Can AI Do Medical Billing and Coding? What Still Needs a Human

AI can do much of the typing and checking in medical billing and coding: draft claims, scrub them for missing data, and suggest codes from a note. It cannot own the final code. A trained person must confirm each code against the clinical record, the payer's rules and the practice's policy before submission.

The useful question for an owner is not "can it?" but "which steps may run unattended and which may not?". A code suggestion is a draft. A submitted claim is a statement the practice signs. Between the two sits a human check, and the work is making that check specific, quick and recorded, rather than a vague promise to "have someone look".

Follow me on Instagram@sagnikteaches

Suggestion versus final code

Most billing AI works in one of three ways. A scrubber checks a finished claim against rules: missing fields, invalid combinations, mismatched dates. A suggester reads a clinical note and proposes diagnosis and procedure codes with the evidence it relied on. An autonomous coder assigns codes and submits without review. The first two are drafts a person can verify. The third moves the decision, and the liability, away from anyone who can explain it.

Connect on LinkedInSagnik Bhattacharya

For a practice owner the safe default is simple. Let AI scrub freely, let it suggest with a person confirming, and allow autonomous submission only for narrow, low-risk, repeat categories after you have audited its output for a period you choose. Even then, audit a sample every month. Best AI tool for medical billing: scrubbing, coding, denials explains which products sit in which group, and AI medical coding software: audit trails and upcoding risk covers what the audit trail should contain.

Subscribe on YouTube@sagnikteaches

For an owner planning staffing, the coder's role changes shape rather than disappearing: less keying, more checking, and a larger need to explain why a code was chosen. Budget for that review time when you price a tool.

The checks, task by task

A billing process has about eight steps. Each can use AI differently, and each has a human check of its own. The table sets out the split; the sections after it take the riskiest tasks in turn.

Registration and eligibility: confirm who and what is covered

AI is good at spotting that a date of birth is missing or a policy number has the wrong length. It is not a source of truth about coverage. The check is to verify eligibility with the payer's portal or the card on the day, and to record the date and the person who checked.

A realistic failure: a tool filled the subscriber field with the patient's name when the policy belonged to a parent. The claim passed the scrubber because the field was no longer empty, then bounced for a subscriber mismatch three weeks later. The fix was a rule that the subscriber field is checked by a person whenever the patient is a dependant.

Charge capture: what was done, not what was written

An AI that reads the note will list the services it finds. If the clinician wrote a template that always says "comprehensive examination", the list will say so every time. The check is for the clinician, not billing staff, to confirm the services actually delivered. Where the note and the memory of the visit disagree, the note gets corrected first, then the claim.

This is also the point where upcoding risk enters. Upcoding means billing a service at a higher level or complexity than was provided or documented. AI does not intend it, but a tool that learns "which codes get paid" from past data can drift towards higher-value codes. The safeguard is a rule, written down, that codes follow documentation and never the other way round. If a suggested code needs details the note does not contain, the answer is a documentation query to the clinician, not a stronger-sounding code.

Code selection: the check that cannot be skipped

A suggester should show its evidence: the sentence in the note that supports each code. The reviewer reads the evidence, not just the code. A workable review has five questions.

1. Does the note, as written, support this code at this level?
2. Is there a more specific code the note also supports?
3. Is any suggested code supported only by a template line or copied text?
4. Are diagnosis codes linked to the procedures they justify?
5. Would an auditor reading only the note reach the same code?

Here is a filled-in example, illustrative. A suggester reads the note for a follow-up visit for a patient with type 2 diabetes and proposes a mid-level visit code plus a separate counselling code. The reviewer finds that the note records one blood glucose review and a medication adjustment, with no separate counselling time recorded. The counselling code is rejected, and a documentation query goes to the clinician: "Counselling time not recorded; please add if it was provided separately." The AI saved ten minutes of reading; the human saved a possible overpayment.

A prompt for a suggester that shows its work

If you use a tool that accepts instructions, ask for evidence and uncertainty, not just codes. The example below is for a de-identified note in an approved tool.

You are assisting a qualified medical coder. From the de-identified visit note below,
suggest diagnosis and procedure codes. For each code give: the code, the exact phrase
from the note that supports it, and your confidence (high, medium, low). List any
code you considered but rejected and why. List information missing from the note that
would change the code. Do not assume anything not written in the note. Do not
choose a higher-level code because it pays more.

A typical response, illustrative, will give a table with codes, quoted phrases and a short "missing information" list. What you fix: remove any code whose quoted phrase is not in the note, downgrade anything marked low confidence to a query, and never accept a code the model cannot quote evidence for.

Modifiers and bundling: rules that live with payers

Modifiers say that a service was altered or done with another, and bundling rules say which services a payer will not pay separately. These rules differ by payer and change, which is exactly where a general AI model is least reliable. It may confidently state a rule that was true last year or applies to a different payer.

The check: a coder verifies every modifier against the payer's current policy document, and your claim scrubber's rules are updated by a named person on a schedule. When the AI says "modifier 25 applies", the coder asks "which payer's policy says so?" If nobody can point to it, the modifier is not applied until someone can.

