A specialty technician named Dana has a patient waiting on a biologic for moderate-to-severe plaque psoriasis, a therapy that runs several thousand dollars a month and will be denied without a clean prior authorization. The old way, Dana would open the chart, scroll through two years of dermatology notes, hunt for the documented topical-steroid trial and the phototherapy course, copy the diagnosis code, retype the relevant labs, and slowly hand-build a clinical justification in the payer's expected format. Twenty-five minutes, give or take, most of it clerical, all of it standing between a real person and the medication their dermatologist already prescribed. Now Dana pastes the chart context into an AI-assisted drafting tool, and in under a minute a clean, persuasive, well-structured justification appears on the screen, citing the diagnosis, the prior therapies, the inadequate response, and the supporting labs. It looks finished. It looks right. And that is exactly the moment this lesson is about, because the draft is only trustworthy if every clinical fact inside it actually came from Dana's chart and not from the model's instinct for what a convincing justification usually says. This lesson teaches the first hands-on step of the goldmine workflow: using AI to draft the clinical justification, grounded in the chart, with the discipline that keeps a fast draft from becoming a fabricated one.

What Justification Drafting Actually Is

A prior authorization (PA) is a payer's requirement that a prescriber and pharmacy prove, in writing, that a requested medication meets the payer's coverage rules before the payer will pay for it. The heart of that proof is the clinical justification: the short, structured narrative that states the patient's diagnosis, the therapies already tried and their outcomes, the relevant clinical findings, and the clinical reasoning that ties the request to the payer's criteria. Historically this narrative was assembled by hand, which is why a single PA could eat up to roughly 25 minutes. AI-assisted drafting collapses that assembly because writing a structured narrative from supplied facts is exactly the kind of generation task a large language model does fast and well. The tool reads the chart context you give it, identifies the load-bearing facts, and arranges them into the payer's expected shape in seconds rather than minutes.

It is worth being precise about which job the AI is doing here, because the program has separated three different AI jobs and the risks differ for each. Extraction is pulling facts out of a source; generation is producing new text; decision support is surfacing a clinical signal for a human to weigh. Justification drafting is primarily a generation job sitting on top of an extraction job. The tool extracts the diagnosis, the therapy history, and the labs from the chart, then generates a narrative around them. Naming this matters because it tells you exactly where the danger lives. A generation model is optimized to produce fluent, persuasive, well-formed text, and a clinical justification that asserts a perfectly qualifying therapy history is more fluent and more persuasive than one that honestly notes a gap. The model has every incentive to write the convincing version and no built-in obligation to write the true one. That gap between fluent and true is the entire subject of grounded drafting.

Grounded Extraction, the Foundation

Grounding means tying every fact the model asserts to a specific place in a source you trust. A grounded justification is one where the diagnosis came from the chart's problem list, the failed therapy came from a documented prescription and a documented outcome note, and the lab value came from an actual result, each traceable back to where it lives in the record. An ungrounded justification is one where some of those facts came from the model's statistical sense of what a psoriasis PA usually contains. The two can look identical on the screen. They read the same, they are formatted the same, they are equally fluent. The only difference is whether the facts are real, and that difference is the whole game in prior authorization, because a justification is a formal document submitted to a payer under a pharmacist's professional credential on behalf of a patient.

The foundation of safe drafting, then, is grounded extraction: giving the model the actual chart context and constraining it to draft only from what you supplied. This is not automatic. If you hand the model a thin slice of the chart and ask it to write a complete justification, it will happily fill the gaps with plausible invention, because a generation model abhors a blank and will reach for the most likely filler. The fabricated failed therapy, the qualifying diagnosis the patient does not carry, the conveniently perfect lab trend, these are not random errors. They are the model doing precisely what it was trained to do, producing the text that best fits the pattern of a successful PA, whether or not that text is true for this patient. Grounded extraction is how you change the model's job from "write a convincing justification" to "write a justification using only these supplied, verified facts, and say so when a fact is missing."

A grounded justification and a fabricated one can read identically on the screen. The only difference is whether the facts are real, and that is the entire job in prior authorization.

