The short answer
“You upload a chart. A few seconds later there is a list of suggested codes, each with a note about why. From the outside that looks a bit like magic. From the inside it is a fairly ordinary chain of steps: the record is taken in, read, mapped to candidate codes, checked against rules, scored, and handed to a person with the evidence attached. The software does the reading. The coder does the judging.”
How AI Coding Software Works
Strip away the branding and almost every AI coding tool does the same five jobs. Where tools differ is mostly in how well they read clinical context, how current their rules are, and how clearly they show their evidence.
The five jobs almost every AI coding tool does
Takes in the record
Charts arrive as notes, dictations, reports, scanned pages and EHR exports. The software turns all of that into text it can work with.
Reads the text
The system works out what the note says: which conditions are present, which were ruled out, which services were performed, which were only planned.
Proposes codes
The clinical facts get mapped to candidate diagnosis and procedure codes from the standard code sets, such as ICD-10 and CPT.
Checks codes against rules
Coding guidelines, payer requirements and documentation standards decide whether the proposed codes hold up.
Hands the result to a person
The output goes to a human reviewer, with enough context for them to say yes or no.
Where tools differ is mostly in how well they do steps two and four, and in what they show you in step five. Step two is the hard part, step four is where accuracy becomes defensible, and step five is where trust gets built or lost.
The First Seconds After Upload
Coding Studio is the workspace where a coder reviews what the system produced. It sits inside RCM.ai and acts as the front end for Chart.ai’s coding output. Here is one chart moving through it, moment by moment.
The chart arrives
Before anything clever happens, there is cleanup. Text is normalized, pages are separated, and the encounter is identified. If a page is missed or two encounters get blended, everything after inherits the mistake.
The system reads
It pulls out the clinical facts: what is documented versus mentioned in passing, what was done versus only planned, details that change the code, and anything contradictory. Denies shortness of breath and reports shortness of breath share most of their words and mean opposite things.
Candidate codes appear
The facts are mapped to possible diagnosis codes, procedure codes and modifiers. Candidate is the right word. Nothing has been checked against the rules yet.
The rules get applied
The candidates meet the Playbook. The system asks whether each code is valid for the date of service, whether the documentation supports it, whether the diagnosis supports the procedure for this payer, and whether anything the rule requires is missing.
Confidence and flags
Each suggestion gets a confidence score, which measures the system's certainty, not the truth. Compliance flags mark a documentation gap, a possible conflict, or a rule that was not satisfied.
It lands in front of the coder
In Coding Studio the coder sees the codes, scores, evidence and flags. Each suggestion points back to the sentence in the record and the rule that applied, so review becomes checking specific claims against specific evidence.
The decision moves on
Accepted codes go into the billing workflow. Rejections and corrections matter too, because a pattern of rejections on one payer usually means a rule needs sharpening.
The full path:
Upload, read, propose, check, score, review, deliver. The software does the first five quickly. The person does the sixth. The seventh is where the work turns into a claim.
Where It Can Go Wrong
Almost all of the heavy lifting happens before a human sees the chart, and almost all of the judgment happens after. Machines are good at reading a lot of text quickly and applying the same rule the same way. People are good at judgment and noticing when something feels off. It would be dishonest to describe the happy path only, so here is where things tend to break, in any tool.
Reading mistakes
Negations, copied forward text and rule out language trip up readers. If the system misreads a note, every later step inherits the error.
Missing context
If the key fact lives in another note, a letter or an earlier result, a tool that only reads one encounter never sees it.
Stale rules
Payer policies and code sets change. A rule library that is not maintained drifts away from reality.
Overconfidence
A system can be sure and wrong at once. This is why confidence works as a routing aid, not as proof.
Weak review habits
If reviewers click accept without reading the evidence, the human safeguard is gone in practice, even if it exists on paper.
Good software is designed around these risks. It shows its evidence, keeps rules reviewed, reads across the history, and makes it easy for a person to disagree.
