AI for Dental Billing: Claim Checks, Statements and Follow-Up

A dental practice can use AI for three billing jobs: checking claims for missing information before submission, wording patient statements clearly, and working an aged-claims queue. A person must choose and approve every procedure code. AI finds gaps and drafts text; it never decides what is billed, so it cannot become a route to upcoding.

The savings are real and modest. Many denials trace to boring causes: a missing tooth number, an absent radiograph reference, a subscriber ID typo, a frequency limit nobody checked. A tool that reads a claim summary against a checklist is good at exactly that kind of boredom. It is poor at judgement, and billing has plenty of that, so the line between the two matters more than the choice of tool.

Follow me on Instagram@sagnikteaches

What to give AI, and what to keep in the practice management system

Your practice management system (PMS) is the source of truth. It holds the clinical record, the fee schedule and the claim. AI sits beside it, working on exports and text, and writes nothing back by itself. That keeps the audit trail where it belongs.

Connect on LinkedInSagnik Bhattacharya

If your software already has an AI feature, learn what it does before adding anything. The pieces on Open Dental add-on tools and Dentrix's AI feature cost show how much of this is already in the system you pay for.

Subscribe on YouTube@sagnikteaches

Handling patient data: the choices that matter

Billing data is protected health information. That changes how you use a chatbot, not whether you can.

• De-identify first. Export the report, then replace patient names, dates of birth, subscriber names and addresses with an account number or row number before pasting. Keep the key in the PMS. Tooth numbers, procedure codes, payer names, dates of service and amounts are often fine to leave, but your compliance contact decides.
• Use a plan with the right terms. Free and individual plans are not designed for patient data. ChatGPT Plus is $20 a month and Claude Pro $20 monthly or $17 a month billed annually, but both are individual plans. ChatGPT Business and Claude Team need at least two seats and do not train on business content by default, yet neither comes with a business associate agreement (BAA), the contract a vendor signs before it handles protected patient information. OpenAI's help centre says it does not offer a BAA for ChatGPT Business, only for sales-managed ChatGPT Enterprise or Edu accounts and its API. Anthropic's BAA covers its API and Enterprise plans, and only once the account's primary owner accepts it. Until a BAA is signed, de-identify everything.
• Turn training off on any consumer plan, and delete chats after use.
• Keep records out of chat history. The record of what you billed lives in the PMS, never in a chat window.

The tutorial on dental practice AI mistakes lists consent and over-automation errors worth reading alongside this.

Pre-submission claim checks that catch the boring errors

The best use of AI in billing is a second pair of eyes that never gets tired. Build a checklist from your own denial history, not from a generic list. Pull the last 90 days of denials and group the reasons. A handful of reasons often account for most of them.

A starter checklist (adapt to your payers)

1. Patient and subscriber details match the insurance card on file (name spelling, date of birth, member ID, group number).
2. Rendering and billing provider details are present and match what the payer expects.
3. Each procedure has a date of service, a code, a tooth, surface or quadrant where the code needs one.
4. Required attachments are listed and available: radiographs, periodontal charting, narrative, where your payers ask for them for that code.
5. Frequency and age limits were checked against the patient's benefits (from your verification, not guessed).
6. Prior authorisation is recorded where the plan required it.
7. The fee billed equals your fee schedule for that code.
8. Coordination of benefits is correct: primary and secondary are in the right order.

A prompt for the check

You are a dental billing checker. Below is a de-identified claim summary and my checklist. For each checklist item, answer "OK", "MISSING" or "UNSURE" and quote the line from the summary you relied on. If a field is not in the summary, say MISSING. Do not suggest or change procedure codes. Do not assume anything not written. End with a list of items for a person to resolve.

Checklist: [paste]
Claim summary: [paste]

A sample, with what to fix (illustrative)

Claim summary: "Acct 1182. DOS 12 Oct. D0120, D0274 (four bitewings), D1110. Primary: Plan A, ID present. Secondary: none recorded. Attachments: none listed. Benefits note: bitewings once per 12 months, last billed 3 Nov last year."

