Medicare Codes for Billing Psychology in Australia: AI Claim Checks

Yes, AI can check psychology Medicare claims for errors before submission, if it works on structured fields (item number, date, start and end time, referral details, session count) matched against the current MBS descriptor you supply. It flags mismatches for you to review. It should never see session notes, choose an item for you or push you towards a higher-value item.

Search results for this topic are mostly CPT-code pages written for a different system. For a psychologist billing Medicare here, the reference is the Medicare Benefits Schedule (MBS) and its item numbers, and the errors that cost money are specific: an item that does not match the time recorded, a referral that has lapsed, a provider-type mismatch. The check below is built around those errors, and the clinician makes every decision. It never replaces the schedule itself, which you verify at the time of claiming.

Follow me on Instagram@sagnikteaches

What AI should and should not touch in a psychology claim

A psychology claim has two kinds of information. Structured fields (dates, times, item numbers, referral details) can be checked by a rule. Clinical content (what was discussed, formulation, risk) must never leave the clinical record for a billing check. The whole design below rests on keeping those apart.

Connect on LinkedInSagnik Bhattacharya

A client code such as C017 is enough. The check never needs to know who the client is, only that C017 had her ninth session of the year on a given date. If your practice management system exports a claim list, delete identity columns in the spreadsheet before doing anything else with it.

Subscribe on YouTube@sagnikteaches

The item families, and the one rule about numbers in this tutorial

Under Better Access, face-to-face psychological services sit in two item groups on MBS Online. Group M6 (Psychological Therapy Services) is for eligible clinical psychologists, with items in the 80000 range: on the schedule checked for this tutorial (4 November 2026), item 80000 is a consulting-rooms session of at least 30 minutes but less than 50, and item 80010 a session of at least 50 minutes. Group M7 (Focussed Psychological Strategies) is for registered psychologists and other eligible allied health professionals, with the psychologists' items in the 80100 range: item 80100 runs from 20 minutes to under 50. Video and phone versions sit in a separate telehealth group (M18) with numbers in the 91000s, such as 91166 for a clinical psychologist's 30-to-50-minute video session.

The explanatory notes allow up to 10 individual and 10 group services per calendar year, with an initial course of up to six individual services before the referrer reviews. MBS Online also lists a Better Access change from 1 November 2025, and schedule fees were last updated on 1 July 2026. Rules move, so do not take these figures from a tutorial or a model's memory: paste the descriptors from MBS Online yourself, with the date you copied them, and treat that pasted text as the source of truth for every check below. Fees are left out here on purpose, because the check only needs durations, provider types and limits.

This is the single most important habit. A model asked "what is the item for a 50-minute session" will answer confidently from training data that may be a year out of date. A model asked "does this session match this pasted descriptor" is doing a comparison, which it does well. Keep the second kind of question and avoid the first.

A reference table you build once from the current schedule

Create a small table in your sheet, one row per item you use. Columns: item number, provider type it applies to, minimum duration, setting (rooms, other, telehealth if applicable), annual limit wording, date copied, and a link to the MBS Online page. Fill it from the live schedule. Illustratively, a row looks like: item [number from MBS], applies to [provider type], minimum [x] minutes, [setting], limit as per current Better Access rules, copied [date].

Put a "review by" date on the table, a month ahead, and make a diary note to recopy it. Schedule changes are announced in advance, and a check built on last year's table is worse than none, because it gives false comfort. If you use practice software with a built-in claim engine, ask the vendor how quickly it updates item rules; the comparison in AI software for psychologists covers what such systems include.

The checks, in order

Run these on a claim list before submission. Each is a rule a spreadsheet can test, and AI helps write the formulas and explain failures.