Chart.ai Core Coding
Chart.ai is the Healthcare Intelligence layer Grelin built on its Playbook engine. It is the part that reads. It looks across a patient’s full history, assembles one picture of every condition, setting and year, and reasons across it against clinical and payer rules. Coding Studio is where much of that output arrives for review. A few things define how Chart.ai handles the coding step.
It is evidence first
Every finding traces to the rule it came from and the patient text behind it, so a coder can verify quickly and the work holds up in audits and appeals.
It reads the whole patient
Because it assembles the full history, it can surface conditions documented elsewhere but never coded, and check coverage requirements that depend on prior treatment.
It is fast enough for the visit
Checks finish in about a second, so gaps can be fixed before the provider signs the note instead of weeks later.
Its rules are reviewed
Rule changes are made centrally, and each one is checked by a certified coder and a clinician before it goes live.
A human has the final word
Chart.ai produces intelligence. It never diagnoses, never writes to the chart on its own, and never acts in its own name. The coder reviews and decides.
What to look for when you watch a demo
Try the demo on one of your own messy charts, not a clean sample. That is where differences show up.
Does it show the evidence for each code?
A code with no source is a claim you cannot check.
Does it read beyond the single visit?
Ask how it handles information from earlier records.
Does it flag gaps, or only produce codes?
Gaps are where denials are prevented.
Can you disagree easily?
Accepting, rejecting and correcting should be quick and obvious.
What happens after you reject something?
Good systems learn from corrections through governed rule updates.
Who writes and reviews the rules?
Look for named expertise and a clear process.
Where does the output go?
It should land in the tools your team already uses.
The bottom line
The best AI coding software makes the last step easy: clear evidence, honest flags, and a quick way to say yes, no, or not quite. Read the whole patient, apply reviewed rules, show the work, and let an expert decide.
Frequently Asked Questions
How does AI coding software work?
Most tools follow the same sequence. They take in the record, read the clinical text to identify conditions and services, map those facts to candidate diagnosis and procedure codes, check the candidates against coding and payer rules, and pass the result to a human reviewer. Tools differ mainly in how well they read clinical context, how current their rules are, and how clearly they show their evidence.
What is Coding Studio?
Coding Studio is the workspace where an expert coder reviews AI coding output. Charts are pre-processed in seconds, and each one arrives with suggested codes, confidence scores, the supporting evidence, and compliance flags. The coder accepts, rejects or corrects each suggestion. It sits within RCM.ai and also acts as the front end for Chart.ai's coding output.
What happens right after I upload a chart?
The chart is taken in and cleaned up, the system reads the narrative and extracts the clinical facts, candidate codes are proposed, and those codes are checked against the relevant rules. Each suggestion gets a confidence score and any compliance flags, and the result is presented to a coder with the evidence attached. Nothing is final until a person decides.
Does the AI finalize codes without a coder?
In Chart.ai and Coding Studio, no. Suggested codes are delivered for an expert coder to accept or reject, and the human has the final word. Chart.ai produces intelligence. It never diagnoses, never writes to the chart on its own, and never acts in its own name.
What does a coder actually see in Coding Studio?
For each chart, the coder sees the suggested codes, a confidence score for each, the evidence linked to each suggestion, and compliance flags where a rule or documentation gap needs attention. Because every suggestion points back to a sentence in the record and a rule, review becomes checking specific claims instead of rebuilding the coding from scratch.
How is Chart.ai different from a claim scrubber?
A scrubber checks a finished claim against general rules. Chart.ai works earlier, reading the patient's full history and checking it against specialty and payer rules before the claim exists. It also surfaces clinical findings and documentation gaps, each linked to its evidence, rather than only checking claim formatting and edits.
What should I ask when evaluating AI coding software?
Ask whether it shows the evidence behind every code, whether it reads beyond a single visit, how it flags documentation gaps, who writes and reviews its rules, how it handles payer and code set updates, how easy it is for a reviewer to disagree, and how accuracy is measured. The best test is your own charts, judged against your own coders.
Related Article
How Playbooks Encode Payer and Specialty Logic