Normal Weekly Earnings: AI Can Assemble the Evidence, Not Set the Figure

The useful AI workflow for normal weekly earnings does not calculate NWE. It makes the evidence ready for calculation. That distinction matters because a clean spreadsheet can still produce the wrong answer if the pay period is unrepresentative, the overtime evidence is incomplete, or a payroll code has been mistaken for a statutory category.

Under the current Safety, Rehabilitation and Compensation Act 1988, Compilation 82 (the SRC Act), sections 8 and 9 govern NWE and the relevant period. Section 19 then uses NWE within the incapacity compensation framework. The sequence is evidence, legal characterisation, human calculation and determination. AI belongs in the first step.

That makes this workflow different from an incapacity calculation cross-check. A cross-check tests arithmetic after a person has selected the inputs and performed the calculation. An NWE evidence pack tests whether the human decision-maker has the right records, the right date coverage and a traceable basis for classifying each input before any figure is set.

What belongs in an NWE evidence pack?

Start with the statutory structure, not the fields exported by payroll. Section 8(1) uses average weekly hours, the average hourly ordinary time rate of pay and qualifying allowances. Section 8(2) adds an overtime component where the employee was required to work overtime on a regular basis. A payroll system might label an amount OT, ALLOW or SHIFT, but that code does not decide whether the section 8 requirements are met.

For a self-insured bank, that may mean reconciling payroll, rostering, HR and approval records without treating any imported claim-system label as conclusive.

Section 9(1) identifies the default relevant period, subject to the rest of section 9, as the latest two-week period before the date of injury during which the employee was continuously employed by the Commonwealth or a licensed corporation. That is not necessarily the last two payslips. Pay cycles can straddle dates, and section 9 contains rules dealing with pay variations and periods when earnings were reduced or absent because the employee was away from work.

The evidence pack should therefore have four linked layers:

1. A period coverage map. Show the date of injury, the apparent section 9 period, every pay cycle that overlaps it, and any absence, pay variation or employment break that may affect the period. Do not silently fill a missing date.
2. A pay-element register. List ordinary hours, ordinary rates, overtime hours, overtime rates, allowances and other payments exactly as the records describe them. Put the source beside each item and keep legal classification blank until a person completes it.
3. An authority trail. Link overtime to the roster, approval, contract term or other material said to show it was required and regular. Link each allowance to the employment instrument and its stated purpose.
4. A decision hand-off. Present unresolved questions, competing records and source locations without recommending a period, including an amount or applying the formula.

Comcare's NWE guidance says the decision-maker is responsible for determining the figure. It also says an initial NWE determination should identify the period used, explain why it was selected and summarise the data considered. The guidance is scheme guidance, not the Act, but it gives the evidence pack a practical destination: reasons that can be traced back to records.

How do you stop AI from choosing the answer?

The control is a zero-math hand-off. The AI may extract, reconcile, sort and flag. It may not total earnings, apply the section 8 formula, nominate the relevant period, classify overtime as required and regular, or decide that an allowance qualifies.

That boundary should appear in the prompt and in the output schema. If the model is given a column called included in NWE, it will try to complete it. Replace that column with human classification required and require a source reference for every factual entry.

The first prompt builds the pack. De-identify the records first using the placeholder discipline in the de-identification toolkit. A human must compare every extracted item with the original de-identified record and make every statutory classification.

You are preparing an evidence pack for a human decision-maker under sections 8 and 9 of the Safety, Rehabilitation and Compensation Act 1988.

The material is de-identified. Use only these placeholders:
[EMPLOYEE_NAME], [CLAIM_NUMBER], [DATE_OF_INJURY], [PAY_PERIOD_START], [PAY_PERIOD_END], [EMPLOYMENT_INSTRUMENT].

From the supplied records, produce:
1. A date coverage map showing the date of injury, the apparent two-week period before injury, overlapping pay cycles, absences and pay-rate changes.
2. A source ledger for ordinary hours, ordinary rates, overtime hours, overtime rates and each allowance.
3. For every item, the source file, page or row, exact record wording, and any conflicting source.
4. A list of missing records and questions requiring human resolution.

Do not calculate NWE or any component. Do not recommend a relevant period. Do not decide whether overtime was required or regular. Do not decide whether an allowance qualifies. Write "human classification required" wherever legal characterisation is needed.

Then apply a provenance lock. No factual proposition enters the pack unless it carries a file name and page, row or cell reference. If the source cannot be located, the item goes to the gaps register rather than the evidence table.

This is especially important for section 8(5). That provision is not a general licence to choose the most favourable, convenient or statistically stable period. It operates where, because of the shortness of the relevant period, the calculated result would not fairly represent the weekly rate at which the employee was being paid, and it permits another period considered reasonable for reaching a fair representation. Comcare's guidance gives examples of longer periods, but those examples are not automatic rules.

