A retail pharmacist named Marcus has a patient at the window holding a prescription for a brand-name SGLT2 inhibitor, a newer diabetes medication, and the claim has rejected with a prior-authorization requirement. Marcus opens the AI assistant and asks what the plan's coverage criteria are. In two seconds it answers with confidence: the plan requires a documented trial of metformin and a second oral agent, plus an A1c above a stated threshold, before it will cover this drug. It sounds authoritative, it sounds specific, and it would let Marcus build the justification immediately. There is just one problem, and it is the problem this entire lesson exists to solve: that answer may be the model's best guess at what a plan like this usually requires, assembled from the statistical average of every coverage policy it ever absorbed, rather than what this specific plan actually publishes for this specific drug this year. If Marcus builds a perfect justification against an invented criterion, he gets a fast, clean, confident denial, and the patient waits longer than if he had never used the tool. Matching the request to the real payer rule, not the model's guess, is the second hands-on step of the prior-authorization goldmine, and it rests on one idea: ground the model on the actual criteria, never trust its memory.

Why the Payer Rule Is the Hard Part

A prior authorization (PA) is a payer's demand that you prove the request meets its coverage rules before it pays. The justification you drafted in the previous lesson is only as good as the rule it is matched against, because a flawless clinical narrative aimed at the wrong criterion is a flawless route to denial. And the rule is genuinely hard to pin down. Each payer, and often each plan within a payer, publishes its own coverage criteria, frequently buried in long policy documents, organized differently from every other payer, and revised on their own schedule. The criterion that applied last quarter may have changed. The criterion for this drug may differ from the criterion for the drug next to it on the same formulary tier. Step-therapy requirements, quantity limits, age restrictions, diagnosis requirements, and documentation expectations all live in these documents, and finding the exact one that governs this request is precisely the tedious, error-prone work that historically consumed a large share of the 25 minutes a PA used to take.

This is exactly the kind of search-and-match task AI can accelerate dramatically, and that is the promise. But it is also exactly the kind of task where a language model's deepest weakness shows up, because the model has read enormous quantities of coverage-policy language during training and has formed a powerful statistical sense of what coverage criteria usually look like. Ask it about a plan's rule and, unless you have grounded it on that plan's actual current document, it will generate the most likely-sounding criterion, fluent and specific and plausible, drawn from that average. The danger is not that the model refuses to answer; it is that it answers confidently with a criterion the payer never published. That is the fabricated coverage criterion, one of the named clinical hallucinations, and in the PA workflow it is a direct cause of avoidable denials.

The Model's Memory Is Not the Source of Truth

The single most important idea in this lesson is the distinction between a model answering from its memory and a model answering from a grounded source. When you ask a plain language model a question with no source attached, it answers from its parameters, the patterns baked in during training. Those patterns are a blurry, averaged, and frozen snapshot of an enormous amount of text, with no awareness of which plan you mean, no access to this year's revision, and no ability to tell you it does not actually know this specific plan's rule. It will not say "I am not sure"; it will produce its best statistical guess in the same confident tone it uses for things it has genuinely grounded. For a coverage criterion, that confident guess is worse than useless, because it looks exactly like a real answer and it sends you off to build a justification against a rule that does not exist.

The source of truth for a payer rule is the payer's current published policy, and nothing else. Not the model's memory, not last year's version, not a similar plan's criteria, not what the drug's manufacturer says is typical. The job of grounding is to put that actual document in front of the model and constrain the model to answer only from it. When the model has the real policy text and is told to cite the specific criterion it relied on, two things change: the answer is anchored to the truth, and you get a citation you can verify. When the model does not have the document and is answering from memory, neither is true, and the fluency of the answer is actively misleading because it disguises a guess as a fact.

A language model asked about a payer rule with no source attached does not tell you it is guessing. It produces its best statistical average of what such a rule usually says, in the same confident voice it uses when it is right.

Grounding and Retrieval in Plain Terms

The technique that fixes this has a name worth knowing: retrieval-augmented generation, usually shortened to RAG. The plain-language version is simple. Instead of asking the model to answer from memory, you first retrieve the relevant source document, the actual payer policy for this drug and plan, and you give that text to the model along with your question. The model then generates its answer using the retrieved text as its evidence rather than its training-baked guess. Retrieval finds the truth; generation phrases it. The model becomes a reader and summarizer of a document you supplied rather than an oracle pronouncing from memory, and that is a profound difference in trustworthiness even though the interface looks the same.