1. Time against item. Is session end minus start at least the minimum for the item billed? A 28-minute session billed against an item whose descriptor requires 30 minutes is the classic error.
2. Time against a higher item. Does the recorded time support a longer-session item that you did not bill? Report this for information only, never as a suggestion to change upward; the clinician decides what was delivered.
3. Provider type against item series. Is the item in the family that matches the provider's registration or endorsement?
4. Referral present and in date. Is there a referral on file, from an eligible referrer type, in hand by the first claimed session, and covering this session within the number of services it referred for?
5. Session count. Is this session within the annual number allowed under the current rules, counting sessions from other providers where known?
6. Date sanity. Is the date of service on or after the referral date, not in the future, and not a duplicate of another claim line for the same client code?
7. Setting and mode. Does the item match how the service was delivered (in person or telehealth) as recorded in the booking?
8. Notes exist. Is there a note for the session in the practice software, yes or no? Do not read it, only confirm it exists.

Main worked example: one month, 142 claim lines

All figures illustrative. A two-psychologist practice has 142 claim lines for October. The practice manager exports a list with client code, date, start time, end time, item number, referral date, referrer type and session number, with identities removed. She pastes the first twenty rows plus the pasted reference table into a business-tier AI account and uses the prompt below.

You are checking Australian psychology claim lines before submission.
Use ONLY the reference table and the claim rows I paste. The reference table is the current MBS
as copied on [date]. Do not use outside knowledge of item numbers, durations or limits.

For each row, test:
1. Duration (end - start) against the minimum for the item billed.
2. Provider type against the item's provider type.
3. Referral date on or before the service date.
4. Session number against the annual limit wording in the table.
5. Duplicate client code + date.

Output a table: row number, client code, item, test failed (or PASS), one-line reason.
Do not suggest a different item. Do not suggest changing times. If you cannot test something, say "UNABLE TO TEST".

[reference table]
[claim rows]

A realistic output flags 6 of the 20 rows. Row 4 fails the duration test: recorded 27 minutes against a 30-minute minimum. Row 9 shows a referral date after the service date, which looks like a typing error in the referral field. Row 12 is a possible duplicate, and rows 15 and 16 are "UNABLE TO TEST" because the limit wording in the table is ambiguous about sessions delivered by another practice. Row 18 is flagged for provider type: a clinical-psychology item billed under a general registration, which is serious if true and may simply be the wrong provider selected in the export.

What the manager does next matters more than the output. She opens each flagged row in the practice software and checks the session record herself. Row 4 turns out to be a real 27-minute session; the psychologist confirms, and the line comes out of the claim run rather than being left as billed, because no item on the practice's reference table fits a 27-minute session by that provider. What happens to it next is the psychologist's call, made against the schedule. Row 9 is a typing error and is corrected. Row 12 is a genuine duplicate and is removed. Row 18 was an export error. Across the whole month she repeats the check in batches of twenty and finds 14 flags in 142 lines (9.9 percent), of which 9 were real errors worth fixing. The check did not make any decision; it made her look at 14 lines instead of 142.

Doing the same check with formulas, so AI is optional

Most of the tests above are simple enough for a spreadsheet, and rules are more repeatable than a chat model. Ask AI to write the formulas once, check them, then use the sheet each month. In a layout with start time in column D, end time in E, the item's minimum minutes pulled by a lookup in F, and referral date in G, the duration test might be:

Duration (min)   =  (E2 - D2) * 1440
Duration OK?     =  IF(Duration >= F2, "PASS", "CHECK: " & ROUND(Duration,0) & " min")
Referral OK?     =  IF(G2 <= C2, "PASS", "CHECK: referral after service")

Where C is the service date. Have the AI write the lookup from your reference table and the duplicate test with a COUNTIFS on client code and date. Then test the formulas on six rows where you know the answer, including one you have deliberately broken. If you want the sheet to explain a failure in plain English, use AI on the flagged rows only. Some practice management systems also run their own checks; use both, because they fail in different ways.

Never upcode: building the rule into the process

A tool that tells you a session "could be billed at a higher item" creates a pressure that does not belong in clinical billing. The service billed must be the service delivered, supported by the recorded times and notes. Four design choices keep AI on the right side of that line.

