Every experienced coder has a browser habit nobody put in the job description.

One tab for the CPT description. One for the NCCI edit table. One for the payer's LCD. One for last month's denial that looked exactly like this one. By the time the code gets entered, there have been twelve tabs open and forty minutes spent confirming something the coder was probably already right about.

This closes the series. Issue 5 covered what's newly possible with general-purpose AI in healthcare operations. Issue 6 went deep on data infrastructure, the dashboards and databases. This issue closes it out with the part that touches revenue the most directly: coding lookups, denial pattern research, and prior authorization documentation.

The theme underneath all three issues hasn't changed. AI isn't replacing coding expertise, denial judgment, or clinical documentation review. It's closing the distance between knowing something and confirming it.

Coding-related issues drive an estimated 25 to 30 percent of initial claim denials, according to AAPC's Coding and Compliance Report. Not because coders don't know the rules. Because the rules live across dozens of payer policies, NCCI edit tables, and LCDs that update on their own schedule, and no one carries all of it in their head at once.

That's a lookup problem, not a knowledge problem. And lookup is exactly where AI earns its keep before it earns anything else.

"What are the current NCCI edit conflicts when CPT [code A] is billed with CPT [code B]? What modifier, if any, is typically required, and what documentation supports billing both on the same date of service?"

The output isn't a coverage determination. It's a fast first pass that tells the coder whether this combination is worth a closer look, before the claim goes out instead of after it comes back. The coder still makes the call. They're just making it with fewer open tabs.

Every revenue cycle team has denials that repeat. The same payer, the same reason code, the same service line, month after month, treated each time as a one-off because nobody had the hour it takes to pull the pattern out of a spreadsheet.

"Here's a de-identified table of denials from the past quarter [paste columns: payer, denial reason, CARC/RARC code, service line, amount]. Group these by payer and denial reason, rank them by total dollars at risk, and tell me the top three patterns worth investigating. For each, suggest what upstream step might be causing it."

This is the same instinct as the dashboard work from Issue 6, applied to a question that's usually asked too late. Most teams find the pattern during the annual review. The tool exists now to find it during the month it's happening, while there's still time to fix the workflow instead of just appealing the claim.

Prior authorization is where this adds up fastest, because the numbers are no longer subtle. According to the AMA's 2025 Prior Authorization Physician Survey of 1,000 physicians, the average practice completes 40 prior authorizations a week and spends 13 hours of physician and staff time on it. Thirty-two percent of physicians say requests are often or always denied, and 74 percent say denials have gotten worse over the past five years.

Most of that time isn't spent deciding whether care is appropriate. It's spent assembling the documentation to prove it, often after a denial has already arrived and the appeal clock has started.

"For a prior authorization request for [service] under [payer]'s published policy, list the clinical documentation elements typically required to establish medical necessity: conservative treatment history, imaging findings, symptom duration, failed therapies, and anything else specific to this service. Format it as a pre-submission checklist my team can run before sending."

The same approach works in reverse for a denial that's already landed:

"Draft a medical necessity appeal letter template for a [payer] denial of [service]. Structure it with: claim identifiers left blank, the payer's stated denial reason, a medical necessity argument section, a list of supporting documentation to attach, and a professional closing. Keep it factual, not persuasive."

Neither prompt predicts what a payer will do. Both close the gap between what your documentation says and what the policy requires, before that gap becomes someone's Tuesday afternoon three weeks from now.

The same three rules from every issue in this series, because they don't change with the use case.

Don't put patient data in. General-purpose AI tools aren't HIPAA-covered by default. Work with de-identified data and structure-only descriptions. PHI requires an enterprise-tier tool with a signed Business Associate Agreement.

The output is a draft. A flagged NCCI conflict, a denial pattern, an appeal letter: all of it gets reviewed by someone who knows the payer, the patient, and the policy. AI can surface what's likely relevant. It can't be accountable for the claim.

And decide who owns the review. The risk in this kind of work was never that AI gets it wrong occasionally. It's that a confident, fast, wrong answer gets trusted because it arrived quickly. Confident and wrong is a dangerous combination in prior authorization, and speed makes that combination easier to miss, not harder.

Three issues, one thread. Issue 5: the blank page is no longer the barrier it was. Issue 6: the infrastructure you'd have needed an analyst to build is now something you can draft yourself. Issue 7: the lookups, patterns, and documentation that used to eat a coder's or a prior auth team's day are now a starting point instead of a search.

None of it replaces the judgment that made your best coders and auth specialists valuable in the first place. If anything, it protects that judgment, by taking back the hours that used to go to finding the answer instead of applying it.

The organizations that get value from this aren't the ones with the newest tool. They're the ones who were already clear on where their time was actually going, and pointed the tool at exactly that.

Which denial on your team is the one everyone recognizes on sight, because it's happened a dozen times and nobody's fixed the upstream cause?

How many of your coder's twelve tabs are really a lookup, and how many are actually a judgment call?

And when the tool is confidently wrong, who's positioned to catch it?

Sources: AMA, "2025 Prior Authorization Physician Survey" (1,000 physicians, results published 2026); AAPC, Coding and Compliance Report (coding-related issues driving 25 to 30 percent of initial claim denials).