Not demos — production pipelines where a wrong answer is a compliance finding or a denied claim. Where the models failed, what finally changed, and what HIM leaders should do about it this quarter.
I've spent the last two years building AI into real production work: auditing complex inpatient records on the HIM side, and working claims, payment posting, and payer audits on the revenue side. Not demos — pipelines that have to be right, because a wrong answer is a compliance finding or a denied claim. And for most of those two years, I kept hitting the same wall.
The models could read a multi-page inpatient chart. What they couldn't do was think critically across it. They couldn't hold the ED note, the labs, the op report, four days of ICU flowsheets, and the discharge summary in tension against coding guidelines the way a credentialed coder does. The longer and more complex the chart, the more the reasoning fell apart. A model would apply a guideline correctly on page 5 and contradict itself by page 40, or miss the one data point that changed everything.
For a while I assumed the fix was bigger context windows. It wasn't. The charts fit; the sustained reasoning didn't.
With the newest generation of frontier models — most recently Anthropic's Claude Fable 5 — that wall has moved. And the specific things that changed map almost perfectly onto the problems HIM departments and coding operations struggle with most.
Three capabilities matter here, and none of them is "it knows more codes."
Sustained, long-running work. The newest models are built to execute complex tasks over extended periods: planning an approach, checking their own progress against the goal, and refining the work as they go — with the gap versus prior models widening as tasks get longer. In practical terms, this is the difference between a tool that can look at one chart and a tool that can be assigned a project: a pre-bill audit of a service line, a quarterly coding quality review, a denial root-cause analysis across an entire payer.
Vision. So much of revenue cycle work is still trapped in documents that were never designed for machines: scanned EOBs, faxed records, handwritten orders, payer correspondence. Stronger vision capability means more accurate extraction from exactly the messy, real-world documents that have historically broken automation projects. Anyone who has tried to automate payment posting from paper EOBs knows the bottleneck was never the posting logic — it was reading the document correctly in the first place.
Depth on knowledge work. Coding is not pattern matching. It is applying a dense, constantly changing rule set — Official Guidelines, Coding Clinic, NCCI edits, LCDs, payer policy — to ambiguous clinical narratives. Models that reason more carefully over long documents and complex guideline interactions get materially better at exactly the cases that matter most.
If you code or audit for a trauma center, you already know these cases. They're the ones where the answer lives in five different documents — which is exactly where AI used to fall apart.
Mechanical ventilation over or under 96 hours is one of the largest single payment swings in inpatient coding, often $20,000 or more, and a perennial RAC target. Getting it right means reconstructing the vent clock from the ED intubation time, the anesthesia record, days of ICU flowsheets, and any extubation and re-intubation events. Miss one re-intubation buried in a flowsheet and the DRG is wrong. That's not a coding-knowledge problem; it's a multi-document reconstruction problem.
An OIG audit target for decades. The PCS root operation — and frequently the DRG — turns on whether the operative note supports actual excision: documented instrument, depth, tissue cut away. The phrase "sharp debridement" alone doesn't settle it. The evidence has to be read critically in the surgeon's own language, not keyword-matched.
Principal diagnosis sequencing in a polytrauma admission, where the splenectomy-versus-femur call swings the DRG dramatically. Sepsis in the trauma patient, where POA status and organ dysfunction linkage live across the ID consult, lactate trends, and vasopressor records. Acute blood loss anemia hiding in a hemoglobin trend no physician explicitly documented. ATN versus unspecified AKI — an MCC versus a CC — sitting in CK trends and nephrology notes. Every one requires holding a dozen data points in tension across a long record. That is precisely the capability that just changed.
Audit readiness first, because that's where the pressure is. Between SMRC reviews, TPE, RAC, and payer-specific ADRs, departments are being asked to defend more claims with the same staff. A model that can sustain a long review means second-level review stops being a sampling exercise. Instead of auditing 30 charts a quarter and extrapolating, you can realistically pre-screen every chart in a high-risk category — botulinum toxin injections under an active LCD, spinal fusions, high-weighted DRGs — before the claim goes out the door. The audit posture shifts from reactive to preventive.
Coding quality programs change shape. Today, a quarterly QA cycle is weeks of coordinator time: pulling samples, reviewing, writing findings, building education. A long-horizon model can do the first pass across the full population, flag genuine discrepancies, and draft the findings report — leaving credentialed auditors to do what only they can do: adjudicate the hard calls, contextualize trends, and educate coders. The auditor's judgment becomes the scarce resource applied where it matters, not the bottleneck applied everywhere.
On the revenue cycle side, document-heavy workflows are the obvious early wins. EOB conversion and payment posting, correspondence triage, prior authorization packet assembly, charge capture reconciliation — all "read messy documents, apply rules, produce structured output" problems. They were partially automatable before. Better extraction accuracy plus the ability to self-check arithmetic and balancing makes full automation with human exception handling a realistic near-term architecture rather than a roadmap slide.
At the facility level, an abstract is the ultimate multi-document reasoning task. A registrar synthesizes pathology reports, imaging, operative notes, and oncology treatment summaries into hundreds of discrete data fields: AJCC TNM staging, Summary Stage, site-specific data items that change by primary site, and first-course treatment assembled from records generated over weeks. With abstraction backlogs measured in months at many facilities and credentialed registrars in chronically short supply, a model that has already performed casefinding from the pathology feed, assembled the relevant documents, and drafted the abstract for registrar validation changes the timeliness math entirely — including the CoC accreditation standards for abstracting timeliness and the follow-up rates that surveys actually check.
The implications run upstream to central registries too. State and SEER registries spend enormous effort consolidating multiple facility reports of the same tumor into a single record — reconciling conflicting staging, treatment, and demographics submitted by three different hospitals about one patient. That is, again, a long-context critical reasoning problem, and the same applies to death clearance follow-back and the flood of ePath reports needing casefinding review. The registrar's role shifts toward what certification was always for: validating staging logic, resolving ambiguous primaries, and querying physicians — consolidating exceptions instead of everything.
What arrived isn't a replacement for coders and registrars — it's a tool finally powerful enough to take the repetitive volume off our teams, so the humans can focus on the judgment calls that were always the real job.
None of this means autonomous coding without human accountability. The compliance reality has not changed: a credentialed coder or auditor must own the final code assignment, and your compliance plan should treat model output the way it treats any other decision-support input — validated, monitored, and overridable. The departments that get burned will be the ones that confuse "the model can sustain long work" with "the model no longer needs oversight."
Privacy diligence gets more specific, not less. Frontier models can ship with vendor-side safety monitoring and data retention terms that differ from older model tiers. If you operate under a Business Associate Agreement — and for any PHI workflow, you must — confirm exactly how those retention terms interact with your BAA before routing clinical documents through a new model. This is a contract review conversation, not a technical one, and it should happen before the pilot, not after.
Be realistic about cost. Frontier capability is priced like frontier capability. The right strategy is not to route everything through the most powerful model — it's to benchmark where the accuracy delta actually changes outcomes (fewer posting exceptions, fewer audit findings, fewer denied claims) and pay for capability only where it earns its keep.
HIM has spent a decade hearing that AI was coming for medical coding. After two years of building with it, I can tell you what actually arrived is something more useful.
CODEMED pairs AHIMA/AAPC-certified coders, ODS-C registrars, and auditors with production-tested AI tooling — so you get the speed without the compliance risk.
