Coding and billing is documentation work at scale, and language models are good at reading documentation. Summarizing a long operative note, pulling out the elements that support an E/M level, drafting a denial appeal that quotes the payer's policy, explaining a remark code you have not seen in a year: those are minutes now, not half-hours.
The hard limits: models do not have the current code set memorized. ICD-10-CM changes every October and again in April, CPT changes every January, and NCCI edits update quarterly. A model trained before those updates will confidently hand you a deleted code, a wrong descriptor, or a modifier pairing that fails an edit. Every code it suggests gets verified in your encoder and the official guidelines, no exceptions.
And the compliance line: you code what the provider documented, not what the model infers. An assistant that 'finds' a diagnosis the note never states, or nudges every visit toward a higher level, is generating False Claims Act exposure with your credential on it. PHI stays inside your organization's approved systems; consumer chatbots get fictional notes only.
Quick wins this week
- Paste a fictional or fully de-identified progress note and ask the model to list the medical decision making elements it finds (problems, data, risk) without assigning a level, then level it yourself and compare.
- Give it a denial with the claim adjustment reason and remark codes plus the payer's published policy, and ask for an appeal letter that quotes the policy's own criteria.
- Ask it to explain a guideline you keep looking up, such as sequencing rules for a chronic condition or when modifier 25 is supported, then check the explanation against the official guidelines and CPT Assistant or Coding Clinic.
- Turn the annual ICD-10-CM addenda for your specialty into a one-page cheat sheet of new, revised, and deleted codes, verified line by line against the official file.
- Build a study plan for the CPC, CCS, or CPB exam from the published exam blueprint and your weak domains.
What AI can do for medical coders and billers, task by task
Reading long documentation faster, inside approved systems
Summaries of operative reports, discharge summaries, and long encounter notes belong only in tools your organization has covered under a Business Associate Agreement: Copilot inside your tenant if IT has enabled it, or AI built into your encoder or EHR. Use the summary as a map to the note, not as the source; code from the documentation itself. Check that the summary did not collapse two procedures into one or drop a laterality or a complication.
E/M leveling support without letting the model pick
Under the guidelines in place since 2021, office visits level on medical decision making or total time, and most other settings followed in 2023. Ask the model to extract the problems addressed, the data reviewed or ordered, and the risk elements, quoting the note for each, and to list documentation gaps as questions. Then apply the MDM table yourself. A model asked for the level will drift toward the higher one, and it cannot see what the provider did not write.
Denials, appeals, and payer correspondence
Paste the claim adjustment reason and remark codes, the payer's published policy or the Medicare coverage determination, and a de-identified clinical summary, and ask for an appeal that addresses each criterion in order. Verify every citation by opening it. Know the deadlines the model does not: timely filing limits vary by payer, and Medicare appeals move through fixed levels with fixed windows. Finish anything with real patient detail inside your billing system.
Compliant provider queries and CDI education
Give the model the de-identified clinical indicators and the documentation gap, and ask for a query in the format the AHIMA and ACDIS compliant query practice guidance describes: non-leading, clinical indicators listed, multiple choice with clinically reasonable options plus 'other' and 'unable to determine', and no mention of reimbursement. Read it for leading language before it goes out; a query that suggests the answer is a compliance finding waiting to happen.
Keeping up with code sets, NCCI, and payer policies
Upload the current ICD-10-CM Official Guidelines, the NCCI policy manual chapters you work in, and the LCDs and payer policies you fight most into NotebookLM and ask questions that must be answered from those documents with the passage shown. For a fast first pass on a payer's public policy, Perplexity finds the document; you read it. Check the effective date on everything, because the model cannot tell you what changed this quarter.
Working alongside computer-assisted and autonomous coding
Computer-assisted coding has been in encoders for years, and autonomous coding vendors (Nym, Fathom, and CodaMetrix among them) now code some encounter types with a human reviewing exceptions. Your role shifts toward auditing: learn the confidence thresholds, sample the auto-coded claims regularly, and watch for systematic drift such as a specialty where every visit lands at the same level. Report patterns to compliance rather than fixing them silently.
Denial analytics and reporting
Export de-identified aggregate data (denials by reason code, payer, provider, and CPT range) and ask Copilot in Excel or Julius to find the concentrations and explain them in plain language for a revenue-cycle meeting. Check the arithmetic and the joins; models still miscount and mislabel columns. No patient rows leave the billing system, and the file you upload should not have names, dates of birth, or account numbers in it.
Certification study and continuing education
Give the model the published exam blueprint for the CPC, CCS, CPB, or a specialty credential, your weak domains, and your schedule, and ask for a plan with spaced review and practice questions with rationales. Verify every rationale against the current code book and guidelines, because practice questions built from stale memory teach stale codes. The certifying body's own practice exams remain the source of truth for what is tested.
