed AI use as something the pharmacy designs and maintains, not something it hopes emerges from careful individuals. The accreditor is, in the end, asking whether that intentionality exists, and the documentation is how the pharmacy proves it does. A pharmacy that can show the policy, point to the trained staff, produce the review cadence, and describe the incident pathway has answered the question before it is fully asked.

The Record as Protection, Not Burden

It is worth confronting the natural resistance directly, because a pharmacist drowning in a queue does not want one more thing to document. The reframe that makes the discipline stick is that the audit trail does not exist to satisfy a bureaucrat; it exists to protect the pharmacist, the pharmacy, and the patient, and it does so in concrete ways. Consider what the record does when something goes wrong. A payer audits a batch of prior authorizations and challenges one: the contemporaneous record showing the grounded facts, the verified criterion, and the named sign-off is the pharmacy's defense, the proof that this was a competently handled request and not a careless one. A patient-safety event is investigated: the override log showing the pharmacist caught and corrected three fabricated facts that month is the evidence that the verification gate was real and working, which protects the individual pharmacist personally. The accreditation review arrives: the record is the difference between passing and failing.

The deepest point is that the documentation discipline and the safety discipline are the same discipline, which is the resolution the whole program has been building toward. The act of recording what you verified forces you to actually verify it; a workflow that produces a real audit trail cannot have a rubber-stamp verification gate, because the gate would have nothing true to log. So the record is not a tax on the safe workflow; it is part of what makes the workflow safe, because it makes the verification visible, attributable, and therefore real. This is why the program can promise that the same effort that documents AI use also makes it safer. A pharmacy that builds the audit trail as a byproduct of a grounded, human-verified, signed prior authorization workflow gets all three things the credential promised at once: the speed, from collapsing the clerical assembly; the safety, from the targeted verification gate; and the proof, from the contemporaneous record. When the URAC reviewer pulls that rheumatoid arthritis prior authorization at random, the pharmacy does not scramble. It opens the record, walks her through what the AI produced and what the pharmacist verified and changed and signed, and demonstrates, in one request, the governed competence the whole accreditation is asking for. That is the goldmine fully dug: fast, sound, and provable, with a patient on therapy faster and a record that proves no corner was cut to get them there.

Key Takeaways

• For an accreditor, a payer auditor, or a safety investigation, the record is the work, because it is the only thing they can examine; a verification that happened but was not recorded did not happen, and careful-but-undocumented is indistinguishable from careless.
• URAC's Health Care AI Accreditation puts a pharmacy on the user track, which evaluates whether the pharmacy uses the tool competently and under governance, and the evidence it accepts is documentation, making the record the single deliverable the accreditation judges.
• A defensible prior authorization (PA) record has five elements: what the AI produced, the source grounding (the chart traces and the payer policy version, proving retrieval-augmented generation, or RAG, was used), what the human verified, what the human changed, and who signed with a timestamp.
• The corrections and overrides are among the most important evidence in the file, because they prove the verification was real and not a rubber stamp.
• The audit trail must be captured as a byproduct of the workflow as it runs, not reconstructed afterward, because a contemporaneous record is trustworthy while a from-memory reconstruction is exactly what an auditor distrusts, and manual documentation gets skipped under queue pressure.
• The four-stage workflow (assemble, match, verify, submit) is already the right shape because each stage is a natural logging point; governance is the discipline of making sure each stage writes its log and none can be skipped.
• Governance above the single record includes written AI-use policy, staff competency documentation, the metrics review cadence as an early-warning system, and a defined incident pathway, which together prove a governed program rather than one lucky record.
• The documentation discipline and the safety discipline are the same discipline: a workflow that produces a real audit trail cannot have a rubber-stamp gate, so the record both proves and produces the safety, delivering the credential's three promises at once, speed, safety, and proof.