• The prompt forbids item suggestions, as above.
• The sheet flags only failures, never opportunities.
• A flagged row is corrected towards what the record shows, whether that is lower, higher or unchanged, by the clinician.
• Recorded start and end times come from the booking or the session itself, not from memory after the fact.

If a time was recorded wrongly and the session was in fact longer, the correction belongs in the clinical record with a dated note first, then the claim. Do not edit a time to fit an item. The same principle applies in other billing contexts, as what still needs a human in medical billing and coding explains.

Referral and session-count traps to look for

Referral problems are the trap to design for first. The MBS explanatory note on Better Access referrals (MN.6.3) says the psychologist must have the referral in hand at the first consultation, that a Mental Health Treatment Plan on its own is not a referral, that the referral should state the number of services referred for, and that it must be kept for 24 months from the first service. Check these in your own process against the current note:

• A referral that arrived after the first session and was never dated, so the first claim line has no valid referral.
• A treatment plan filed as if it were the referral, with no separate signed and dated referral letter behind it.
• A referral for a set number of services that has been used up, with the next session booked anyway. A new referral is needed before that session is claimed.
• A new referral that replaces an old one without the session count being reset or carried over correctly, because the rules for counting are specific.
• A referral from a type of referrer who is not eligible for that item.
• A client who has used sessions with another provider earlier in the year, which the client may not mention. Ask at intake and recheck at session six.
• Telehealth sessions recorded as in-person because the booking system defaults to rooms.

Build a "session count" column that increments per client code within the calendar year, and a "referral on file" column that reads yes or no. Colour a row when the count reaches two short of the annual limit as listed in your reference table, so the clinician talks to the client about next steps early. Illustratively, a client at session eight of ten whose referral has no review planned is better told in session eight than in session eleven.

Where AI scribes and notes fit, and where they stop

If a practice uses an AI scribe, the notes it produces are clinical records and fall under the same consent and confidentiality expectations as any note. They are not claim data. Keep the billing check on structured fields only, so there is never a reason to paste a note. What psychologists must consider when using scribes, including what the regulator expects, is set out in can psychologists in Australia use AI scribes? and the consent steps in AI scribe notes, reports and consent.

One useful boundary: the only thing from the note the claim check may use is the fact that a note exists and the time of the session. If the scribe records a start and end time, the practice manager may use that to cross-check the booking time; the content stays where it is.

Three practice shapes, three set-ups

• A solo psychologist. One sheet, one reference table, a monthly check before the claim run. Twenty minutes. No AI needed after the formulas are built.
• A group practice with an admin team. The admin runs the check weekly, flags go to the treating psychologist by client code, and a log records every correction and who made it.
• A practice using a billing service. Ask the service whether it runs similar checks and what it sees. Never send session notes to a billing provider that does not need them. Ask which AI tools it uses and on which plans.

For other admin tasks worth automating around the same data, see seven admin jobs beyond session notes.

A second month: the same check after a schedule change

Rules change, and the check has to change with them. Suppose the practice reads in an MBS Online fact sheet, or an announcement from the Department of Health, Disability and Ageing, that an item descriptor or a limit will change from the first of next month. The manager sets a four-step routine, illustrative but repeatable. First, copy the new descriptors into a second tab of the reference table with an effective date. Second, add an "effective from" column to the lookup so that a session dated before the change uses the old row and one dated after uses the new. Third, run the previous month's claims through the new rules to see what would flip, which shows how much the change affects the practice. Fourth, tell the clinicians what changes in practice, in two lines each.

AI is useful for the second and fourth steps. Ask it to rewrite the lookup formula so the right row is chosen by date, and to turn a pasted descriptor change into a two-line summary of "what is different for us". Then verify the summary against the source text, as you would any summary. A mistake here is easy to make and expensive: a practice that applied the old minimum duration for a fortnight after a change would have sent a batch of claims on the wrong basis.

