Medical Billing AR Follow-Up: Prioritise Aged Claims With AI

Export the aged AR report with identifiers stripped, bucket claims by days outstanding, score each by balance, age and payer deadline risk, and work the highest scores first. AI can build the scoring sheet, rank the queue and draft payer call scripts and notes, while a person makes every call and updates every claim.

Sorting by age alone is the usual mistake. It sends your best collector to a 200-day $18 balance while a $1,400 claim at 75 days drifts towards the payer's filing or appeal limit. The point of prioritising is that your hours are finite and claims are not equal. The setting here is a medical billing team chasing payers on behalf of practices, and the method treats the data privacy question as a design constraint from the start, not an afterthought.

Follow me on Instagram@sagnikteaches

Strip the report before AI sees it

An AR export often carries patient names, dates of birth, member IDs and diagnosis codes. You do not need any of that to decide what to work next. Build the habit of exporting the minimum: claim ID or an internal reference, practice, payer, date of service, date billed, billed amount, expected amount if you have it, balance, last action date and status. If you must match results back to patients, do it in your own billing system using the claim ID.

Connect on LinkedInSagnik Bhattacharya

Whether any AI tool may see patient information is a contract and plan question, not a prompt trick. Business plans of the main assistants differ, and vendors sign a business associate agreement (BAA), the contract under which a service provider may handle a healthcare client's patient information, on particular products only. Anthropic, for example, offers one for its API and for Claude Enterprise once the account owner accepts it in the plan's settings, with some features excluded, and not for its consumer plans; OpenAI's help centre describes a self-serve BAA for its API that needs an established history of API usage. The tutorials on which Claude plans offer a BAA and ChatGPT plans for small practices set out what to check. Until you have that in writing, and your client practices agree, treat consumer tools as for de-identified data only, and never paste a patient name, date of birth or member ID into one.

Subscribe on YouTube@sagnikteaches

Ageing buckets, and what each one means for action

Buckets are the common language of AR, and each implies a different job. Agree with your team whether you measure age from the date of service or the date billed, and use the same on every client report.

Turnaround times and filing limits differ by payer and contract, so keep a payer reference table in your own documents with the timely-filing and appeal limits from each contract or provider manual. AI cannot supply these, and a plausible wrong limit is dangerous. It can, however, help you structure the table and spot gaps.

A scoring sheet that decides the queue

A score turns judgement into a number your team can sort. Keep it simple enough to explain in a minute. Below is a version I would start with, where each claim receives points on three factors and the total ranks the queue.

Sort by total, then by balance. Note that "over 120 days" scores 2 in the age factor, not 4. That is deliberate: very old claims are less likely to pay, so age alone should not pull them to the top unless a deadline or a high balance does. Every number here is a starting point; adjust after a month by looking at which scores actually paid.

Illustrative worked example: a billing team handles four small practices with 400 open claims above 30 days. The old method sorted by age and gave the team 120 claims over 90 days first. The score puts 34 claims at 11 points or above, of which 12 have a deadline within 30 days. Those 12 hold $9,600 of balances. The team works those 12 on Monday, the next 22 by Wednesday, and sweeps the rest by lower score for the rest of the week.

Building the sheet and the ranking with AI

You do not need software beyond Excel or Google Sheets. Use AI to build the formulas, not to hold the data. Describe the columns and the rules, and ask for formulas you can paste. The best AI tools for Excel compares the assistants that can write and explain these formulas; whichever you use, describe every column by letter so the formulas come back right.

I have a sheet with columns: A claim ID, B practice, C payer, D date of service, E date billed, F balance, G last action date, H payer deadline date, I fixable (Y/N).
Write Excel formulas for new columns:
J age in days from E to TODAY()
K age bucket label (0-30, 31-60, 61-90, 91-120, 120+)
L balance points: 1 under 100, 2 up to 300, 3 up to 800, 4 up to 2000, 5 above
M age points: 31-60 = 1, 61-90 = 3, 91-120 = 4, over 120 = 2, otherwise 0
N deadline points: 5 if H minus TODAY() is 30 or less, 3 if 60 or less, otherwise 0
O fixable points: 2 if I is Y
P total score = L+M+N+O
Give one formula per column, explain any assumption, and tell me what to do if H is blank.

