NDIS Plan Manager Invoices: Checking Claims Against Budgets With AI

An NDIS plan manager can use AI to check provider invoices by extracting the key fields from each invoice, comparing them with the participant's category budget and service agreement, and producing a flag-and-hold list for staff. AI reads and compares; a trained team member approves every payment, and participant details stay in your own system.

The plan manager's position differs from the provider's. You did not deliver the support, so you cannot compare an invoice against a shift note. Your evidence is the invoice itself, the plan and its budgets, the service agreement and the participant's confirmation. That makes the checks narrower and more mechanical, which suits software, and the cost of a wrong payment falls on a participant's budget, which is why the human approval is not negotiable.

Follow me on Instagram@sagnikteaches

What you can and cannot check from where you sit

Plan managers pay invoices on a participant's behalf, claim from the funder and keep records of how the funds are spent. The participant still chooses their providers and supports, and the funder's plan manager guidance asks you to make sure participants are aware of and approve claims, and that each claim fits their plan, their budget and what counts as an NDIS support. So the questions you can answer are about form and fit, not delivery:

Connect on LinkedInSagnik Bhattacharya

• Is the invoice complete? The funder's list of invoice details includes the provider's business name and number, the participant's name and NDIS number, support dates, quantity or hours, the support item number from the catalogue, claim type, total amount and any tax component. Each invoice covers one participant.
• Does it fit the plan? Is the support item in a category the participant has funds for, and does the amount fit what remains in it?
• Does it fit the agreement? Is it a service the participant has agreed with this provider, at an agreed rate and frequency? Clear agreements make this check far easier; the service agreement tutorial covers the wording.
• Is it unusual? A duplicate, a jump in hours, a new item from a long-standing provider.

The question you cannot answer is whether the support was delivered as described. For that you rely on the participant's confirmation and the provider's own records. A good process asks for confirmation for unusual lines and keeps a route open for participants to dispute anything.

Subscribe on YouTube@sagnikteaches

The data you need, set out as tables

Three tables, all in your own system, give the check its reference points. Keep them small and current.

The funder's guidance says every plan has four support budgets (core, capacity building, capital and recurring), though a participant may have funding in only some of them, with 21 support categories beneath them and support items beneath those. Whether funding is flexible or stated changes the check. Flexible funding can be used across the support categories within the same support budget, but not across budgets; stated funding can only buy the support the plan describes. So an invoice in a category that has run dry is not automatically a hold if flexible funding elsewhere in the same budget covers it, while an invoice against stated funding for a different support is. Record which applies in the budgets table. Keep your rates and rules table under a named owner who rereads the funder's current pricing arrangements on a fixed schedule, since price limits and item rules change and the AI tool does not know your version.

Step 1: get invoices into a structured table

Invoices arrive as PDFs, emails and forms. The first job is extraction into one row per invoice line. Many plan management systems have their own capture; where yours does not, a business-plan AI tool with file upload can read a PDF and return fields. Check the output against the document for the first fifty invoices, then sample.

Extract these fields from the invoice below into one row per line item:
provider_name, provider_business_number, invoice_number, invoice_date,
participant_ref (the reference in the header, not the name), service_date,
support_item_number, description, quantity, unit, unit_price, line_total,
tax_amount, invoice_total. If a field is not on the invoice write MISSING.
Do not infer or correct anything. Do not calculate.

INVOICE TEXT:
[invoice text with participant name and address removed]

"MISSING" is a gift: it turns "the tool guessed" into "a person needs to look". Illustrative output for a faulty invoice: ... | [item no.] | Community access | 3 | hours | 62.00 | 186.00 | MISSING | 186.00. The missing tax field and the absent participant number are checked next.

Step 2: the mechanical checks (no AI needed)

Put the table in a spreadsheet or your system's rules engine and run exact checks. These are cheap, fast and explainable, which is what you want when a provider asks why an invoice was held.