For a pharmacist or technician, this can be as formal as a vendor tool that maintains a live, indexed library of current payer policies and retrieves the matching one automatically, or as hands-on as you pasting the actual policy text into the assistant yourself before asking it to identify the governing criterion. Both are grounding; both replace the model's memory with a real source. The formal version scales better and is what the later levels of this program build, but the hands-on version is available to you today and follows the same principle: do not ask the model what the rule is, show the model the rule and ask it to find and apply the relevant part. A grounded tool will also, when properly built, tell you when it cannot find a matching criterion in the supplied source rather than inventing one, which is exactly the honest behavior you want. A tool that always produces a confident criterion no matter what you feed it is a tool that is probably guessing.

How to Tell Grounded from Guessing

You will not always know how a given tool works under the hood, so it helps to have a few practical tells that separate a grounded answer from a guess. The first tell is the citation: a grounded answer points to specific policy language, ideally quoted, with a document name, plan, and date attached, while a guess offers a fluent criterion with nothing behind it. If you cannot click through or read the exact passage the criterion came from, treat the criterion as unverified. The second tell is the willingness to say no: a grounded tool, asked about a rule its source does not contain, will tell you it could not find a matching criterion, while a guessing tool always produces one. The third tell is consistency under rephrasing: ask the same question two different ways, and a grounded answer stays anchored to the same cited passage, while a guess may drift, because it is regenerating an average rather than reading a fixed document. None of these tells is foolproof, but together they let you size up an answer in seconds and decide whether it earns your trust or earns a trip to the actual policy.

A Worked Example: the SGLT2 Inhibitor

Return to Marcus and the diabetes patient and run the grounded version. Instead of asking the assistant what the plan requires, Marcus retrieves the plan's actual current coverage policy for this drug, either through a tool that maintains the policy library or by pulling the document from the payer portal, and supplies that text. He then asks the assistant to identify the specific criterion that governs this request and to quote the exact policy language it relied on. Now the answer is anchored. Suppose the real policy requires a documented trial and inadequate response to metformin alone, with no second-agent requirement, and a different A1c threshold than the model's earlier guess. The grounded answer surfaces that real criterion with the quoted text, and Marcus can see immediately that the patient's documented metformin trial satisfies it. The justification he drafts is now aimed at the rule that actually governs the claim, and the PA has a real chance of clean approval.

Contrast the ungrounded version that opened the lesson. The model's confident two-second answer asserted a second-agent requirement and a higher A1c threshold, neither of which this plan actually imposes for this drug. Had Marcus trusted it, he would have spent time documenting a second-agent trial the patient never needed, possibly delayed the request to obtain it, and built a justification around criteria the payer would not even apply, while overlooking the metformin-only rule the patient already met. The likeliest outcome is a denial, a confused back-and-forth, and a patient whose diabetes medication is delayed by the very assistant meant to speed it up. The fabricated criterion did not announce itself. It arrived as a fluent, specific, confident sentence, indistinguishable in tone from the grounded truth, which is exactly why the discipline is to ground first and never let the model's confidence substitute for a citation to the real document.

The Failure Modes of Rule Matching

Rule matching has its own family of failures, distinct from the drafting failures of the previous lesson, and naming them sharpens the verification. The first and worst is the fabricated criterion: the model asserts a coverage requirement the payer never published, generated from its statistical average. This causes the avoidable denial and is the central reason to ground. The second is the stale criterion: the model, or even a poorly maintained tool, returns a real criterion that was accurate once but has since been revised, so the answer is grounded but out of date. This is why grounding must be on the current policy, not just any policy, and why the date and version of the source document matter as much as its content.

The third failure is the mismatched criterion: the model retrieves a real, current criterion but the wrong one, applying the rule for a different drug, a different indication, a different plan within the payer, or a different patient population, so the criterion is genuine but does not govern this request. This is subtle because the answer is a true criterion, just not the applicable one, and it requires the human to confirm not only that the criterion is real but that it is the right one for this exact drug, plan, indication, and patient. The fourth failure is the silent no-match handled badly: the model is asked about a rule that the supplied source does not actually contain, and instead of saying so, it fills the gap with a plausible invention, the same gap-filling instinct from the drafting lesson, now applied to criteria. The defense against all four is consistent: ground on the current, correct source; demand a citation to the exact policy language; confirm the citation governs this specific request; and treat a confident criterion with no verifiable source as a guess until proven otherwise.

