Ask your payroll lead a narrow question this week. If an ex-employee alleged in 2031 that they were underpaid across a period in 2026, could you produce the records for it? Then ask the follow-up that matters now: which of those records will have been generated by a tool, and where does the tool keep them.
The obligation is unglamorous and it is old. What has changed is that a growing share of the artefacts sitting behind it are now produced by systems that were designed to answer a question quickly, not to hold a record for seven years.
What does the Act actually require?
Subsection 535(1) of the Fair Work Act 2009 is one sentence: an employer must make, and keep for 7 years, employee records of the kind prescribed by the regulations in relation to each of its employees. Subsection 535(2) adds that the records must be in the form prescribed by the regulations, if a form is prescribed, and must include any information prescribed by the regulations. Subsection 535(3) allows the regulations to provide for inspection of those records. Subsections (1) and (2) are civil remedy provisions.
There is a second obligation alongside it. Subsection 536(1) requires an employer to give a pay slip to each of its employees within one working day of paying an amount to the employee in relation to the performance of work, and subsection 536(2) sets the form and content requirements.
The Act also anticipates a specific failure. Subsection 535(4) provides that an employer must not make or keep a record for the purposes of the section that the employer knows is false or misleading, subject to subsection (5), which excludes records that are not false or misleading in a material particular. Subsection 536(3) does the same for pay slips.
Three obligations, then: make it, keep it for seven years, and do not knowingly keep one that is wrong.
What happens if the record is not there?
This is the part that most operational discussions miss, and it is the reason the topic belongs in an AI conversation rather than a filing conversation.
Section 557C is titled "Presumption where records not provided". Under subsection 557C(1), where an applicant makes an allegation in relation to a matter in proceedings relating to a contravention by an employer of a listed civil remedy provision, and the employer was required by subsection 535(1) or (2) to make and keep a record, or by regulations made for the purposes of subsection 535(3) to make a record available for inspection, or by subsection 536(1) or (2) to give a pay slip, in relation to that matter, and the employer failed to comply with the requirement, then the employer has the burden of disproving the allegation.
The Act signposts this inside section 535 itself. The note to subsection 535(3) states that if an employer fails to comply with subsection (1), (2) or (3), the employer may bear the burden of disproving allegations in proceedings relating to a contravention of certain civil remedy provisions, and refers the reader to section 557C. The same note appears under section 536.
Subsection 557C(3) lists the provisions that trigger it: section 44 on the National Employment Standards, section 45 on modern awards, section 50 on enterprise agreements, section 280 on workplace determinations, section 293 on national minimum wage orders, section 305 on equal remuneration orders, subsections 323(1) and 323(3) on methods and frequency of payment, subsection 325(1) on unreasonable requirements to spend or pay amounts, and any other civil remedy provisions prescribed by the regulations.
There is one release valve. Subsection 557C(2) provides that subsection (1) does not apply if the employer provides a reasonable excuse as to why there has not been compliance. Notice what the excuse has to be about. It is an excuse for the absence of the record, offered after the fact, in a forum where the alternative is disproving a claim about hours you no longer have data for.
Where does AI actually break this?
Not in the way people expect. The risk is rarely that a model calculates something wrong, although that happens. The risk is that the useful output of an AI-assisted process is not the same object as the record the Act requires, and nobody notices the difference until it is needed.
Five failure modes are worth naming.
The output was never a record. An award interpretation assistant tells a manager that [EMPLOYEE_NAME] is entitled to a particular penalty rate for a shift. The manager acts on it. The reasoning existed in a chat window that has since rolled off. What was kept is the payment, not the basis for it. When the question is asked years later, the payment is the allegation and the basis is gone.
The record lives in a vendor's retention window, not yours. Many AI products default to retention measured in days or months, and some enterprise configurations shorten it deliberately for privacy reasons. That is a sound privacy posture and a direct collision with a seven-year statutory obligation if the product is holding something that forms part of the employee record. The fix is not longer retention everywhere. It is deciding which artefacts are records and exporting those into the system that already holds records.
The record survives the contract but you do not have it. Where an AI or HR technology vendor is replaced, exports are usually negotiated for current-state data. The obligation runs for seven years from the making of the record, which will typically outlast the contract. Ask what happens to historical artefacts at termination before you sign, not during the migration.
A summary replaced the record. Summarisation is what these tools are best at, and a summary of a roster, a time record or a leave decision is not the prescribed information. It reads like a record and fails the content test.