Prompts for medical coders and billers
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
MDM element extraction with no level assigned
Act as a coding auditor reviewing evaluation and management documentation. Here is a fictional or fully de-identified note with no real patient data: [PASTE THE NOTE]. List, in three sections, the problems addressed, the data reviewed or ordered, and the risk elements you can find, quoting the exact sentence from the note that supports each item. Do not assign a level and do not assign codes. Do not infer anything the note does not state; if a problem or a data element is implied but not documented, list it separately as a question I could ask the provider. Finish with any documentation gaps that would prevent supporting the level the provider chose: [LEVEL THE PROVIDER SELECTED].
Tip: Level it yourself from the extraction, then compare with the provider's choice. The model is a reading aid, not a leveling tool.
Appeal letter that quotes the payer's policy
You are a billing specialist writing a first-level appeal. Codes billed: [CPT OR HCPCS CODES] with diagnoses [ICD-10-CM CODES]. Denial reason and remark codes: [CARC AND RARC CODES AND TEXT]. Payer policy or coverage determination, pasted from the published source: [PASTE THE POLICY OR LCD TEXT]. De-identified clinical summary: [DE-IDENTIFIED CLINICAL SUMMARY]. Write an appeal under 450 words that addresses each relevant criterion in the policy in order, quoting it and then stating the supporting facts from the summary. Cite nothing outside the policy and the summary. If a criterion is not clearly met, tell me in a separate note. Formal tone, no criticism of the payer, and a closing line that requests the specific action.
Tip: Add real dates, account numbers, and identifiers only inside your billing system. The chatbot never sees the actual claim.
Compliant provider query draft
Act as a clinical documentation specialist who follows the AHIMA and ACDIS compliant query practice guidance. De-identified clinical indicators from the record: [CLINICAL INDICATORS]. The documentation gap: [WHAT IS UNCLEAR OR MISSING]. Write a query that states the clinical indicators, asks an open, non-leading question, and offers clinically reasonable options in multiple-choice form including 'other, please specify' and 'unable to determine'. Do not mention reimbursement, quality scores, or the impact of any answer. Do not suggest a preferred answer. Keep it under 150 words. Then flag any phrase in your own draft that a compliance reviewer might read as leading.
Tip: Read the flagged phrases first. If you would be uncomfortable showing the query to an auditor, rewrite it.
Guideline explainer that cites only what I paste
You are a coding educator. Below is a section of the official guidelines or a payer policy: [PASTE THE GUIDELINE OR POLICY SECTION]. My question: [YOUR QUESTION]. Answer using only the pasted text, quoting the sentence you rely on for each point. If the text does not answer the question, say so plainly. If you add anything from your general knowledge to make the answer coherent, put it under a separate heading 'Unverified, from general knowledge' so I can check it against the current code set and Coding Clinic or CPT Assistant. Do not state any code number that is not in the pasted text.
Tip: The 'Unverified' section is where the stale codes go. Check every item in it before you rely on the answer.
Code update cheat sheet from the official addenda
Act as a coding manager preparing an update briefing for [SPECIALTY]. Here is the official addenda or update summary: [PASTE THE OFFICIAL ADDENDA TEXT]. Build a table with columns for code, descriptor, change type (new, revised, deleted), effective date, and a one-line note on when a coder in this specialty would use it. Include only codes that appear in the pasted text. Do not add codes, descriptors, or guidance from memory. Then list any change in the text that alters a sequencing or coding rule rather than a code, and quote it.
Tip: Verify the table against the official file line by line before it goes to the team. One transposed digit becomes a hundred wrong claims.
Certification study plan
Act as a study coach for the [EXAM, E.G. CPC, CCS, OR CPB] exam. Published exam blueprint: [PASTE THE BLUEPRINT OR DOMAIN WEIGHTS]. My weakest domains: [WEAK DOMAINS]. Exam date: [EXAM DATE]. I can study [HOURS PER WEEK] hours a week around a full-time job. Build a week-by-week plan that weights time by the blueprint and my weak domains, uses spaced review, and includes a timed practice block each week. Then write five practice questions on my weakest domain with rationales, marking any rationale or code you are not certain is current so I can verify it in the code book.
Tip: Use the certifying body's practice exams as the source of truth. The model is for scheduling and warm-up, not for deciding what a code means today.
Want a prompt for something else? Use the Prompt Builder.
Skills to build
Verifying every code against the current set
Why: A model's code knowledge is frozen at its training cutoff, and it fills gaps with plausible inventions: a real-looking code with the wrong descriptor, a deleted code, a modifier pairing that fails NCCI.
How: Treat every code the model mentions as a lead. Confirm it in your encoder, the official guidelines, and Coding Clinic or CPT Assistant before it touches a claim, and keep a running list of the stale ones you catch.
Holding the compliance frame: documentation drives the code
Why: The False Claims Act does not care who suggested the code, and 'the AI recommended it' is not a defense. A model infers; a coder reports what the provider documented.
How: Add 'do not infer anything the note does not state' to every documentation prompt, and route anything the model surfaces that is not documented into a compliant query rather than a code.
Reading for what the model added
Why: The dangerous output is the confident extra: a specificity the note lacks, a 'probable' diagnosis promoted to confirmed, a laterality that was never stated.
