AI agent · Healthcare · Review against rules
Coding review
The agent reads each coded episode alongside its clinical documentation and checks every code against the rules in force. Episodes that agree pass; disputed codes go to a coder with the reason and the source passage.
Typical volumes for this process, not a client figure.
A sample gets reviewed because reviewing all is impossible.
All of it reviewed; the disputed ones go to a coder.
Where the time goes today
After a stay or a visit, coders translate the clinical documentation into diagnosis and procedure codes. Those codes decide what the provider is paid under case-mix or tariff systems, and they feed activity statistics and research. An audit team then reviews a small sample of episodes, chosen at random or by risk, because checking every one would need nearly as many auditors as coders.
What the sample misses is where money and accuracy go. A comorbidity recorded in a nursing note but never coded, which lowers the group the episode falls into. A procedure coded without the accompanying act the rules require. A level of care the documentation does not support, which a payer audit later recovers. Errors cluster by specialty, by coder and by documentation habit, and a random sample finds them slowly. When the payer's audit finds them first, the correction is expensive and sometimes comes with a penalty.
How the agent works
- Read the episodeThe agent reads the coded record together with the documentation for that episode: discharge summary, operative notes, results, prescriptions, nursing notes.
- Check each codeFor every code it asks whether the documentation supports it, whether the sequencing follows the rules, whether a required accompanying act is present, and whether the documented care supports the level coded.
- Find what is missingIt looks for documented conditions and acts that affect grouping or payment but carry no code.
- Route the disputesEpisodes where every code agrees pass. Disputed codes go to a coder with the proposed correction, the rule and the source passage, prioritised by impact and by the submission deadline.
- Record the outcomeThe coder accepts, amends or rejects each proposal. The decision is stored, added to the test set, and used for feedback on documentation to clinical teams.
What stays with a person
Coders decide every correction. The agent proposes; a qualified coder accepts, amends or rejects, and only then does the code change in the system. Where documentation is ambiguous, the question goes to the clinician through your usual query process. The agent can draft the query, but a coder reviews it before it reaches a clinician.
Coding managers own the interpretation of the rules, local conventions, and any decision to resubmit episodes that have already been paid. Resubmission after payment has financial and relationship consequences with payers, so it is always a person's call.
What it reads, what it produces
| It reads | It produces |
|---|---|
| Coded episodes from the coding or patient administration system | An episode-level review with every code marked agreed or disputed |
| Clinical documentation: summaries, operative notes, results, prescriptions, nursing notes | Proposed corrections, each with the rule and the source passage |
| The classifications and coding rules in force, and your local conventions | A prioritised worklist for coders |
| The grouping or tariff logic that turns codes into payment | A report of recurring errors by specialty and documentation type |
| Findings from earlier internal and payer audits |
Controls that come with it
- No code changes without a coder's acceptance.
- Rules and local conventions are versioned, and each review records the version it used.
- Changes above a set financial or grouping impact are reviewed by a senior coder.
- Coders audit a weekly sample of episodes the agent passed, to measure what it misses.
- Every proposal traces back to a source passage and a rule, and corrections after submission follow your normal resubmission procedure with manager approval.
How you know it works
- Share of episodes reviewed, from the current sample to all
- Agreement between the agent's proposals and coders' decisions, per error type
- Errors found by payer or external audit, before and after go-live
- Days from discharge to final coding
- Corrections by specialty, to target documentation improvement
Is your process ready?
- The coding rules and local conventions are written and current, including the ones coders learn by word of mouth.
- The coding system and the record expose episodes and documentation through an interface or export, and the licences for both systems and for the classification permit software use.
- A coder can confirm or reject a proposal in a minute or two, because the source passage is attached.
- Episode volume is too high for full manual review, which is exactly why only a sample is reviewed today.
- Coders and auditors agree on the answer in clear cases. Where they do not, write the convention down first.
The five candidacy checks are explained, with an exam, in the free Module 01.
What goes wrong
- Results and letters arrive after discharge and the agent reviews an incomplete episode. Set a waiting period before review.
- Local conventions were never written down, the agent contradicts accepted practice, and coders lose trust in the first weeks.
- Coders are buried in low-value disputes. Prioritise by impact and tune the threshold.
- The classification is revised and the rules switch on the wrong date for episodes that span the change.
Questions we get
Does the agent code episodes itself?
Not in this process. It reviews coding done by coders or by other tools, proposes corrections and never changes a code. Drafting codes for uncoded episodes is a different process with different controls and a different test set. Some providers do both; keeping them separate means the review stays independent of the drafting.
Will it push coding upwards to raise income?
It checks agreement with the documentation in both directions. A code the notes do not support is disputed just as firmly as a documented condition left uncoded, and both are measured before go-live. The aim is coding that survives a payer audit, not a higher case-mix. Overcoding found after payment costs more than it earned.
Which classification systems does it work with?
The classification and grouping rules in force where you operate, provided they are written down and your licence allows software use. Rules and local conventions are loaded as a versioned rule set, and every review records which version it used, so episodes that span a revision are checked against the right one.
How do we measure it before trusting it?
Take past episodes that have already been audited and whose final codes are agreed. Run the agent on them and compare its proposals with the agreed codes, error type by error type. After go-live, coders audit a weekly sample of passed episodes. The same set is rerun whenever the classification or the underlying model changes.
Want this agent on your process?
Tell us about your version of this process — volumes, systems, what goes wrong. A person answers with an approach and a price, usually within two working days, or tells you it is the wrong project.