AI-assisted medical coding: suggestions your coders can defend

Clinical coding is one of the few places in a hospital where a software error has a directly calculable dollar value. Under activity based funding, the 2025-26 National Efficient Price is $7,258 per NWAU(25), a 5.9% increase and the largest since national ABF began. A tonsillectomy is 0.7421 NWAU, about $5,386. A hip replacement at minor complexity is 4.0251 NWAU, about $29,214. Miss an additional diagnosis that shifts complexity level and you have moved five figures on a single episode.

That is why coding keeps coming up when we talk to health CIOs about where AI might actually pay for itself, and it is why we keep giving the same answer: build a suggestion engine, not a coding engine. This post is part of our Practical AI in Health series, and it covers the architecture we would put in front of a coding team, what the evidence supports, and what it costs.

The problem is scale, not competence

Separations grew about 7%, from 11.3 million in 2018 to 12.1 million in 2022, while the coder workforce ages and universities close their health information management programs. Half the previously offered undergraduate HIM courses were discontinued because enrolments were too low, which the AIHW flagged in its analysis of the coding workforce shortfall. Northern Health in Victoria moved off its legacy coding system for exactly this reason. Its health information services director Odette Taylor said the AI component reduces the need to scale up the coding workforce as separations rise, with real-time feedback to coders supporting accurate DRG and NWAU reporting.

So the demand is real. The question is what you can safely hand to a model.

Why fully automatic coding is not defensible today

The strongest single citation here is the NEJM AI benchmark by Soroush et al., which tested GPT-3.5, GPT-4, Gemini Pro and Llama2-70b on generating ICD-9-CM, ICD-10-CM and CPT codes. GPT-4 was the best of them and still managed only 33.9% exact match on ICD-10-CM. The models also produced codes that were valid-looking but non-billable, or simply fabricated. That was code lookup from a supplied description, which is easier than abstracting a code from a discharge summary.

Now add the Australian complication. Admitted episodes here are coded to ICD-10-AM/ACHI/ACS Thirteenth Edition, in force for separations from 1 July 2025, with AR-DRG V12.0 sitting on top. AR-DRG V12.0 assigns Diagnosis Complexity Levels across 11,057 in-scope ICD-10-AM codes. A general-purpose model trained on the American clinical modification does not know that classification, and the editions change on a three-year cycle. Whatever you build has to be versioned to the edition in force on the separation date, not to whatever the model absorbed during training.

The assisted configuration is a different story. A 13-week randomised controlled trial with ten certified coding specialists, published in npj Digital Medicine, found AI-assisted workflows significantly reduced coding time while maintaining accuracy. Satisfaction varied with coder experience and certification, which matters more than it sounds: a tool your senior coders distrust will be worked around.

HIMAA’s national guideline for AI-generated clinically coded data lands in the same place, listing human expertise and human-in-the-loop among its seven considerations, and noting honestly that published evidence of benefit in the Australian context is still lacking. Nobody has proven this here yet. Design accordingly.

The architecture

The pattern we would build looks like this:

• Documents in via HL7 v2 or FHIR from the EMR and PAS: discharge summary, operation reports, progress notes, pathology. Getting these flowing reliably and repeatably is ordinary integration work, which is what we built Centazio for.
• A code index built from the licensed IHACPA code lists for each edition, held separately per edition, with the separation date deciding which index is queried.
• Retrieval first, generation second. The model never emits a code string freely. It nominates candidates from the retrieved set, and anything not present in that edition’s index is discarded before it reaches a human.
• A rules layer applying Australian Coding Standards checks: principal diagnosis sequencing, additional diagnosis criteria, mandatory ACHI pairings.
• A coder-facing panel that shows each suggested code with the exact text span it came from, so review means reading two sentences rather than re-reading the record.
• An append-only audit log, written before the coder sees anything.

The audit record is the part people skip and later regret. It should look roughly like this:

{
  "episode_id": "",
  "separation_date": "2026-03-14",
  "classification_edition": "ICD-10-AM 13th",
  "grouper_version": "AR-DRG V12.0",
  "model_id": "coder-assist-2026-02",
  "index_snapshot": "sha256:",
  "suggested": [{"code": "J45.0", "confidence": 0.82, "evidence_doc": "DS-4471", "span": [1180, 1246]}],
  "coder_id": "",
  "action": "rejected",
  "final_code": null,
  "reviewed_at": "2026-03-19T09:41:12+11:00"
}

That log gives you two things. Externally, it demonstrates that a qualified human assigned every code and that the tool suggested rather than decided. Internally, it is your quality feedback loop. Coding errors are systematic, not random: external audits of 55 hospitals across two states, covering 6,300 records, found a 5.9% predicted DRG mismatch rate, and at one hospital a 5.6% mismatch still translated to nearly $8 million less funding, because one error in a high-value DRG repeated many times. A log keyed by code, coder and DRG surfaces that repetition. Spreadsheets of aggregate accuracy do not.

Regulatory and privacy position

A coder-facing suggestion tool generally sits outside TGA medical device regulation under Exclusion 14G, which covers software for the administration or management of health processes including financial records, claims and billing. That exclusion holds only if every function meets the criteria and the tool does not influence clinical decision-making. The moment someone asks for the suggestions to be surfaced to treating clinicians during the episode, you have changed the intended purpose and you need to re-examine the classification.

Governance can follow the ACSQHC’s AI clinical use guide, released in August 2025 and structured around before you use, while you use, and after you use. The three phases map cleanly onto evidence review, coder training with monitoring, and scheduled re-evaluation after each classification edition change.

Privacy is the part that should reach the board. These systems ingest full discharge summaries, operation reports and progress notes. Health service providers were the most-breached sector in the OAIC’s 2025 notifiable data breach statistics, at 19% of notifications (225 of 1,205). Decide early whether inference runs in Australian-hosted infrastructure, who can query the index, and how long request payloads are retained. Those answers are harder to change later.

What it costs, honestly

Budget for licensed IHACPA classification products, a re-index and re-validation every three years when the edition changes, and inference costs that scale with document volume rather than with episodes. The largest cost is not the model. It is assembling a gold-coded evaluation set from your own separations, ideally auditor-reviewed, plus the coder hours to run a proper before-and-after measurement. Without that set you cannot tell whether the tool is helping or quietly teaching your team to accept plausible wrong codes, which is the failure mode that worries us most.

Start on a single high-volume specialty, measure minutes per episode and post-audit mismatch rate against a matched baseline, and keep the coder as the person who assigns the code. That configuration is the one the evidence supports.

PicNet builds production AI systems for Australian organisations. Talk to us about what a first project could look like.

Tagged: #clinical-coding #icd-10-am #activity-based-funding #human-in-the-loop #health-it