Scrubbing: check what it did not flag

Scrubbers earn their keep, and they are the lowest-risk AI in billing because they stop errors rather than create them. The human check has two parts. First, review every flagged claim, because flags are where the money is. Second, sample unflagged claims, say 20 a week chosen at random, to see what slipped through.

Illustrative sum: a practice submits 400 claims a week and 6% are rejected on first pass, so 24 claims need rework at about 12 minutes each, which is about 4.8 hours a week. A scrubber that catches half of them early saves about 2.4 hours a week, perhaps 125 hours a year. The weekly sample of 20 claims takes about 40 minutes, so keep it; the saving is real but not free.

Denials and appeals: AI drafts, a person attests

Grouping denials by reason is where AI shines, because it can read hundreds of remittance codes and tell you that 40% of denials this quarter were missing prior authorisation numbers. Drafting an appeal letter is also a strong use. But an appeal makes factual claims about the care provided, so the person who signs must check each one against the record.

The check: the appeal letter quotes only facts from the record; dates, codes and amounts match the claim; and the payer's appeal deadline is on the calendar. See AI denial management for medical billing for the workflow, including how to group and track outcomes.

Patient statements: accuracy over elegance

AI can rewrite a confusing statement into plain English and draft the wording for a payment plan offer. It must not calculate or restate balances on its own, because a small arithmetic slip in a message to a patient causes complaints. The check is that figures come from the billing system and are pasted into the template, and that someone approves payment-plan terms before they are offered. Patient statements and payment plan wording has sample text.

How to run the human check without slowing everything down

A check that takes too long gets skipped. Make it fast by being specific, and by spending effort where errors cost most.

Tier the claims. Say green claims are routine, high-volume and low-value, with a clean scrubber result; amber are anything with modifiers, new codes or a payer you rarely bill; red are high-value procedures, appeals and anything involving a clinician query. Green gets a weekly sample, amber a coder's review, red a coder and the clinician.

Record the decision. For each reviewed claim, store the AI suggestion, the final code and the initials of the reviewer. When a payer queries a claim two years later, you can show the person made the decision.

Track overrides. The share of AI suggestions the coder changes is your quality signal. If a suggester is overridden 30% of the time, it is a poor fit for your specialty or your notes. If it is overridden 1% of the time, either it is excellent or the reviewer has stopped reading. Test the second possibility by seeding a few deliberately wrong suggestions in a training batch and seeing whether they are caught.

A month of numbers for a five-clinician practice

Say a five-clinician practice bills 1,800 claims a month. It introduces a scrubber in week one and a code suggester in week three, reviewing as follows. All numbers are illustrative.

Notice the last row. Documentation queries went up, and that is a good sign, not a bad one: the tool is exposing thin notes that used to be coded on assumption. Notice also that coding-related denials barely moved. A tool cannot fix a coding problem that starts in the documentation. It took the second month of working on templates with the clinicians to bring that number down.

A seven-day test before you trust a coding tool

Before any suggester touches live claims, run it in shadow mode: it suggests, your coder codes independently, and nobody submits from the AI's version. After a week, compare line by line. The comparison tells you more than any vendor demo.

Take 100 closed claims from the last quarter, de-identified, with known outcomes, and include a few categories you worry about: a visit with two problems addressed, a procedure with a modifier, a note that was copied forward, and a note that is simply thin. Score each suggestion as matching the coder, better than the coder (a missed specific code), or worse (unsupported or too high). The cases where the tool is worse matter most, so read each of those aloud with the coder and note why.

Add a deliberate trap. Insert five notes where the documentation does not support the usual code, for instance a long template with one line of actual assessment. A tool that codes the template rather than the assessment will show it here. If it codes high on three of the five, it fails the test, however good its average looks.

Illustrative result: of 100 claims, the tool matched the coder on 71, found a missed specific code on 9, and was wrong on 20, of which 6 were too high. Six upward errors in 100 is a pattern to understand, not a rounding error. Ask the vendor whether the model was tuned on payment outcomes, and whether you can set it to prefer the lower code when the evidence is ambiguous. If the vendor cannot answer, the tool is not ready for your claims.

How the checks differ by practice type

The table above applies everywhere, but the weak points differ, and a practice should weight its review accordingly.

A primary care practice bills many visit levels, and the risk sits in time-based and complexity-based choices. The reviewer spends most effort on whether the note supports the visit level, and on whether several problems addressed in one visit were documented separately. AI is useful for flagging missing chronic-condition detail, as a documentation prompt for the doctor.

A physiotherapy or therapy clinic bills timed services, so the check is minutes. If the note says 38 minutes of treatment and the claim has three timed units, the reviewer confirms the arithmetic against the payer's unit rules. AI can compute units from a time log, but the person confirms the rule that applies.

A specialist procedure practice carries higher values, more modifiers and more prior authorisations. Here the human check is heaviest: every procedure claim is red-tier, and the authorisation number is matched to the claim by a person. The AI's best use is a reminder system that flags a procedure booked without an authorisation on file.

