AI coding intensity: the $942M optimisation nobody treated

AI coding intensity: the $942M optimisation nobody treated

There is a number in a September white paper that should interest anyone who has ever shipped a system that optimises a proxy metric. Over two years, inpatient cases billed as medically complex rose from roughly 37% to about 40% across the Blue Cross Blue Shield system — and the treatments attached to those cases did not move with them.

That gap is the whole story, and the Blue Cross Blue Shield Association put a price on it: about $942 million in incremental claims costs. Here is why the finding is bigger than a healthcare fight, why the number is contested, and what a developer should take from it.

The mechanism is a tiered price list, not a fee schedule

US hospitals are not paid per procedure. Each inpatient admission is assigned a diagnosis-related group (DRG), and the payer pays a fixed weight for that group. For a large family of DRGs the weight has severity tiers: no complication or comorbidity, then a complication or comorbidity (CC), then a major complication or comorbidity (MCC). One additional documented secondary diagnosis is often all it takes to move a claim up a tier, and up a tier is worth a lot more money.

So the incentive structure is legible. Billing a patient as more complex than their treatment supports is not fraud by default — but it is directly rewarded, immediately and per claim. You do not need a human being to be dishonest. You need software that is good at finding anything codable.

That is what AI-enabled revenue cycle management (RCM) does. Ambient scribes listen to the clinician–patient conversation and draft structured notes; coding AI then scans those expanded notes alongside lab results and problem lists to propose additional billable diagnoses. The clinical documentation gets denser, and a denser record contains more codable conditions. BCBSA notes that a secondary diagnosis "may be derived from single laboratory values, making it particularly well-suited for detection by AI tools" — and more than 60% of hospital systems are now using AI coding tools, per the association's release.

Clinical discordance, measured in a table

The methodological question is how you distinguish "patients got sicker" from "coding got greedier". BCBSA's answer is to look for cases where the diagnosis rose but the treatment that diagnosis implies did not. They call this clinical discordance.

Their September 2026 white paper, "Hospital Coding Intensity Analysis: Major Bowel Procedures" (authors Chris Birkmeyer, David Wennberg, Keith Kamons and Luke Chalker), drills into DRGs 329–331 — surgical procedures mainly used to treat colon cancer or diverticular disease, over half of them endoscopy-based. Compared with a 2023 baseline, hospitals classified 55,158 additional cases as complex, generating $653 million in incremental reimbursements, an average of $11,800 per excess complex case. Within that bowel family, the highest-tier (MCC) share went from 20.2% to 22.7% while non-complex fell from 36.6% to 32.8% — $60.8 million of incremental cost from one DRG family.

Then the table that matters. For 2025 bowel procedures, comparing the hospitals in the top 25% by complexity growth against everyone else:

More diagnoses. Less ICU. Less transfusion. Same length of stay. A validation check on one specific code makes it concrete: posthemorrhagic anemia. In the top-growth hospitals the anemia diagnosis rate was 13.7% versus 9.9% — 38% higher — but the transfusion rate among diagnosed patients was 16.9% versus 19.3%. Diagnose more anemia, transfuse less of it.

In code, the check is two aggregates and a ratio — no chart review, no clinician in the loop:

-- Did the diagnosis rise while the care it implies stayed flat?
WITH cohort AS (
  SELECT hospital_id, year, high_growth,
         complex, icu_used, transfused, los_days
  FROM   bowel_claims              -- DRGs 329-331
  WHERE  year IN (2023, 2025)
)
SELECT high_growth,
       count(*)                                              AS cases,
       avg(100.0 * complex)                                   AS pct_complex,
       avg(100.0 * icu_used)                                  AS pct_icu,
       avg(100.0 * transfused)                                AS pct_transfused,
       percentile_cont(0.5) WITHIN GROUP (ORDER BY los_days) AS median_los
FROM   cohort
WHERE  year = 2025
GROUP  BY high_growth
ORDER  BY high_growth DESC;

Read the output as a shape, not as a verdict: pct_complex climbs, pct_icu and median_los do not. Nobody has to believe anything about AI. The query is auditable, it runs every quarter, and it cannot be argued with — which is exactly why it is a usable control.

