Corti's Symphony Shows Why Medical Coding AI Must Reason, Not Just Label

Most AI coding tools work like a very fast lookup table: take a clinical note, find the most probable ICD-10 or CPT code from training data, return it. That approach has a fundamental flaw. Coding guidelines change constantly. ICD-10-CM adds and revises codes each October, CPT updates ship every January, and payer-specific rules shift quarterly. A model trained on last year's data can be confidently wrong on today's chart.

Corti, the Copenhagen-based clinical AI company, launched Symphony for Medical Coding on April 1, 2026, built around a different premise: agentic AI medical coding should work the way an expert human coder works — reasoning through evidence, checking the index, validating against guidelines, and reconciling conflicts — rather than matching patterns from historical training data. The results are striking enough to warrant attention from any RCM leader or CDI specialist evaluating AI coding tools this year.

What the "Code Like Humans" Research Actually Found

Symphony is built on Code Like Humans, accepted at EMNLP 2025. The study used 1.8 million patient encounters — the largest peer-reviewed study of AI medical coding of its kind. The core finding: AI systems that treat medical coding as a classification problem are structurally inadequate. Guideline changes break them. Corti claims Symphony outperforms OpenAI, Anthropic, Amazon, Oracle, and Microsoft by up to 25% on clinical accuracy benchmarks.

How Symphony's Four-Agent Pipeline Works

Symphony uses four AI agents that execute in sequence:

• Evidence extractor. Reads the clinical note and isolates documented conditions, procedures, and relevant history.

• Index navigator. Searches the ICD-10 alphabetical index for candidate codes based on extracted evidence.

• Tabular validator. Checks candidates against guidelines, excludes clauses, and specificity requirements.

• Code reconciler. Sequences the final code set, flags ambiguities for human review, and produces auditable output.

Why This Matters Beyond Accuracy Scores

The clinical stakes go beyond billing accuracy. Corti identified three times as many suicide attempts as had been officially coded in Danish patient data — cases present in clinical notes but missed by coders under time pressure. Those missed codes affect public health data, quality metrics, and resource allocation. For US health systems, this has direct financial implications under HCC, value-based contracts, and RADV audit scrutiny.

What It Means for Coders and CDI Specialists

Symphony operates via API through the Corti Console, supporting ICD-10-CM, ICD-10-PCS, and CPT. Every assigned code is linked to the clinical evidence supporting it, and ambiguities are flagged for human review rather than silently resolved. In a RADV audit or payer review, a coder who can point to the documented evidence for each code is in a fundamentally stronger position.

Evaluating AI Coding Tools in 2026

If you are evaluating agentic AI medical coding platforms this year, ask any vendor whether their system reasons through coding decisions or retrieves from a training distribution. Ask what happens to accuracy when ICD-10 guidelines update. Ask whether outputs are auditable. The answers will tell you whether you are buying a pattern-matcher or a system that reasons alongside your coders. Medikode's automated medical coding platform is built on the same premise.