1. Required fields present (business number, participant number, item number, dates, quantity, total).
2. Arithmetic: quantity times unit price equals line total; lines sum to invoice total.
3. Duplicate: same provider, participant, item, date and amount already paid.
4. Item in plan: the support item belongs to a category in the participant's budgets.
5. Funds available: line total is within the funds it can draw on (the category for stated funding, the wider support budget where funding is flexible), allowing for amounts committed on other pending invoices.
6. Plan dates: service date falls inside the plan period, and inside the right funding period where the plan releases funding in periods.
7. Rate: unit price at or under the current price limit for the item (where one applies), per your rates table.
8. Agreement: provider and item appear in the agreements table, and quantity is consistent with it.

Each failed check sets a flag with a reason code. Do not ask an AI tool to do the arithmetic. It can get it wrong, and a spreadsheet formula is both free and exact. For help building the formulas, cleaning messy data covers the preparation that makes a lookup reliable.

Step 3: where AI earns its place

Three jobs are fuzzy enough that formulas struggle and AI helps.

Reading descriptions against items. An invoice line described "weekend outing and lunch" under a personal care item looks odd. A prompt that compares the description with the item name and says whether they appear consistent can flag it for a person. It gives a hint, not a ruling, because item rules are detailed and context matters.

Spotting patterns across invoices. A pasted table of six months of one provider's lines for one participant can be summarised: hours per week trending upwards, new items appearing, a round-number total repeating. Patterns are a reason to ask a question, not evidence of anything wrong.

Drafting the query. Writing to a provider to explain why an invoice is held and what is needed takes ten minutes when done by hand. A drafted message from the reason codes, with a person editing, takes two.

Draft a polite email to a provider about a held invoice. Facts only:
invoice number [no.], held because [reason codes in plain words],
we need [what is required] by [date]. Do not suggest the claim is
wrong or accuse anyone. Mention the participant by reference only.
Under 120 words. Sign off "[your name], payments team".

Illustrative output: "Hello, invoice [no.] is on hold. The line for 12 March does not show a support item number, and the invoice total does not match the sum of the lines. Please send a corrected invoice by [date], and we will process it as soon as we receive it. Thanks, [your name], payments team." Short, specific and free of blame, which is the tone to keep.

Worked example: a Monday batch of 60 invoices

Take an illustrative plan management team of three with 60 invoices in Monday's batch, averaging 2.4 lines each, so 144 lines. Extraction runs in the system and a person spot-checks eight invoices against the PDFs; all match. The mechanical checks flag 19 lines across 14 invoices: 6 with missing required fields, 4 duplicates (an invoice sent twice by the same provider), 3 items outside any funded category, 3 over the remaining balance and 3 with a rate above the limit in the reference table.

Next the AI step reads the 19 flagged lines and the surrounding history, and returns a one-line suggestion for each, such as "likely resubmission of an invoice paid last week" or "item description does not match item name". A team member goes through them in about 40 minutes. Eleven go back to providers with the standard hold message, 4 duplicates are rejected with a note, 2 are cleared after the participant confirms by phone that the service happened, and 2 are paid because the reference sheet was out of date. Those last two are valuable: they show the reference sheet needs updating, and the fix goes in the log.

The remaining 125 lines pass the checks and go to a second person for a sample review (say 10 per cent, 13 lines) before payment is approved. The illustrative time for the batch falls from about five and a half hours of manual checking to about two and a half, and more of the time goes into the invoices that deserve a look. Those are illustrative figures; measure your own before and after.

Flag-and-hold: the rules that protect participants

A hold list is a promise to treat each flagged invoice with care. Write the rules down.

• Hold reasons are specific. Each flag has a code and a plain-language explanation, so a provider can fix it.
• Timing is part of the rule. The funder's plan manager guidance says to pay providers or reimburse participants within 2 business days of receiving payment from the funder, so the checks and any hold belong before you lodge the claim, not after the money arrives. Tell providers and participants promptly when an invoice is held, and recheck that guidance when you review your rules.
• The participant can challenge. Participants choose their providers and supports. If you hold an invoice for budget reasons, tell the participant what remains and what the options are, rather than quietly refusing.
• Nobody pays on a tool's say-so. Approval needs a name and a date in the log, with a second approver above a set value.
• Rejections are explained. A duplicate is rejected with the invoice number it duplicates.