Test the result on ten rows by hand before trusting it. A frequent slip is a date stored as text, which makes the age formula return an error or a nonsense number; another is the blank deadline, which should default to a documented rule, not to zero silently. If the sheet already exists in some form in your practice management system, ask for a column mapping instead of a rebuild.

A scored queue, filled in

To see the sheet working, here are five rows from an illustrative export, already scored. Identifiers are claim references only, and every figure is invented for the example.

Notice what the sort does. C-0610 is the oldest by far and ranks last, because it is small and has no deadline pressure. C-1041 and C-0933 tie at the top: one is large and mid-aged, the other is moderate but close to its limit. A pure age sort would have started with C-0610. That is the whole argument for scoring, in one table.

The payer reference table that feeds the deadline column

Your deadline points depend on a table you maintain by hand from each payer's contract or provider manual. Columns: payer, contract date, timely-filing limit, corrected-claim limit, appeal limit and where each came from. Mark each limit with the document and the date you read it, and a reviewer's initials. The AI can help you format and sort the table, and can check your sheet for blank cells, but it must not fill in a limit from memory.

When a limit changes in a new contract year, the dates in the AR sheet need recalculating, so build the deadline column as a formula that looks up the payer in this table. Test it by changing one payer's limit and watching the scores move. If nothing moves, the lookup is broken.

Deciding when a balance is not worth another call

Small balances need a rule, or they absorb a team's time. Illustrative: if the expected recovery is under the cost of the work to get it, move the claim to a write-off review instead of the queue. Cost of a call might be roughly twenty minutes of a collector's time plus follow-up; if a collector's loaded cost is $30 an hour, a single call is about $10, and three calls on a $40 claim is not a good trade. Write your own threshold per client, since contracts differ on what the practice allows you to adjust without approval, and keep a written record of every write-off request and its approver.

AI can list the claims under the threshold that have had two attempts, ready for a manager's review. It should never mark anything written off; that is a decision with a client attached.

Inheriting a backlog from a new client

When a new practice arrives with 900 claims aged over 90 days, the instinct is to start at the top of the list. Resist it. Triage first: score the whole backlog, then split it into claims still within a payer's limit, claims likely past it, and claims where the status is unknown. Work the first group immediately, check the second for exceptions (a prior submission proof or a payer error can sometimes reopen the window, according to the payer's own rules), and sample the third by calling about a few to learn what is happening. Report the findings to the practice in a one-page summary with counts and dollar totals per group, so expectations are set before the first call. The onboarding steps around that are in the client onboarding tutorial.

Using AI to spot patterns in a de-identified report

Once the export carries only references and amounts, AI can show patterns a busy team misses. Ask it to group by payer and status, and to point at clusters.

Here is a de-identified list of aged claims with columns: payer, practice, age in days, balance, status text. Group by payer and then by status, and tell me (1) which payer has the highest total balance over 90 days, (2) the three most common status phrases, (3) any practice where more than a third of its balance sits over 90 days. Give counts and totals. Do not guess reasons.

An illustrative answer might say that one payer holds 41 per cent of the over-90 balance, that "pending additional information" is the most common status, and that one practice has half its balance past 90 days. The reasons are yours to investigate; the AI has told you where to look. If "pending additional information" appears often, a pre-submission check may be missing a document, and fixing it at source removes claims from future reports. Check the totals against the sheet with a SUM before you act on them, because models do sums unreliably.