A self-pay or cash practice has little payer coding, so the AI question is mostly about statements, receipts and payment plans, where figures come from the billing system and a person approves any change to terms.

What a good documentation query looks like

When a suggester exposes a gap, the coder writes to the clinician. A good query is short, neutral and does not lead the clinician towards a code. Compare two versions, both illustrative.

Weak: "Please document that this was a moderate complexity visit so we can bill the higher code." That steers the clinician towards a payment outcome, which is exactly the upcoding pattern an auditor looks for.

Better: "The note records a blood pressure review and a medication change. Were any other problems assessed at this visit? If so, please document the assessment." That asks for facts and leaves the code to follow. Keep the query, the reply and the final code together in the record. If the clinician says nothing else was done, the code stays where the note supports it.

Where it goes wrong, in practice

Copied-forward notes. A suggester reads a note that was copied from last month and codes a service that was not repeated. The check is to compare note text between visits for identical paragraphs.

Rule drift. The tool's payer rules age. Put a quarterly review on the calendar and have the vendor confirm in writing when its rules last updated.

Pasting into the wrong tool. A staff member pastes a note with a name into a consumer chatbot to "get a quick code". The fix is a written rule, a safe tool that is just as easy to use, and a monthly reminder. Tools that handle patient data need a proper agreement; what to ask before buying any AI tool for a medical practice lists the questions to put to a vendor.

Over-trusting confidence scores. A "92% confident" label is not a measured accuracy figure for your practice. Ask the vendor how it is calculated, and measure your own override rate instead.

A short policy you can adopt this week

AI in billing and coding: practice rules
1. AI may draft, scrub, suggest and group. It may not choose a final code.
2. A named coder confirms every code against the clinical note before submission.
3. Codes follow documentation. If the note does not support a code, query the
   clinician; never choose a code to fit the payment.
4. No patient identifiers go into any tool without a signed data agreement.
5. Appeals are signed by a person who has checked every fact against the record.
6. We store the AI suggestion, the final code and the reviewer's initials.
7. Monthly: review the override rate, sample 20 unflagged claims, check the rule
   update date with the vendor.

That one page answers most of what an auditor, an insurer or a new hire will ask. If you would like help adapting it to your claim flow and choosing where a tool actually fits, that is the kind of work my AI implementation consultation covers.

Owner questions on AI and billing

Who is responsible if an AI-suggested code is wrong?

The practice that submits the claim. A vendor's tool does not take over your liability, so read what your contract says about errors. That is why a named person confirms codes against the clinical note before submission, and why you keep a record of who approved what.

Can I let AI submit claims without a person looking?

For clean, repeat, low-value claims some practices allow automatic submission after the practice's own rules pass, and audit a sample afterwards. For anything involving code choice, modifiers or high-value procedures, a person approves first.

Will AI coding increase my denials?

It can if it codes from thin notes, or reads the same note differently each time. Track denial rate by reason for three months before and after. If coding denials rise, tighten the human check rather than blaming staff.

Is it safe to paste a patient's note into a chatbot to get a code?

Not with identifiers, and not on a consumer plan. Use a tool that has a signed data agreement, or de-identify the text first. Even then, treat the answer as a suggestion to be checked against the code set and payer policy.

Further reads

• A Worked AI Implementation Plan for a Small Medical Practice — A worked rollout order for a small medical practice.
• Patient Data and AI: A Confidentiality Checklist for Small Practices — Confidentiality checks before billing data reaches an AI tool.
• Medical Practice AI Budget: What to Fund First and What It Costs — Where billing support sits among other AI spending.
• AI Platform for Doctors: One Suite or Separate Scribe and Phone Tools? — Whether a billing add-on or a standalone tool fits.
• Best AI for Medical Practices: Eight Tool Types and Who Each Suits — Billing support compared with the other seven tool types.
• Medical Billing Proposal: Draft Practice Quotes With AI — Draft a medical billing proposal with AI: percentage vs flat-fee pricing, a scope table, an onboarding timeline and a fully worked quote.
• Medicare Codes for Billing Psychology in Australia: AI Claim Checks — A pre-submission claim check for Australian psychologists: match MBS item numbers to times, referrals and session counts, flag mismatches and never upcode.
• Use AI to Write a Care Plan Review From Daily Notes, Safely — Draft a care plan review summary from a month of daily notes without exposing identities, then trace every line back before sign-off.
• Using AI for Medical Billing: Spot Missed Charges From Visit Notes — A weekly reconciliation that finds delivered-but-unbilled services, with the spreadsheet step, the AI step and the owner's review.
• AI for Medical Billing Companies: Seven Jobs to Automate First — Seven jobs for a medical billing company to automate first, ranked by risk, with sample prompts, a worked denial-letter example and the human checks each needs.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: a national payer's fraud-and-abuse guide (April 2026 edition) for the upcoding definition and its correct-coding FAQ library for modifier 25 and bundling edits, both read 4 November 2026; the review steps are general billing-compliance practice, and no accuracy statistics are quoted because none were verified.