Four situations that test the rules

Rules look tidy until a real case arrives. These four illustrative situations show how the checks and the human decisions fit together.

A participant near the end of the plan period

An invoice arrives for $1,240 in a category with $1,300 remaining, but the plan ends in nine days and two more invoices from the same provider are expected. The mechanical check passes, because the funds are available. A good process adds a "plan end date within 30 days" flag, because the question is no longer whether this invoice fits but whether the remaining funds are being used as the participant wants. A person rings the participant, explains where the budget stands and asks whether any other supports are still due. The AI tool's part is minor: it drafts the call notes and the follow-up message. The decision is a conversation with the participant.

A rate change mid-month

A provider's invoices jump from one unit price to a higher one on the first of the month. Your reference table shows the new limit is higher too, so the check passes. The pattern step notices the jump anyway and asks whether the agreement has been updated. A person checks the service agreement, finds the old rate still stands, and queries the provider. The provider explains the agreement was renegotiated and sends the signed copy. The agreements table gets the new rate. Nothing was wrong, but the question was worth asking, and the answer now sits on file.

A participant disputes a line

The participant tells you a Saturday session on an invoice never happened. You hold the invoice, tell the provider what the participant said, using the reference rather than speculation, and ask for the support log and time records for that date. You do not pay until the matter is resolved, and you keep a dated note of each contact. AI is useful only in drafting the neutral request, and you keep it neutral: "The participant has told us they do not recall this session. Please send your record for that date." The finding, whichever way it falls, is for people to reach.

A sudden surge from one provider

One provider's invoices for a participant triple in a month. The pattern summary flags it. It might be a plan change, a new package or an error. A person reads the agreement, speaks to the participant and asks the provider for an explanation. If something looks seriously wrong, such as invoices for supports the participant says they do not receive, your organisation's escalation procedure applies, and you take advice on whether and where to report. The tool does not decide that; it raises the question early.

Month-end reconciliation with AI help

Once payments flow, reconciliation stops mistakes quietly compounding. At month end export three lists: invoices paid, claims lodged with the funder and money received. A formula join shows paid invoices with no matching claim, claims with no matching payment, and receipts with no matching claim. Those are timing differences or errors, and a person works through them. AI can help by drafting a short explanation of each unmatched item for the file, once you have established the cause. A tidy reconciliation log also feeds your records if the funder reviews your claims.

The same export feeds the monthly financial statement that the funder's guidance says each participant must receive: how much of the plan has been spent, whether spending is consistent with expectations for the funding period, the remaining balance, and any concern about overspending. AI can draft the plain-language covering note from those figures ("Your core budget is 62 per cent spent with five months of the plan left, which is a little ahead of an even pace"); a person checks every number against the system before it goes out, because a wrong balance in a statement is a wrong promise to the participant.

Illustrative numbers: 312 invoices paid in the month, 309 matched to claims, 3 unmatched. Of the three, one was lodged late (claimed next week), one was rejected by the funder because of a plan end date and needs a reply to the provider, and one was a duplicate payment caught and recovered. Three lines out of 312 is under one per cent, but each carries a story that someone needs to close.

Training new team members on the checks

A new payments officer is the weak point of any process, so teach with real material. Prepare ten anonymised invoice examples with known problems: a missing number, a duplicate, an item outside the budget, a plausible but wrong rate, a round-number total that does not add up. Ask the new person to run the checks and make a decision on each, then compare with the answer key. An AI tool can generate fresh practice invoices with planted errors, which saves you inventing them, but make sure the examples are invented and not real invoices with names removed. After a week of supervised batches, move to unsupervised work with a second approver above your set value for the first month.

Protecting participant privacy

Plan managers hold a lot: plan details, budgets, provider lists, sometimes sensitive notes. Build the checks so most of them use a participant reference, not a name. Do not paste plan documents into a chat tool. Use a business plan with data terms your adviser has read; business plans such as ChatGPT Business, Claude Team and Microsoft 365 Copilot do not train on business content by default. Give each staff member their own login, and keep a short record of which tool is used for which task. The tutorial on safe uses under the privacy principles goes through what to tell staff and participants.

