01 the business
Community hospital
220 beds, coding team of 12
The coding team runs a two-week backlog, which delays billing and reporting. Vague documentation leads to queries and denials. Leadership has heard that AI can code, but early tests with a chatbot produced codes that did not exist.
02 before
The problem
Coders read long records to assign diagnosis and procedure codes, and backlogs delay billing. Missed or vague documentation leads to wrong codes and denied claims. General AI chatbots are not reliable at this on their own: they often produce codes that do not exist.
- A two-week coding backlog.
- Coders spend most of their time finding evidence in long records.
- Documentation gaps are found after claims are denied.
03 the build
What the AI does
- Step 1: Reads the finished encounter or admission record once it is signed.
- Step 2: Finds the statements that support a diagnosis or procedure, such as the reason for a test or the operation performed.
- Step 3: Looks up candidate codes in the official code set, such as ICD-10-CM, ICD-10 and OPCS-4, or CPT, rather than writing codes from memory.
- Step 4: Ranks the candidates and shows the exact sentence that supports each one.
- Step 5: Flags documentation gaps, such as an unspecified condition, so a query can go to the clinician.
paused: waiting for a person to approve
How it was put together
- Looked up candidate codes in the official code sets instead of letting a model write codes from memory.
- Showed the exact sentence that supports each candidate.
- Ranked candidates and flagged documentation gaps for queries.
- Logged every coder decision to measure accuracy over time.
04 human in the loop
Where a person stays in charge
Certified coders choose every final code. Queries to clinicians are written or approved by a coder, and the tool never changes the clinical record.
- Certified coders choose every final code.
- Queries to clinicians are written or approved by a coder.
- The tool never changes the clinical record.
05 rollout
From kickoff to live in 8 weeks
- Weeks 1–3Pick two high-volume specialties and set up code set look-ups.
- Weeks 4–6Run alongside coders and measure agreement on a sample.
- Weeks 7–8Use for those specialties; add more as agreement holds.
06 after
What changes
Coders spend their time on judgement rather than searching, and backlogs shrink. Codes are better supported by the record, which helps when claims are reviewed.
Assume 12 coders each coding 40 charts a day, and evidence-finding support raising that to 48.
12 × 8 extra charts = 96 more charts coded a day, enough to clear a 2,000-chart backlog in about 21 working days.
What to measure
- Coding backlog in days.
- Charts coded per coder per day.
- Agreement between suggestions and final codes.
- Denials for coding reasons.
07 evidence
What the research says
Asked to produce billing codes from descriptions, GPT-4 matched the correct code exactly only 33.9% of the time for ICD-10-CM and 49.8% for CPT.
NEJM AI, Large language models are poor medical coders (Soroush et al.) (opens in a new tab) · 2024
HealthCare.gov insurers denied 20% of in-network claims in 2023, and fewer than 1% of denied claims were appealed.
KFF, Claims denials and appeals in ACA Marketplace plans in 2023 (opens in a new tab) · 2025
Moving US healthcare admin transactions from manual to automated could save $20 billion a year; a prior authorisation takes providers about 24 minutes by phone, fax or email and 16 minutes through a portal.
08 lessons
Mistakes to avoid
- Letting a model write codes. Use the official code sets.
- Measuring speed only. Track accuracy and denials too.
- Skipping coder input. They know where the hard cases are.
09 questions
Common questions
- Can AI code on its own?
- Not reliably. A 2024 NEJM AI study found GPT-4 matched the exact ICD-10-CM code only 33.9% of the time when generating codes from descriptions. Code look-ups plus a coder's decision are far safer.
- Which code sets are supported?
- ICD-10-CM and CPT in the US, ICD-10 and OPCS-4 in the UK, and others your payers use.