The daily queue: a rhythm that survives interruptions

A ranked list is not a plan until it fits into a day. Split the queue into three blocks. First, a 45-minute block for the highest-scoring 8 to 10 claims, which are mostly calls. Second, a block for online status checks on the middle band. Third, an end-of-day 15 minutes to update notes and the next-action dates. Each claim touched gets a next-action date, so it drops out of the queue until that date.

Without next-action dates, the same claims re-enter the ranking every morning and the team re-handles them. Add a column "next action date" and filter out any claim whose date is in the future. Suppose a claim was called on Monday and the payer said to allow 10 working days; the next action is the date after that, and nobody touches it in between unless something changes.

Payer-specific call scripts

A call is cheapest when the caller knows the questions before dialling. Build a script per payer, because each has its own menu, documents and phrases. Keep it in your own library and update it after each call. AI can draft the structure from your notes; it should never invent a payer's phone menu, department names or policy.

Write a call script for a billing specialist calling a payer about an unpaid claim, using only the information below.
Payer: [payer name]. Claim reference: [claim ID]. Date billed: [date]. Balance: [amount]. Last action: [text].
Include: opening (who I am, which practice I represent, tax ID verification placeholder), the three questions to ask (claim received? current status and any pending request? next step and date), what to write down (call reference number, representative first name, date and time, promised action and date), and a closing line.
Leave bracketed placeholders for any information I have not supplied. Do not invent payer procedures.

An illustrative script for a claim stuck at day 70 would have the caller confirm receipt, ask whether the claim is pending, denied or in review, and ask for the specific missing item if one exists. A good version ends with: "Can you confirm what you will do and by what date, and give me a reference number for this call?" That last line creates the next-action date on your side and a record on theirs.

Variants by situation

• Claim not on file: ask for the date the payer's system received anything, the submission route used, and whether it was rejected at the front end. Then check your clearinghouse report before resending.
• Pending review: ask what is being reviewed, what documentation they want, and where to send it.
• Underpaid: ask for the allowed amount and the remark that explains the difference, then compare with your contract rate.

Notes that the next person can use

Poor notes ("called, will call back") cost more than a missed call because the next person starts again. Give the team a note template and have AI turn rough call jottings into it, with identifiers removed. Fields: date and time, payer, representative first name, reference number, what was asked, what was said, outcome, next action and date, and anything to attach.

Turn these rough call notes into the standard note format. Use only what is written; if a field is missing, write "NOT RECORDED". Do not include patient names.
Notes: "spoke to [rep name] ref 88213, claim in review, they want op report, send by fax or portal, allow 14 days, will resend monday"

Illustrative result: Date and time: [fill]. Representative: [rep name]. Reference: 88213. Status: in review. Requested: operative report. Send via: portal or fax. Expected: up to 14 days. Next action: resend documents on Monday; recheck after 14 days. The "NOT RECORDED" rule shows exactly what was forgotten, which is useful for training. Copy the note into your billing system, not into a chat window.

A week of the routine, with numbers

Illustrative: a team of three has 400 open claims over 30 days, a balance of $210,000 and 14 per cent of it over 90 days. On Monday the lead exports the report, strips identifiers, runs the scoring sheet and assigns the top 34 claims across the three people. By Friday the team has worked those 34 high-score claims and 70 mid-score status checks, and logged next-action dates. In week two the lead compares what was collected from the high-score group against the rest. If the top group yields more cash per hour, the scoring works; if not, the deadline factor or the fixable factor needs retuning.

The measures to keep are few: cash collected per hour worked, the percentage of AR over 90 days, the number of claims touched within seven days of becoming high priority, and the number of claims written off after the final attempt. Report them monthly to the client practices; the tutorial on a monthly dashboard for a medical practice shows how.

Five ways a scored AR queue misleads you

The first is a de-identification gap: a notes column or a free-text field that still holds a name or a phone number, copied into a prompt. Run a search on the exported file for common patterns, such as digits in a date-of-birth format, before the first paste, and keep the export template locked. The second is the score trusting a wrong deadline: if the payer reference table is out of date, the sheet puts the wrong claims first. Review that table each quarter. The third is a team that treats the ranking as an order instead of a guide. A senior biller sometimes knows a payer is about to change a process; allow an override column with a reason, and review the overrides monthly.