The pack can identify fluctuating shifts, periodic overtime, a pay change, an absence or a short employment history. The delegate decides whether the statutory trigger is met and what period, if any, is reasonable.

What does the workflow look like on a fictional file?

Consider this fictional, de-identified scenario. [EMPLOYEENAME] has an accepted claim under [CLAIMNUMBER], with [DATEOFINJURY] recorded consistently across the claim form and medical material. Payroll provides two payslips, a six-week transaction export, rosters, an enterprise agreement extract and a manager's overtime approval log.

The apparent two-week period contains annual leave, one weekend shift and an allowance coded OPS. The six-week export shows OPS in four pay cycles, but the payslips do not explain its purpose. Overtime appears three times, while the approval log contains only two matching entries.

The AI does not decide that six weeks should replace the section 9 period. It produces a period coverage map showing the annual leave and the overlapping pay cycles. It records the repeated allowance, links it to the enterprise agreement clause if one can be found, and flags that its purpose remains unclear. It also identifies the unmatched overtime entry and asks for the source said to establish authorisation.

The resulting checklist is short and useful:

• Is [DATEOFINJURY] consistent across the claim, payroll and employment records?
• Does the period map reflect section 9 rather than the payroll cycle label?
• Are absences and pay variations visible rather than averaged away?
• Does every overtime entry have evidence relevant to whether it was required and regular?
• Does every allowance have a source describing why it was paid?
• Are conflicts and missing records separated from confirmed facts?
• Has a human selected the period, classified the inputs and performed the calculation?

The second prompt red-teams the pack, a narrower cousin of the determination evidence check skill file. A human must open each cited record, correct any false extraction and decide whether further evidence is required.

Review this de-identified NWE evidence pack for source integrity only.

For each entry:
1. Check whether a source location is stated.
2. Compare the entry with the quoted source text supplied.
3. Flag date gaps, duplicated records, inconsistent hours, unexplained pay codes and unsupported descriptions of overtime or allowances.
4. Separate confirmed record content from assumptions and human decisions.

Return three tables: verified extracts, conflicts, and missing evidence.

Do not calculate an amount. Do not select or recommend a relevant period. Do not classify an allowance or overtime item. Do not predict a determination. Every flagged item must be reviewed against the original record by a human.

This pass does not replace reading the file. It should catch an output that says "regular overtime" when the source merely says "overtime".

The current Comcare NWE advice form provides useful evidence fields for Australian Government agencies. It is not a universal form for every self-insured licensee, which must check its own scheme process, licence settings and controls.

Do this Monday

1. Remove calculation from the AI step. Write down that the output is an evidence pack, not an NWE worksheet. Delete any prompt instruction that asks the model to total, average, include or exclude amounts.
2. Create the period coverage map. Require the date of injury, section 9 period, overlapping pay cycles, absences and pay variations to appear on one view. Make missing dates visible.
3. Add source coordinates. Require a file and page, row or cell for every extracted fact. Route entries without a location to a gaps register.
4. Separate record labels from legal categories. Keep the payroll code in one field and the human classification in another. Never let OT become "required regular overtime" by default.
5. Test one fictional file. Use only invented facts and placeholders. Insert an unsupported allowance and an unmatched overtime entry, then check whether the workflow finds both without choosing the answer.
6. Name the decision owner. Record who selects the relevant period, classifies each input, performs the calculation and signs the section 8 determination.

A section 8 decision falls within the Act's definition of a determination in section 60(1). The notice requirements in section 61 require a human process that states the terms and reasons of the determination and preserves reconsideration and review rights. A source-linked pack supports that process without supplying the reasons.

Bottom line

NWE errors are often evidence errors wearing the clothes of arithmetic. AI can make the source trail visible before a delegate chooses the period or touches the formula. It cannot decide that overtime was required and regular, classify an allowance, invoke section 8(5), or determine the figure. Build the evidence pack first, then keep the determination and calculation human.

This article is general information and education only. It is not legal advice, and it is not advice about any individual claim. Decisions under the Safety, Rehabilitation and Compensation Act 1988 are made by human decision-makers on the individual merits of each claim, and claimants have reconsideration and review rights in respect of determinations. Seek advice specific to your scheme and circumstances.

References

1. Federal Register of Legislation, Safety, Rehabilitation and Compensation Act 1988, Compilation No. 82, effective 1 July 2026. https://www.legislation.gov.au/C2004A03668/latest/text
2. Comcare, Scheme guidance: Calculating normal weekly earnings, document SRC315, last updated August 2023. https://www.comcare.gov.au/scheme-legislation/src-act/guidance/calculating-normal-weekly-earnings
3. Comcare, Normal weekly earnings advice, August 2025. https://www.comcare.gov.au/sites/default/files/docs/normal-weekly-earnings-advice-form.pdf