The check might return: "Frequency: UNSURE. Last bitewings were billed 3 Nov last year, and DOS is 12 Oct, which is under 12 months. Attachments: MISSING for D0274 if the payer requires radiographs. Secondary: UNSURE, no coverage recorded." That is a good result. It did not rewrite anything. It found that a claim on 12 October falls eleven months and nine days after the last bitewings, which a rushed human might miss. A person now checks the benefit rule and either holds the claim or bills it knowing the risk.

What to correct: a weak prompt invites "this looks fine". If it returns OK without quotes, treat that as a failure and tighten rule one: no quote, no OK.

Claim narratives: draft from the chart, never to justify a code

Narratives explain why a service was needed. Some codes, such as certain periodontal treatments, crowns and build-ups, often need one. AI can turn documented facts into clean prose. It must never make the facts stronger than they are.

Write a claim narrative for a dental insurer from only the clinical facts below, in plain professional language, under 90 words. Do not add findings, measurements, symptoms or history that are not listed. Do not state that treatment was "necessary" unless the dentist's note says so in those words. If any fact needed for this type of narrative is missing, list it under "MISSING" instead of guessing.

Facts: [dentist's documented findings, de-identified]

Sample facts, illustrative: "Tooth 19, large distal carious lesion on radiograph dated 12 Oct, extending into dentine, existing amalgam failing with recurrent decay, patient reports cold sensitivity." Sample output: "Radiograph dated 12 Oct shows a large distal carious lesion on tooth 19 extending into dentine, with recurrent decay under a failing amalgam restoration. The patient reports cold sensitivity. Restoration is planned." What to fix: if the tool wrote "to prevent pulpal involvement", delete it unless the dentist wrote it. The dentist signs the narrative, not the tool.

A fair rule: no clinical sentence leaves the practice that a dentist has not read. If a narrative is routine, build a template with blanks and have the dentist fill the findings. AI then only tidies grammar.

Patient statements that people actually understand

Statements are where billing meets the patient, and unclear ones create calls. The amounts come from the PMS. AI's job is the words around them.

What a good statement says

• What was done, in plain language, with the date.
• What the insurer paid, what is still pending, and what the patient owes.
• Whether the balance is final or may change, and when it will be final.
• How to pay, by when, and who to call.

Rewrite this patient statement message in plain, kind language at a reading age of about 12. Keep every amount, date and account number exactly as given. Do not add amounts, deadlines or consequences that are not in the text. Do not include any threat. Tell the patient what to do next and who to call. Keep it under 120 words.

Original: [paste the template text with placeholders like {amount_due}]

Before, illustrative: "Balance due for DOS 10/12 after ins adj. Pt resp $142.00. Remit within 30 days." After: "Hi {first_name}, thanks for visiting us on 12 October. Your insurance has paid its part. The remaining amount is {amount_due}. We would be grateful if you could pay by {due_date}. You can pay online at {link} or call us on {phone} if you would like to talk through it." Note what is deliberately absent: no "overdue", no "collections", no interest. If your practice's policy includes late fees, those words come from your written policy, not from AI.

The tutorial on explaining payment plans to dental patients pairs well, because a clear statement can include one sentence about what you offer. If balances often arise from missed appointments, make the statement wording match your written no-show policy so the two documents say the same thing.

An aged-claims queue worked once a week

An aged claim is one submitted and still unpaid. Most PMS tools can export an ageing report or claims list to CSV or Excel, though the name differs by system. Pick the columns once and keep them stable: account number, payer, date of service, date submitted, amount, claim status, last action date, next action.

Worked example with numbers (illustrative)

Say a three-dentist practice has 46 open claims over 30 days old, totalling $38,400. A billing coordinator exports the list, replaces names with account numbers, and asks AI to bucket and prioritise it. The result:

A single prompt produces this from the export, and the coordinator's first hour goes on the six oldest claims, because their filing deadlines are the ones that expire. The AI adds no money. What it does is make the work orderly, so nothing sits unread for four months. Over 90 days, the exposure is $7,400, and a claim lost to a filing limit is $0 you cannot get back.

A bucketing prompt for the aged-claims export

Below is a de-identified aged-claims list. Sort it into buckets: 31-60 days, 61-90 days, over 90 days, denied. For each bucket give the count, total amount and the five oldest by date submitted. For each claim list a suggested next action from: check portal, phone payer, resubmit, correct and resubmit, appeal, write off for review. Do not suggest codes or clinical changes. Flag any claim where date of service is more than 9 months old as "CHECK FILING LIMIT". Output a table.