A fourth is the denial that sits in the aged report. Some "unpaid" claims are denied and need an appeal or correction, not a status call; route those to denial management. And a fifth is chasing balances that belong to the patient: once the payer has processed the claim, the remaining amount is a patient-statement matter, handled under your client's policy, not a payer call.

If you would like to map your aged report to a working queue and decide what can safely be de-identified, that is what my AI implementation consultation covers. The first useful step is a scoring sheet tested on one practice's data, with identifiers removed.

AR follow-up questions billing teams ask about AI

Can I paste my AR report into ChatGPT?

Not with patient names, dates of birth or identifiers unless you are on a plan and agreement that covers that data. The safer routine is to export only claim ID, payer, dates, amounts and status, then match results back to patients in your own system. Check your client contracts first.

What are the standard ageing buckets?

Most reports use 0 to 30, 31 to 60, 61 to 90, 91 to 120 and over 120 days from the date of service or claim submission. Check which date your system uses, because a bucket measured from service date looks older than one from submission date.

Should the oldest claims always be worked first?

Not always. A small old balance may cost more to chase than it returns, while a mid-aged high-value claim approaching a payer's deadline may matter more. A score that combines balance, age and deadline risk beats sorting by age alone.

Can AI call the payer for me?

AI can prepare the call: a script, the questions to ask and a notes template. A person still makes the call, records the reference number and rep name, and updates the claim. If a vendor offers automated payer calls, check your client contracts and each payer's rules before considering it.

How do I know the prioritisation works?

Compare cash collected per work hour, the share of claims touched within their target window and the over-90-day percentage before and after, over at least two months. If cash per hour does not rise, change the weights rather than adding more tools.

Further reads

• Medical Billing Proposal: Draft Practice Quotes With AI — Quote a practice for AR work before you start.
• Best AI Tool for Medical Billing: Scrubbing, Coding, Denials — Which AI tools suit scrubbing, coding and denials.
• Optometry Insurance Verification: A Front-Desk Workflow With AI — Verify coverage up front so fewer claims age.
• Can ChatGPT Read PDFs, Spreadsheets and Photos? What Breaks — What breaks when AI reads exports and scans.
• Physical Therapy Billing: AI Checks for Claims and Time Entries — Pre-submission checks that keep claims out of the aged report.
• Medical Practice Credentialing: Track Payer Renewals With AI — Track payer enrolment so claims are not held up.
• AI Accounts Receivable: Predict Late Payers and Act Early — The same prioritising logic applied to ordinary invoices.
• AI Claim Scrubber: What It Catches and What It Misses — Modifier, bundling and diagnosis edits are caught; eligibility, documentation and fresh payer rules often are not. Test one on 50 claims.
• ChatGPT for Medical Billing: Appeals and Claim Checks Without PHI — A redaction method, appeal and claim-check prompts, and the review step that lets a billing team use ChatGPT without pasting patient data.
• Prior Authorisation With AI: Drafting Letters and Tracking Status — A payer checklist, a redaction habit, a letter prompt and a status tracker speed up prior authorisation without handing patient data to a chatbot.
• 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 Tools for Paediatricians' Offices: Seven Admin Jobs — Seven front-office jobs for AI in a paediatric practice, from phones to rotas, each with an example and a rule for using it safely.
• AI for ABA Therapy Practices: Seven Admin Jobs and Three to Keep Human — Seven admin jobs an ABA practice can give to AI, three that stay with the BCBA, and the business associate agreement check before any client data moves.
• 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: Anthropic privacy centre article on business associate agreements, OpenAI help article on BAAs for the API (both checked 4 November 2026); ageing buckets follow common accounts-receivable reporting practice.