How to Prompt for a Grounded Draft

The difference between a draft you can trust and one you have to distrust often comes down to how you set up the request. A weak prompt invites fabrication; a disciplined prompt fences it out. The first principle is to supply the source. Paste in the actual relevant chart context, the diagnosis and code, the documented therapy history with dates and outcomes, the pertinent labs, and the prescriber's note, rather than asking the model to write a justification for a drug from its general knowledge. The model cannot ground on a source you did not give it, and a justification written from the drug name alone is a justification written from the model's memory of similar cases, which is fabrication by another name.

The second principle is to constrain the model explicitly. Tell it, in plain terms, to use only the supplied facts, to never assert a clinical fact that is not in the provided context, and to flag any place where the payer's typical criteria would expect a fact that the chart does not contain. A good instruction reads something like: draft the clinical justification using only the clinical facts in the context below; do not add any diagnosis, therapy, or lab value that is not present; where a relevant fact appears to be missing, insert a clearly marked placeholder rather than inventing a value. This last instruction is the quiet hero of grounded drafting. A model told to flag gaps instead of filling them turns its greatest weakness, the urge to complete the pattern, into a useful signal that tells you exactly where the chart is thin before you submit. The third principle is to ask for traceability: request that the draft note, for each clinical assertion, where in the supplied context it came from, so your later verification has a map instead of a wall of prose.

What a Grounded Prompt Looks Like in Practice

The shape of a disciplined prompt is concrete enough to memorize. It opens with the role and the constraint, then supplies the source, then states the gap rule. In plain language it runs: you are drafting a clinical justification for a prior authorization; use only the clinical facts provided below; do not introduce any diagnosis, prior therapy, lab value, or date that is not present in the provided context; where a fact the payer would typically expect is missing, write a clearly marked placeholder such as MISSING: prior systemic therapy not documented rather than inventing a value; for each clinical assertion, append a short tag indicating which note or result it came from. Then the chart context follows. A prompt built this way does three jobs at once: it tells the model what to write, it forbids the model from reaching beyond the source, and it converts the model's gap-filling instinct into an explicit flag you can act on. The difference between this prompt and a casual "write me a PA justification for this drug" is the difference between a tool you can verify in two minutes and a tool whose output you have to distrust from the first word.

A Worked Example: the Psoriasis Biologic

Return to Dana and the psoriasis patient and walk the grounded version end to end. Dana assembles the chart context first: the plaque psoriasis diagnosis with its code, a documented twelve-week trial of a high-potency topical corticosteroid with a note recording an inadequate response, a documented course of narrowband phototherapy with a note recording persistent disease, a recent body-surface-area assessment, and the dermatologist's note requesting the biologic. Dana pastes this into the tool with a grounded prompt: use only these facts, flag any gap, note the source of each assertion. The tool returns a justification that states the diagnosis, recounts the topical and phototherapy trials with their inadequate outcomes, cites the body-surface-area figure, and frames the request as a documented step through the conventional therapies the payer expects before a biologic. Each sentence carries a small source tag pointing back to the note it came from. The draft took under a minute and reads cleanly. This is the workflow doing what it is supposed to do.

Now contrast the ungrounded version, the one that should haunt every PA drafter. Suppose Dana, in a hurry, had simply typed "write a prior authorization justification for this biologic for plaque psoriasis" with no chart context. The model, asked to be persuasive with nothing to ground on, produces a beautiful justification: it asserts a documented topical trial, a phototherapy course, a methotrexate failure, and a qualifying body-surface-area score, all of them fluent, all of them formatted perfectly, and at least one of them, the methotrexate failure, entirely invented because this patient never took methotrexate. That invented systemic-therapy failure is not a harmless flourish. It is a misrepresentation submitted to a payer, and when the payer's reviewer checks the record and finds no methotrexate, the result is a denial, a delay for a patient whose disease is active, and a credibility problem for every future submission from that pharmacy. The two drafts looked equally finished. One was grounded and true; the other was fluent and false. Nothing on the screen told them apart, which is precisely why the grounding has to happen at the input, not be hoped for at the output.

The Failure Modes to Design Against

Grounded drafting is a defense against a small set of specific, nameable failures, and knowing them by name makes them easier to catch. The headline failure is the fabricated clinical fact: a failed therapy that did not happen, a qualifying diagnosis the patient does not carry, a comorbidity invented to strengthen the case. This is the generation risk in its purest form, and it is dangerous precisely because the fabrication makes the justification stronger, so the model is most tempted to invent exactly the facts that matter most. The second failure is the subtle extraction error feeding the draft: a real lab value misread, a therapy date shifted, a dose transcribed wrong, so that the narrative is built on a fact that exists but is slightly off. This is quieter than outright invention and easier to miss because the fact is genuinely in the chart, just not as stated.