List: [paste]

Follow-up drafts

Once sorted, AI can draft the message for each situation. Typical templates are a portal message, an email to a payer representative, and a call script. Keep them short and factual: "Claim for DOS {date}, account {acct}, submitted {date}, amount {amount}. Status shows pending. Could you confirm it was received and whether anything further is needed?" The payer's reply matters more than your wording. Log the outcome in the PMS the same day, with the person's name and the reference number from the call.

Appeals and denials

Denials fall into a handful of types: missing information, frequency limit, not a covered service, bundled into another code and lack of medical necessity. The first is a fix, the second and third are often patient conversations, and the last two are appeals.

AI can read a denial text and tell you which type it is, then draft the first paragraph of an appeal from the facts. The dentist or the billing lead reads every line. Three rules keep this clean:

1. Every clinical claim in the appeal appears in the chart, with the date, and the chart extract is attached.
2. The appeal cites the plan's own wording where it helps, copied from the document, not from memory or from the chatbot.
3. The deadline for appeal is read from the denial letter and diarised, not assumed.

Realistic failure, described generically: a draft appeal says "the patient's periodontal condition required treatment" and the chart does not record pocket depths. The payer asks for the charting, which does not exist, and the appeal fails. The fix is to build the appeal from documented measurements only, and if they are missing, to learn from it by adding a charting check to clinical workflow.

Three smaller jobs that fit the same workflow

Reading an explanation of benefits against what you posted

When an insurer pays, the explanation of benefits lists each line with an allowed amount, the insurer's payment and the patient's share. Mismatches between what you expected and what was paid, called short pays, are common and easy to miss. If your clearinghouse or PMS exports payments as a table, give AI the de-identified rows with your expected amounts alongside and ask it to list lines where paid differs from expected by more than $5 or 10 per cent. Illustrative output: "Acct 1182, D1110: expected $62.00, paid $48.00. Difference $14.00. Acct 1190, D0274: expected $58.00, paid $58.00. OK." A person then checks whether the difference is a contracted write-off (fine) or an underpayment (worth a call). The tool is only doing arithmetic and comparison here, which is where it is dependable, though you should spot-check a handful of rows against the originals every time.

Children's accounts and shared responsibility

Statements for children often go to a parent or guardian, and in some families to two. AI does not know who is responsible, and a statement sent to the wrong person discloses a child's treatment and balance. Keep the responsible party as a field in your PMS and make the template pull the name from there. Ask AI to draft the wording only, with a blank for the addressee, and have a person check the addressee on the first statement for any account with split responsibility.

Month-end summary for the owner

A short monthly paragraph for the practice owner is a good use of a tool that never tires of formatting. Give it the de-identified totals you already export (production, collections, adjustments, ageing buckets and denials by reason) and ask for six lines: what moved, what is above 90 days, the top three denial reasons and one action for next month. Illustrative result: "Collections were 96 per cent of net production. Over-90-day claims fell from $9,600 to $7,400. Top denial reasons: missing attachment (7), frequency limit (4), subscriber ID mismatch (3). Next month: add radiograph check to the pre-submission list." The numbers come from your export, so check the sums against the PMS before you share the summary. If the export is messy, the tutorial on cleaning messy data shows how to tidy columns before you ask any tool to read them.

When statement wording meets a real conversation

Statements start phone calls. A short call script, drafted once, keeps those calls calm and consistent: confirm the caller's identity using two details before you discuss anything, explain the line items in plain language, state what insurance paid and why, offer the payment options your practice actually has, and write the outcome into the account notes. AI can draft the script and an FAQ for the front desk ("Why is my balance different from the estimate?"). The usual honest answer is that an estimate is not a guarantee of payment, and that final amounts depend on what the insurer pays. Put that sentence on your treatment estimates and your statements so it never comes as a surprise.

A realistic mistake: a statement template tells every patient "insurance has paid its portion", but some accounts have a pending claim. Patients ring in confused and angry. The cure is to make the template conditional, one version when the claim is paid and another when it is pending, and to have AI draft both, rather than writing one cheerful message for all cases.