The white paper's own summary of why they read this as documentation behaviour rather than illness:

The consistent inverse relationship between diagnosis-based complexity and both aggregate resource utilization and diagnosis-specific procedural intervention is the strongest indicator that coding escalation reflects documentation practice changes rather than actual patient acuity shifts.

That is a Goodhart's law result with a table attached. The metric that pays the hospital is coded complexity. The thing that costs the hospital is treatment intensity. Optimise the first and the second is free to fall — and it did.

The counter-argument is real, and it is also an AI story

The American Hospital Association published a fact sheet in August calling the upcoding charges "unsubstantiated", and the case is not weak:

• Hospital case mix index — how sick the inpatient population is — rose about 5% between 2019 and 2024 per an AHA/Vizient analysis, and AHA data attributes 19% of hospital expense growth over that span to caring for sicker, more complex patients.
• Lower-acuity care keeps migrating to outpatient settings, so what remains inpatient is naturally more severe.
• The Medicare Payment Advisory Commission reported in 2025 that upcoding contributed to $40 billion in overpayments to Medicare Advantage plans, one payer settled a Justice Department upcoding suit in March 2026, and Massachusetts sued another in May 2026.

Three honest caveats on BCBSA's side, all of which they state or which follow from the method:

1. It is claims data, not charts. BCBSA acknowledged the limitation — clinical documentation would be the more direct test of whether patients were genuinely sicker. Fierce Healthcare reported Chalker conceding the same point.
2. It is a white paper from an interested party, not peer-reviewed, produced by an association whose members pay the bill.
3. Some of the increase is correct coding. Hashmi, BCBSA's VP of clinical affairs, put it carefully: "There may be an element of correct coding there, but the likelihood that this is technology-enabled upcoding is higher, in my view." Thinner charts before 2023 would mean part of the $942M pays for care that always happened.

And the strategy response is AI too, on both sides. BCBSA wants explicit rules for hospitals running coding AI, and wants payment tiers restructured so one extra diagnosis pays less. The AHA wants Congress and the administration to stop insurers using "downcoding" edits that auto-reject claims. As TechCrunch put it, Dr. Shiv Rao of Abridge warned of "a horrible dystopic future nobody wants to live in" with "bots fighting bots, agents fighting agents" — while allowing it might also cut costs. Chalker declined the war framing, calling it "a completely one-sided blood bath".

What this is actually teaching

Two things, and neither is about healthcare.

Incentives are part of your threat model. Every engineer eventually ships a system that optimises something. When the reward function and the ground truth come apart, the gap is where the incident happens — and no amount of model quality fixes it. The optimiser here is not misbehaving; it is doing exactly what the DRG weight table pays it to do. Ask what your system maximises, what actually costs money, and whether anyone is checking that the two still move together. A dashboard showing "complex case rate: up 3 points" with a second dashboard showing "ICU utilisation: flat" is a monitoring gap, not a billing dispute — and the cheapest fix is a second number, not a cleverer pipeline. There is a version of that argument about resisting machinery for its own sake in the art of simple system architecture.

Counters are arms races too. The insurer's response is not a rule change, it is an automated analysis pipeline looking for the same asymmetry. Every deployment on the provider side makes the payer-side detector better, and vice versa. BCBSA's own framing is that the trend "hasn't stopped" and is not confined to inpatient care — so the interesting engineering question is not who wins but what a verification layer looks like when both sides are optimising. Simon's lethal trifecta reasoning applies with the roles swapped: you have a metric that pays, an input that is noisy, and an automated actor optimising the first against the second.

If you work on claims, revenue, risk scoring, fraud detection, ad ranking or anything else where a proxy drives money, the useful takeaway is one sentence long: when a number starts moving, check whether the thing it is supposed to represent moved with it. BCBSA just published the receipts for doing exactly that, at scale, on $942 million of somebody else's money.

Sources

• Blue Cross Blue Shield Association: Analysis examines AI in hospital billing as spike in complex patients adds nearly $1 billion in extra costs
• Blue Cross Blue Shield Association: Hospital Coding Intensity Analysis: Major Bowel Procedures
• Blue Health Intelligence / BCBSA: Rising Coding Intensity and Its Impact on Health Care Affordability
• Fierce Healthcare: Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care: analysis
• TechCrunch: Insurers claim AI is already increasing healthcare costs
• American Hospital Association: Fact Sheet: Artificial Intelligence and Coding Intensity