How to explain a flag to a clinician without a fight

Psychologists, reasonably, dislike being told by a spreadsheet that their claim is wrong. The way the flag is worded matters. Compare "Claim error: item invalid" with "Session recorded 27 minutes; item billed lists a 30-minute minimum. Please confirm the session length or the item." The second is a question, names the facts, and leaves the decision with the clinician. Ask the AI to turn every flag into that second form using only the row data, and to keep each under thirty words. A weekly message listing flags by client code, with the one-line reason and a column for "confirmed / corrected / not billed", takes ten minutes to send and closes the loop in a way that a sticky note does not.

Keep a count of flags per month and watch the trend. If the same type of error recurs, for example referrals dated after the service, the fix is upstream, in the intake form or the booking template, rather than in more checking. In the worked month above, 9 real errors out of 142 lines is a fair starting point; the aim is to see that number fall over three months because the intake process improved.

Which AI account, and what it costs

Because the check uses structured, de-identified fields, the privacy risk is low, but the habit still matters: use a paid business plan with training off by default rather than a personal login. ChatGPT Business is $25 per user per month billed monthly or $20 annually and, like Claude Team, needs two seats; Microsoft 365 Copilot Business is $21 per user per month on annual billing (list prices in USD checked this month; confirm local pricing on the vendor pages). If your practice already has Google Workspace or Microsoft 365, ask the tool built into it first. A formulas-only approach costs nothing beyond your existing software.

A pre-submission checklist for each claim run

1. The reference table was copied from MBS Online within the last month, and the date is on it.
2. The export carries client codes only, with no names, dates of birth or Medicare numbers.
3. Duration, provider type, referral, session count, duplicate and mode checks have been run.
4. Every flagged row has been opened in the practice software and decided by the treating psychologist or the practice owner.
5. No claim was changed to a higher-value item without a recorded reason in the clinical record first.
6. Corrections are logged: row, change, who, date.
7. No session notes were pasted anywhere outside the practice software.

If you would like help setting up a check like this, or mapping which admin steps in your practice suit AI, that is what my AI implementation consultation covers. The overall picture is in the AI guide.

Questions psychologists ask about AI and Medicare claim checks

Can AI tell me which MBS item number to bill?

It can help you check a choice, but it should not make it. The correct item depends on your registration type, the session duration, the setting, the referral and the current schedule wording. Use AI to compare a structured claim line with the item descriptor you paste from the current MBS, then decide yourself.

Do I need to paste session notes into AI to check claims?

No, and you should not. A useful check needs only structured fields: date, start and end time, item number, referral date and referrer type, and the session number for the year. Clinical content stays in your practice software. If an AI scribe is in use, check its settings and your consent process separately.

What are the most common Medicare psychology claim errors?

The usual ones are an item that does not match the session length, a session after the referral has run out or without a valid referral, a provider-type mismatch, a wrong date of service, and a count of sessions beyond what the current rules allow. Check every one against the schedule rather than from memory.

Is it upcoding if the AI suggests a longer-session item?

It is if the session was not that long. Billing a higher-value item than the service delivered is a compliance problem whichever tool suggested it. Set the rule that the time you billed must be supported by the recorded start and end times, and that any flag moves you towards the lower item or a correction, never towards a higher one by default.

Further reads

• AI Note-Taking Tools for Therapists: A Confidentiality Checklist — Confidentiality checks for AI note tools in therapy.
• Patient Data and AI: A Confidentiality Checklist for Small Practices — A fuller checklist for keeping client data out of AI tools.
• ChatGPT for Medical Billing: Appeals and Claim Checks Without PHI — Claim checks without patient information, from the billing side.
• AI Denial Management for Medical Billing: Group, Appeal, Track — Handling rejected and queried claims once they come back.
• AI for Testing Psychologists: Report Drafts Without Pasting Raw Scores — The never-paste rule applied to reports.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: MBS Online item pages for 80000, 80010, 80100 and 91166, explanatory notes AN.0.78 (Better Access) and MN.6.3 (referral requirements), and the MBS Online fact sheet list, checked 4 November 2026; the pricing pages for ChatGPT Business, Claude Team and Microsoft 365 Copilot Business, checked the same day. Item numbers and rules change, so recheck them at the time of claiming.