How: Compare the extraction to the note rather than judging it on its own. Anything in the output you cannot point to in the documentation gets deleted or queried.
Writing queries that would survive an audit
Why: A query that leads the provider toward a higher-paying diagnosis turns a documentation improvement into a compliance finding, and models default to helpful phrasing that leads.
How: Keep the AHIMA and ACDIS query guidance next to your workstation, make the model flag its own leading phrases, and have a CDI colleague review your first ten AI-drafted queries.
Auditing automated coding instead of trusting it
Why: Computer-assisted and autonomous coding are usually right, which is exactly what makes drift hard to see, and the pattern that gets an organization in trouble is systematic, not random.
How: Learn the tool's confidence thresholds, sample auto-coded claims on a schedule, track error types by specialty and code range, and report patterns to compliance in writing.
PHI discipline in a job that is made of PHI
Why: Every note you touch is protected health information, the HIPAA minimum necessary rule applies to what you paste into a tool as much as to what you print, and de-identification means removing all 18 Safe Harbor identifiers, not only the name.
How: Work only inside tools your organization has approved under a BAA, keep a fictional note set for learning with consumer assistants, and ask compliance before any new tool touches a real record.
Tools worth knowing
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Claude
A careful writing and analysis assistant that shines on long documents.
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
NotebookLM
A research notebook that only answers from the sources you give it, with citations.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
Julius AI
Chat with your spreadsheets and data files and get charts, stats, and answers back
Cautions for medical coders and billers
A clinical note is protected health information by definition, and consumer chatbots carry no Business Associate Agreement. Never paste confidential patient data into a consumer AI tool unless your organization has approved that tool and signed the agreements. Do real work inside approved systems, and use fictional or fully de-identified notes when you want to learn what a general assistant can do.
Code sets change on fixed schedules and a model's knowledge does not. Expect deleted codes, wrong descriptors, and modifier pairings that fail NCCI edits, all delivered confidently. Verify every code in the current encoder and guidelines, paste the official text into prompts instead of asking for recall, and treat any code number the model produces from memory as unverified until you have looked it up.
A model asked for the level tends to pick the higher one, and a model asked what a note supports will find diagnoses that are implied rather than documented. Both create False Claims Act exposure, and risk-adjustment coding faces its own audits. Code what is documented, query what is unclear, and never accept a suggestion you cannot point to in the note.
When a tool is right most of the time, the reviewer stops reading, and that is when a systematic error runs for months. Your credential and your certifying body's code of ethics still attach to every claim you release. Keep sampling, keep documenting what you find, and escalate patterns rather than quietly correcting them.
Automated claim review flags patterns across a practice's claims, and CMS has said that Medicare Advantage plans cannot base a medical necessity denial solely on an algorithm. Write appeals for the human reviewer who eventually reads them: quote the policy, state the facts, cite nothing you cannot open, and keep the tone flat.
Your 30-day plan
- Week 1: Ask compliance which AI tools are approved, which carry a BAA, and what may be pasted into each. Build a small set of fictional notes for practice. Use a general assistant for two non-PHI tasks: a certification study plan and a cheat sheet from the current addenda.
- Week 2: Run five fictional notes through the MDM extraction prompt, level them yourself, and compare with the model's extraction. Note every element it added that the note did not state.
- Week 3: Build two appeals from payer policies using de-identified summaries and verify every citation. Load the official guidelines and your top three payer policies into NotebookLM for a week of lookups.
- Week 4: If your organization uses computer-assisted or autonomous coding, ask for the confidence thresholds and audit a sample of auto-coded claims. Draft two compliant queries and have a CDI colleague review them.
- End of month: Write down the three tasks where AI saved real time, the stale or invented codes you caught, and what you want compliance to know about the tools in use.
Frequently asked questions
Will AI replace medical coders?
Can ChatGPT assign CPT and ICD-10 codes?
Is it a HIPAA violation to paste a patient note into ChatGPT?
How accurate is AI medical coding?
Should I learn AI if I am a coder or biller?
Terms used on this page
Related roles
- AI for Healthcare AdministratorsAdministrators drown in documents: policies, payer correspondence, board decks, survey prep, budget narratives. AI drafts and summarizes all of it, if you keep PHI inside covered systems and set the rules for your organization before someone else sets them for you.
- AI for PhysiciansThe documentation and inbox load that follows you home is where AI pays off first. It drafts; you decide, verify, and sign, and nothing with PHI leaves the systems your organization has covered under a Business Associate Agreement.
- AI for BookkeepersAI writes the collections email, drafts the categorization rules, builds the month-end checklist, and explains the P&L to an owner who has never read one. You still decide what the transaction actually was and whether the books are ready to close.
- AI for Compliance OfficersAI rewrites the policy nobody reads into one people will, finds the rule change you missed, drafts the training deck, and structures the investigation chronology. It also puts a new line on your risk register: the company's own use of AI, which regulators now expect you to govern.