AI in employment law: unfair dismissal, awards and Fair Work detail

By Michael Nadalin, Founder, Market Lead · 9 October 2026 · 8 min read

A dismissed employee has 21 days from the date the dismissal takes effect to lodge an unfair dismissal application with the Fair Work Commission. That is the number that shapes employment practice. Instructions arrive late, the client's documents are in a phone, and the first real question is not whether the dismissal was harsh, unjust or unreasonable but whether the person is protected from unfair dismissal at all.

This is the kind of work where AI looks like an obvious fit: short deadlines, high volume, modest amounts in dispute, and a statute with a structured set of tests. It is also the kind of work where a general purpose chatbot does real damage, because employment law is full of figures that change every year, instruments that are varied constantly, and provisions that were amended recently enough for a model to describe the old law with complete confidence.

What follows is a practitioner's view of where AI is useful on a Fair Work file, where it is unreliable, and what checking looks like when you are the one whose name is on the document.

Where the time actually goes on an unfair dismissal file

Most unfair dismissal matters resolve at or around conciliation. That means the economically important work happens early: taking instructions, testing jurisdiction, building a chronology, and forming a realistic view of the merits and remedy before the conference. Compensation is capped by the Fair Work Act at the lesser of 26 weeks' pay and half the high income threshold, so there is rarely room for many hours of unrecoverable research.

AI helps most in the compressed front end. Turning a bundle of emails, warning letters, rosters and text messages into a dated chronology is tedious and mechanical, and a model given the documents does it quickly. The same goes for a first pass at the application or the employer response, where the structure is set by the Commission's forms. What AI cannot do is tell you the date the dismissal took effect when the facts are contested, and that date is the one everything else runs from.

The jurisdictional gateway is a checklist, and checklists suit machines

Before the merits, the applicant has to get through a series of gates. Each is a discrete statutory question with its own evidence. This is a good use of AI as a prompt for your own thinking: give it the instructions and ask it to work through each gate, identify which facts are missing, and list the documents that would prove each one.

The caution is that several gates turn on details a model will happily assume. Whether an employer is a small business employer depends on a headcount that includes associated entities and certain casuals. Whether a casual's service counts toward the minimum employment period depends on the nature of the engagement. Whether a redundancy is genuine depends on consultation obligations in the applicable award or agreement and on redeployment across the enterprise and its associated entities. A model will give a clean answer to each. You need it to give you the question instead.

Awards are where general AI tools go wrong most often

Award work is the hardest test for any AI system. A modern award is a long instrument with a coverage clause, classification definitions, penalty and overtime rules, allowances, and schedules. It is varied regularly. Minimum rates move after each Annual Wage Review. A model that learned the instrument from its training data is describing a past version, and it will not tell you which one.

The errors are predictable. The tool picks an award from the job title when coverage depends on the employer's industry and the work actually performed. It quotes a rate that was correct two reviews ago. It applies a penalty rate without checking the span of hours or whether the employee is a shiftworker under that award's definition. It ignores an enterprise agreement that displaces the award altogether. In an underpayment matter each of these compounds across every pay period in the claim.

The same lag affects the Act itself. The Closing Loopholes amendments changed the casual employee definition, added a statutory definition of employment, introduced the right to disconnect and criminalised intentional underpayment. A tool that is not reading the current compilation may explain the earlier position fluently.

Ranking employment tasks by how much you can trust the output

It helps to stop asking whether AI is good at employment law and ask which task you are handing over. The dividing line is simple. Where the tool is working on documents you gave it, it is reorganising known material and errors are easy to spot. Where it is supplying law, figures or authorities from memory, it is generating plausible text and errors look exactly like correct answers.

Applying the harshness criteria in section 387 sits in the middle. A model will produce a tidy analysis against each factor: valid reason, notification, opportunity to respond, support person, warnings, the size of the enterprise and its access to human resources expertise, and other relevant matters. That is a useful structure for a conciliation outline. It is not a merits assessment, because the weight given to each factor is a matter of judgment on evidence, and the tool has only heard one side.

A verification routine that fits a fixed fee file

Verification has to be proportionate or it will not happen. On a matter where the fee is modest, the routine below adds minutes and it covers the errors that cause real trouble: a missed time limit, a wrong award, a stale rate, an authority that does not exist.

Two further points. Employment files are full of sensitive material: personnel records, medical certificates, complaints about named co-workers. Do not put that into a tool unless you know where the data goes, whether it is retained, and whether it is used for training. Your confidentiality obligations do not change because the software is convenient. And several Australian courts have published practice notes or guidelines on generative AI. If the matter is heading to a court as well as the Commission, such as a general protections or underpayment claim, check what applies in that forum before you prepare affidavits or submissions.

What to ask of a tool before you use it on a Fair Work matter

The questions are the same ones you would ask a junior. Where did that come from? Is that the current version? Have you read the case, or just the headnote? A tool worth using on employment work answers from Australian primary sources, shows the provision or passage it relied on so you can check it in seconds, and says so plainly when it cannot find support for a proposition. One that answers from memory and never admits a gap is not suitable for this area, however fluent it sounds.

That is the approach behind Legal Brain: research grounded in Australian legislation and case law, with sources shown so the checking is quick. Whichever tool you choose, the test is practical. Take a matter you finished last year, where you already know the award, the classification and the outcome, and run it through. If the tool gets the coverage question wrong or cites something you cannot find, you have your answer before a client pays for it.

Used well, AI takes hours out of the chronology, the first draft and the conciliation outline, which is where a low value file loses money. It does not replace reading the award, and nothing in the current generation of tools suggests it will.

Frequently asked questions

Can I use ChatGPT to work out which modern award applies to an employee?

You can use it to suggest candidates, but not to decide. Award coverage depends on the coverage clause, the employer's industry or the employee's occupation, and the classification definitions applied to the duties actually performed. A general chatbot tends to guess from the job title and may be working from an outdated version of the award, so read the current instrument yourself.

How long do you have to lodge an unfair dismissal claim in Australia?

An application must be lodged with the Fair Work Commission within 21 days after the dismissal took effect. The Commission can extend time only where there are exceptional circumstances, which is a demanding test. The same 21 day limit applies to a general protections application involving dismissal.

Is it safe to put client employment documents into an AI tool?

Only if you know how the tool handles data. Check whether inputs are stored, who can access them, where they are hosted and whether they are used to train the model. Employment files often contain health information and allegations about third parties, and your confidentiality obligations apply regardless of the software you use.

Will AI replace employment lawyers for unfair dismissal work?

Not on present evidence. AI is good at organising documents and producing first drafts, but the decisive questions are matters of evidence and judgment: when the dismissal took effect, whether there was a valid reason, and what a realistic settlement looks like. Tools that supply law from memory also make errors on rates, thresholds and recent amendments that a lawyer has to catch.