The third failure is the confident gap-fill, where the chart is missing a fact the payer will want and the model, rather than flagging the absence, smooths it over with a plausible-sounding assertion that papers over the hole. This is the failure the gap-flagging instruction is designed to prevent, and it is worth stressing because it is the one that turns a genuinely weak case into a falsely strong-looking one. A case that honestly lacks a required prior therapy should surface that lack so a human can decide what to do, whether to chase the missing documentation, route to a different therapy, or hold the submission. A model that quietly invents the missing therapy hides the very problem the pharmacist needs to see. The fourth failure is tone over substance: a justification so polished and confident that it discourages scrutiny, where the very fluency that makes it persuasive to a payer also makes it persuasive to the drafter, who is lulled into submitting without verifying. The defense against all four is the same posture: ground at the input, flag the gaps, trace the assertions, and treat every drafted clinical fact as a claim to be checked rather than a fact to be trusted.

The Pharmacist Owns the Assertion

The cardinal rule of this entire program lives in the drafting step as surely as it lives anywhere: AI supports the pharmacist's judgment; it never replaces it. The AI assembled and drafted the justification, and that is genuine, valuable work that returned most of the 25 minutes. But the clinical assertion in that justification, the claim that this patient has this diagnosis and tried and failed these therapies and meets these criteria, is asserted by the pharmacist who signs and submits, not by the tool that drafted. "The AI wrote it" is not a clinical position a pharmacist can stand behind to a payer, a board, or a patient. The drafting tool is a faster typewriter with a dangerous imagination, and the professional accountability for every word that goes out under the pharmacist's credential stays exactly where it always was.

This is why grounded drafting is empowering rather than threatening to the pharmacist's role. The tool does not take the clinical judgment away; it removes the clerical assembly that buried the clinical judgment under twenty minutes of copy-and-paste, and it hands back a draft that is fast to produce and fast to verify when it was grounded properly. The pharmacist's attention, freed from the mechanical work, concentrates exactly where it should: on confirming that the load-bearing clinical facts are real, that the gaps are honestly surfaced, and that the assertion the pharmacist is about to make on behalf of a patient is one the chart actually supports. A drafting workflow that respects this boundary makes the pharmacist faster and the submission sounder at the same time. A workflow that blurs it, that lets the polished draft stand in for the pharmacist's verified judgment, trades the speed for a hidden risk that surfaces as a denial, a delay, or worse, at the worst possible moment, when a patient is waiting.

Key Takeaways

• AI-assisted justification drafting is the first hands-on step of the prior-authorization (PA) goldmine: it collapses the roughly 25-minute manual assembly by generating the structured clinical narrative from supplied facts in seconds, but the draft is only trustworthy if every clinical fact inside it is grounded in the actual chart.
• Drafting is primarily a generation job sitting on an extraction job, and a generation model is optimized for fluent and persuasive text, not for true text, so it has every incentive to write the convincing justification rather than the accurate one.
• Grounding means tying every asserted fact to a specific place in a trusted source; a grounded justification and a fabricated one can read identically on the screen, and the only difference, whether the facts are real, is the entire job in PA.
• Prompt for grounding at the input: supply the actual chart context, constrain the model to use only the supplied facts, instruct it to flag missing facts with a marked placeholder rather than inventing them, and ask it to note the source of each clinical assertion for later verification.
• The instruction to flag gaps instead of filling them turns the model's greatest weakness, the urge to complete the pattern, into a useful signal that shows you exactly where the chart is thin before you submit.
• The named failure modes are the fabricated clinical fact (most dangerous because invention makes the case stronger), the subtle extraction error feeding the draft, the confident gap-fill that hides a real weakness, and tone over substance that discourages scrutiny.
• The cardinal rule holds in drafting: the AI drafts, but the pharmacist who signs and submits owns the clinical assertion; "the AI wrote it" is not a position a pharmacist can stand behind to a payer, a board, or a patient.
• Grounded drafting is empowering, not threatening: it removes the clerical assembly that buried clinical judgment, returns a draft that is fast to produce and fast to verify, and concentrates the pharmacist's attention exactly where it belongs, on confirming the load-bearing facts are real.