Nobody can find it. Subsection 535(3) contemplates inspection. A record that exists in an unsearchable form across three systems, one of which is a chat transcript, is a record you will struggle to make available in the period a request allows.
The practitioner play
This is a mapping exercise, and it is finishable in a fortnight for most organisations.
- List the decisions, not the tools. Write down every employment decision now touched by AI that has a pay or entitlement consequence: rostering, shift allocation, time capture and approval, award or agreement interpretation, leave assessment, overtime approval, allowance eligibility. Tools change. Decisions persist.
- For each decision, name the record. Against each, write the specific record the regulations require in relation to that matter. Where a team cannot name it, that is the finding, not a gap in the exercise.
- Trace where that record is written. Follow it to a system, not a person. Note whether the system is the one that already holds employee records or whether it is the AI tool, a spreadsheet, a chat transcript or an inbox.
- Check the clock on each location. For every location that is not the system of record, ask what its retention period actually is and compare it to seven years. Anything shorter is either a migration task or an accepted risk that somebody senior should accept in writing.
- Fix the join, not the tool. Where an AI tool sits in the path, the durable pattern is that the tool proposes and the system of record stores. The output that matters is written into the employee record at the time of the decision, with the inputs it relied on, rather than left in the tool.
- Test one retrieval. Pick a real fortnight from the current financial year for one team, [TEAM], and try to assemble the complete records for it as if responding to an inspection. The exercise takes an afternoon and it is the only part of this list that produces evidence rather than intent.
Where the human decision stays
Two boundaries are worth stating explicitly because automation makes them easy to blur.
The first is that a record must be accurate and the employer is on the hook for knowing when it is not. Subsection 535(4) prohibits making or keeping a record the employer knows is false or misleading. Where an AI tool populates a record, somebody has to remain in a position to know whether it is right, which means the design must keep a human able to check rather than merely able to approve. An approval workflow where the reviewer has no practical means of verifying the input is a rubber stamp with a name attached to it.
The second is that entitlement decisions remain decisions. A model can read an award, retrieve the relevant clause and set out an interpretation. The determination of what an employee is owed is an employer decision with legal consequences, and the reasoning behind it belongs in the record. This site has covered the interpretation side of that separately; the point here is narrower and about what survives afterwards.
What never to automate
Do not let a system decide what is and is not a record. That classification is a legal judgement about prescribed content, and it should be made once, deliberately, by people who have read the regulations, then implemented in configuration.
Do not delete on a model's recommendation. Retention decisions on employee records are governed by a statutory period, and a tool optimising for storage or privacy hygiene has no visibility of that period unless somebody encodes it.
Do not rely on the reasonable excuse in subsection 557C(2) as a control. It exists for genuine misfortune. It is not a substitute for keeping the record.
Bottom line
Section 535 sets a seven-year obligation and section 557C prices the failure to meet it by handing you the burden of disproving whatever is alleged. AI has not changed either provision. What it has changed is where the underlying artefacts are produced and how long they survive, because tools built for a fast answer are frequently configured to forget. The organisations that will be fine are the ones that decided, in advance, which outputs are records and wrote them into the system that already keeps records for seven years.
Do this Monday
- List every pay or entitlement decision now touched by an AI tool, and name the record required in relation to each
- Compare the retention period of every system holding one of those artefacts against seven years, and flag every shortfall
- Confirm your AI and HR technology contracts address what happens to historical records at termination, not just current-state export
- Run one real retrieval test for a single team and a single fortnight, and time it
- Check that any workflow where a person approves an AI-generated record actually gives that person a way to verify it, rather than only a way to click approve
This article is general information for practitioners. It is not legal advice, and record-keeping obligations vary with the instrument, the employee and the regulations that apply. Refer to the current text of the Fair Work Act 2009 and the Fair Work Regulations 2009, or take advice, before acting.
Primary sources
- Fair Work Act 2009 (Cth), sections 535 and 536, current text via the Federal Register of Legislation. https://www.legislation.gov.au/C2009A00028/latest/text
- Fair Work Act 2009 (Cth), section 557C Presumption where records not provided, current text via the Federal Register of Legislation. https://www.legislation.gov.au/C2009A00028/latest/text
- Fair Work Ombudsman, Record-keeping. https://www.fairwork.gov.au/pay-and-wages/paying-wages/record-keeping
TheAICommand. Intelligence, At Your Command.