Tell participants plainly how invoices are checked, in your service agreement: that software helps compare invoices to budgets, that a person approves every payment, and how to ask for a human to review a decision.

Stale rates, pending invoices and false duplicates

Stale reference table. A rate limit changes, your sheet does not, and either valid invoices are held or overpriced ones pass. It shows up as a run of rate flags on one item. Fix with a dated review schedule.

Committed funds ignored. Three invoices each fit the remaining balance but together exceed it. The check must subtract pending approvals, not just paid amounts.

The helpful extraction. A tool "corrects" a missing item number by guessing. Your prompt says not to; sample the output to make sure that holds after each tool update.

Duplicates that are not duplicates. A provider legitimately delivers the same support item on the same day for two shifts. A duplicate check that ignores times will wrongly reject it. Include start and end times where they exist.

Held-invoice backlog. Flags pile up because nobody owns the follow-up. Set a daily follow-up slot, and track the age of each hold.

Checking the process works

Measure five things monthly: share of invoices held and why (by reason code), median time from receipt to payment decision, holds later cleared as false alarms, payments later found wrong, and complaints. If false alarms are more than a third of holds, tune the rules. If any payment is found wrong after approval, trace which check should have caught it and add it. Keep the figures in a one-page summary to share with your team.

For the other half of the picture, the provider's pre-claim check against shift notes is covered in the provider invoicing tutorial; seeing how providers check their own claims helps you write better hold messages.

The funder's pricing arrangements, plan management guidance and record keeping rules are the authority. Check the current pages and ask a qualified adviser about anything that affects how you pay or hold a claim.

Plan manager questions

Can AI approve invoices for payment?

No. A plan manager pays on a participant's behalf and is expected to check that each claim meets the funder's compliance requirements, so a person approves every payment. AI can read an invoice, compare fields to the plan record, and list what looks wrong. The decision to pay, hold or query belongs to a trained team member, recorded with their initials.

What data should the AI tool see?

As little as the check needs: a participant reference, support category, remaining budget, the invoice fields and the service agreement terms. Keep names, addresses and plan documents in your own system. Use business plans with reviewed data terms, and give staff individual logins so the check log is attributable.

Where do the budget figures come from?

From your own records and the participant's plan as held in your plan management system, not from the AI tool's memory. Export the category balances as a table and have the check compare the invoice against it. Rates and rules change, so keep a current reference sheet and update it when the funder's pricing arrangements change.

What should happen when an invoice is held?

Tell the provider and the participant promptly and plainly what is missing or needs confirming, and keep a note of the date and reason. A held invoice with no message to the provider is the start of a late payment complaint. Use a standard message, edited by a person, and a due date for the reply.

Further reads

• AI Accounting for NDIS Providers: Categorising Income and Claims — Categorising NDIS income and claims in the books.
• AI for Disability Services: Service Agreements and Incident Logs — Service agreements and incident logs in disability services.
• How to Automate Invoicing With AI: From Finished Job to Paid — Invoicing automation from finished job to paid.
• NDIS Software for Small Providers: Per-User Costs and What You Get — Software costs to compare before you buy more tools.
• NDIS Providers Automation Assessment Checklist: Which Jobs for AI First? — Rank admin jobs worth automating first.
• How Accountants Use AI to Spot Errors in Client Books — Error-spotting methods you can borrow for payment batches.
• How to Get Clients as an NDIS Provider: AI for Enquiries, Referrers — A workflow for new NDIS providers: directory profiles, support-coordinator outreach and AI-drafted enquiry replies that a person always checks.
• How to Write NDIS Progress Notes: An AI-Assisted Audit Checklist — A provider checklist for audit-ready progress notes, with good and weak examples and safeguards for voice-to-text and AI review.
• Can AI Write NDIS Behaviour Support Plans? What the Commission Says — What the NDIS Commission's February 2026 statement says about AI in behaviour support plans, turned into a workflow a provider can follow.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NDIS provider record keeping requirements and invoice field lists (ndis.gov.au, current as of August 2026); NDIS guidance on plan management and support budgets; vendor data-use pages for business AI plans.