Verification and the Cardinal Rule

Grounding makes a right answer likely; verification makes it certain, and in a clinical and financial document aimed at a payer, likely is not the standard. The verification step here is specific and fast when the grounding produced a citation. The pharmacist or technician opens the payer's actual current policy and confirms three things: that the criterion the tool cited genuinely appears in the policy as quoted, that the version and date of the policy are current, and that this criterion is the one that governs this exact drug, plan, indication, and patient rather than a neighboring rule. When the tool grounded properly and cited the text, this confirmation takes a minute or two and is mostly a matter of reading the quoted passage against the live document. When the tool offered a confident criterion with no citation, the verification is the discovery that there is nothing to verify against, which is itself the answer: do not trust it, go find the real rule.

The cardinal rule of the program governs this step as firmly as any other: AI supports the pharmacist's judgment; it never replaces it. The tool can retrieve and surface the candidate criterion, and that is real, valuable acceleration that returns much of the 25 minutes. But the determination that this is the criterion that governs this request, and that the patient's documented history satisfies it, is a judgment the pharmacist or technician makes and owns, not a verdict the tool delivers. The matched rule is a starting point for human confirmation, never a final answer to rubber-stamp. A PA built on a grounded, cited, human-confirmed criterion is fast and defensible. A PA built on the model's confident memory is fast and fragile, and the patient pays the difference in delay every time the guess turns out to be wrong.

It is worth seeing how this step connects to the ones around it, because the goldmine workflow is a chain and a weak link anywhere breaks the whole thing. The drafting step from the previous lesson produced a justification grounded in the chart; this step ensures that justification is aimed at the criterion that actually governs the claim; the verification step in the next lesson confirms both the clinical facts and the cited criteria before anything is submitted. A justification can be perfectly grounded in the chart and still fail if it is matched to a fabricated rule, and a perfectly matched rule does nothing if the justification asserts a clinical fact the chart does not support. The two grounding disciplines, grounding on the chart and grounding on the payer policy, are separate jobs aimed at separate sources, and a clean PA needs both. That is why this chapter teaches them as distinct steps rather than folding them together: the failure modes differ, the sources differ, and the verification differs, even though the underlying principle, never trust the model's memory over a real source, is the same in both.

Key Takeaways

• Matching the request to the real payer rule is the second hands-on step of the prior-authorization (PA) goldmine, and a flawless justification aimed at the wrong criterion is a flawless route to denial, so the rule the justification targets matters as much as the justification itself.
• Payer coverage criteria are genuinely hard to pin down: each plan publishes its own rules in long documents, revised on its own schedule, and finding the exact governing criterion is the tedious work that consumed much of the historical 25-minute PA.
• A language model asked about a rule with no source attached answers from its training-baked memory, producing its best statistical guess at what such a rule usually says, in the same confident tone it uses when it is right; that fabricated coverage criterion is a direct cause of avoidable denials.
• The source of truth for a payer rule is the payer's current published policy and nothing else, not the model's memory, last year's version, a similar plan's rule, or what is typical; grounding means putting that real document in front of the model and constraining it to answer only from it.
• Retrieval-augmented generation (RAG) is the technique: retrieve the actual policy first, give it to the model as evidence, and have the model read and apply it rather than answer from memory; in practice this can be a vendor tool with a live policy library or you pasting the real policy text yourself.
• The four failure modes are the fabricated criterion (invented requirement), the stale criterion (real but revised), the mismatched criterion (real and current but for the wrong drug, plan, indication, or population), and the silent no-match the model papers over with invention.
• Verification confirms three things against the live policy: that the cited criterion genuinely appears as quoted, that the policy version is current, and that the criterion governs this exact drug, plan, indication, and patient; a confident criterion with no citation is a guess until proven otherwise.
• The cardinal rule holds: the tool retrieves and surfaces the candidate criterion, but the determination that it governs this request and that the patient satisfies it is the pharmacist's judgment to make and own, never a verdict to rubber-stamp.