Avoiding upcoding and other compliance traps

Upcoding means billing a more expensive code than the service documented. An AI tool does not do it on purpose, but the risks are real:

• A prompt that says "find the highest-paying code for this treatment" is a request for upcoding. Never write one.
• A narrative that overstates findings is a false statement to a payer, even if a person typed the facts.
• A "suggestion" feature in software that picks codes is a tool to review, not a tool to trust.

Write a one-page rule for the team: AI never picks, changes or recommends codes; AI never receives clinical facts that are not already documented; the person who clicks submit reads the claim. When your software adds an AI feature, read its documentation and ask the vendor how it is trained, whether it writes to the record, and what the audit trail shows. The comparison of AI insurance verification tools asks those questions of the vendors that touch benefits.

Rolling it out over three weeks

1. Week 1: pull 90 days of denials, build the checklist, agree the data rules (de-identification, tool, plan, BAA).
2. Week 2: run the pre-submission check on 20 claims alongside your normal process, and compare what it flagged with what staff caught.
3. Week 3: add statement templates and the weekly aged-claims review.

Measure the first-pass acceptance rate, days in accounts receivable and total over 90 days, before and after. Do not expect miracles in a month; a practice that already checks well may see little change, and that is fine. If the check flags nothing that staff had not already caught after 50 claims, stop using it and save the subscription. A practice with a thin front desk and a backlog of statements will see more. If you want help deciding where this earns its keep, my AI implementation consultation covers exactly that, through the contact page.

Questions dental teams ask about AI and billing

Can AI choose procedure codes for me?

It should not. Codes come from the clinical record and the dentist's documentation, and a person confirms them. AI is better used to compare what was documented with what is on the claim and to flag gaps, never to decide what to bill.

Can I upload an ageing report to ChatGPT?

Only once it is de-identified, unless the practice has a signed business associate agreement with the vendor, which OpenAI offers on sales-managed ChatGPT Enterprise accounts but not on ChatGPT Business. Replace names and dates of birth with account numbers before pasting, and ask your compliance contact before any real patient data goes in.

Will AI write appeal letters for denied claims?

It can draft the structure from facts you supply, such as the code, date and denial reason. The dentist or billing lead must confirm every clinical statement and attach the real documentation, because the appeal is a statement to a payer.

How long does it take to set this up?

A simple pre-submission checklist and statement templates can be running in an afternoon. An aged-claims queue that you review weekly takes about a week to settle, mostly in agreeing which columns your system's export includes.

Further reads

• Dental Treatment Plan Follow-Up: A Message Sequence With AI — Message sequence for accepted and unaccepted treatment plans.
• AI Scribe for Dentists: Will It Write Back to Your Software? — Whether an AI scribe will write back to your software.
• Dental Fee Schedule: Set Fees and Explain Them to Patients With AI — Set fees and explain them, so statements surprise nobody.
• Dental New Patient Forms: A Digital Pack and an AI Completeness Check — Capture insurance details correctly at the first visit.
• How to Use AI in a Dental Clinic: A Four-Week Starting Plan — A wider four-week plan for AI across the practice.
• Dental Morning Huddle: A One-Page Daily Brief AI Builds From the Schedule — Surface today's balances and pre-authorisations at the huddle.
• Dental No-Show Policy: Write It, Word It and Send It With AI — Word the no-show policy so statements and policy agree.
• Outsourced Medical Billing or In-House AI Tools? A Decision Rule — A cost-per-claim sum and a four-test rule for deciding between an outsourced billing company and in-house AI claim tools.
• AI Denial Management for Medical Billing: Group, Appeal, Track — Export your remits, group denials by reason code, draft appeals with AI and track outcomes on one sheet, with a worked month in numbers.
• AI in Healthcare Billing: Patient Statements and Payment Plan Wording — Draft statement and payment plan wording with merge fields, reconciliation checks and a privacy-safe review route.
• Dental Lab Invoice Template: AI-Drafted Statements and Chasers — Per-case invoice fields, a monthly dentist statement with ageing, a four-step chaser sequence and a dispute log, drafted with AI and reviewed.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT pricing pages and OpenAI help-centre article on business associate agreements, Claude pricing page and Anthropic privacy centre article on BAAs, a dental payer's published covered-procedures list for CDT code descriptors (all checked 4 November 2026); general dental revenue-cycle practice.