Consultation Does Not Open the Algorithm, practitioner guidance from TheAICommand
← HR & AI
Workforce

Consultation Does Not Open the Algorithm

Australia's first minimum standards order commenced on 17 August 2026. It names algorithms twice, and both times to exclude them from disclosure, while leaving the consultation obligation intact. For HR teams consulting on an AI rollout, that is the clearest official illustration yet of where the confidentiality carve-out falls, and why it is no reason to run a thinner consultation.

People & Culture. Written for Australian HR and people teams. General information only. Not legal or HR advice. Employment decisions stay with people.

Quick answer

Australia's first minimum standards order commenced on 17 August 2026 and excludes platform algorithms from disclosure, twice. The consultation obligation itself survives in full. For HR, the lesson is to write a per-change disclosure schedule that separates what you will provide in writing from what you will withhold, and record the basis for each.

An instrument now names algorithms. Only to exclude them.

Australia's first minimum standards order commenced on 17 August 2026. Made by a Fair Work Commission Full Bench on 11 August 2026, the Interim On-Demand Delivery Employee-like Worker Minimum Standards Order is the first industrial instrument here written for a workforce whose work is allocated by an app. It carries a consultation obligation. It also says, twice, that the operator need not disclose "information about its platform such as any algorithms used therein".

The order does not bind you. It binds digital labour platform operators, and protects employee-like workers, who are independent contractors engaged through a platform under Part 3A-2 of the Fair Work Act, not employees. What it gives you is an illustration.

Every AI consultation eventually meets the same question from the floor: how does it decide? The honest answer is often that the model belongs to the vendor. The problem is what happens next, when "we cannot share the model" becomes the reason a pack that should carry four required items shrinks to a slide about efficiency. The carve-out is narrow. Everything on the other side of it does not move.

An editorial split showing a disclosure line falling between a change description and a closed system
The carve-out is one line in the clause. It is not the whole clause.

What commenced on 17 August 2026?

The Fair Work Ombudsman describes minimum standards orders as legally enforceable orders on pay and conditions for certain regulated workers and businesses, and says this is the first one issued.

Clause 5 is a consultation term, and not a thin one. It fires on a definite decision to exit a market, drop a category of delivery product, change the class of vehicle a worker must use, or permanently change the application's hours in an area, where that is likely to have significant effects. Clause 5.3 then requires written notice of the changes and their expected effect, a reasonable opportunity for feedback, and genuine consideration of it.

Then clause 5.5: "Clause 5 does not require a DLPO to disclose any confidential or commercially sensitive information including, but not limited to, information about its platform such as any algorithms used therein." Clause 6.8(b) repeats the exclusion for the platform feedback forum.

Two things limit the weight that carve-out can carry. No reasons for it appear anywhere: the clauses sat in the draft order reflecting the consent position of the Transport Workers' Union, Uber and DoorDash, and at paragraph 28 the Full Bench was not persuaded any other alteration of substance should be made to the draft. It added at paragraph 27 that "Interim" went into the title to underscore a foreshadowed review.

The Act behind it is silent too. The word "algorithm" does not appear anywhere in the Fair Work Act 2009, Compilation No. 73, in force 7 July 2026 and current as at 12 September 2026. Nor did the Commission have to include consultation: section 536KL(1) lists it among the matters an order may include terms about, making it permitted rather than required. It put a consultation term in an instrument that did not demand one, then drew the line inside it.

Why does a delivery order matter to your consultation clause?

Because your own duty carries the same two lines, and most teams read them as one.

The model consultation term (Schedule 2 of the Fair Work (Model Terms) Determination 2025, F2025L00202, in force 25 February 2025 and current as at 12 September 2026) requires the employer to "provide, in writing, to the relevant employees and their representatives (if any): all relevant information about the change including the nature of the change proposed; and the reasons or justification for the change; and information about the expected effects of the change on the employees; and any other matters likely to affect the employees". Immediately after comes the carve-out: "However, the employer is not required to disclose confidential or commercially sensitive information to the relevant employees or their representatives (if any)."

Four things you must provide in writing. One category you need not. They are separate obligations, and the second does not shrink the first. Whether a given item is confidential or commercially sensitive is a question of fact about that item, decided by you and defensible later. It is not a permission that arrives with the vendor contract.

Three boundaries, because each gets conflated with this one. Whether the clause fires at all is a different question, covered in the clause you already signed; this piece takes the duty as given and asks what it makes you hand over. The NSW power to inspect a digital work system is a safety statute running the other way, opening the system to a permit holder. And what RG 271 reserves for a person is an ASIC guide on consumer complaints, not an industrial obligation.

How do you write a per-change disclosure schedule?

One page per change. Forty minutes, and it replaces an argument with a record.

  1. Fix the change and the instrument. Name [SYSTEM_NAME], the affected group, and the clause that binds you, quoted rather than paraphrased. Silence in your agreement imports the model term.
  2. Draft the four required items before considering any withholding. Nature, reasons, expected effects, other matters likely to affect employees. Writing these first stops the carve-out becoming the starting point.
  3. List each item you intend to withhold on its own line. Not "the model". The weights, the training data, the scoring thresholds, the contract price. A category cannot be assessed. An item can.
  4. State the basis and the decider for each. Confidential or commercially sensitive is the test. Name who applied it, the date, and one sentence of why. An unattributed withholding is the weakest line in the record.
  5. Write what travels instead. This step keeps the consultation whole: decision logic in plain words, the inputs, what is measured, the override route, the review route.
  6. Set a review date and record what came back. The term requires prompt and genuine consideration, and reasonable steps to communicate the outcome.
ItemDispositionWhat travels instead
What [SYSTEM_NAME] does, and to which workProvide in writingNot applicable
Reasons or justification for the changeProvide in writingNot applicable
Expected effects on roles and hoursProvide in writingNot applicable
Other matters likely to affect employeesProvide in writingNot applicable
Model weights, architecture, training dataAssess item by itemDecision logic in plain words, and the inputs used
Vendor pricing and commercial termsAssess item by itemThe effect on roles, workload and hours
Proprietary scoring thresholdsAssess item by item, never as a blockWhat is measured, and what triggers human review
Security configuration and access controlsWithholdWho can see outputs, and how access is reviewed

A worked example. [SYSTEM_NAME] will rank and route inbound requests for [TEAM], work the team currently picks up in order of arrival. In the consultation session, [EMPLOYEE_NAME] asks how the ranking works.

The wrong answer is that it is proprietary. The schedule already holds the right one. The weights are withheld, assessed as commercially sensitive by [MANAGER_NAME] on the recorded date. What travels instead: the inputs the system reads, the plain statement that it ranks by predicted handling time and stated urgency, the threshold that sends an item to a person, the override route, and the review date. Nobody saw the model. Everybody can now argue about the right things.

What the order's automated design shows, and what you cannot copy

The same order is an automated-first employee relations architecture a tribunal has signed off. Clause 5.3 says the feedback obligation is met by a reasonable opportunity to give feedback "via email, a digital or online platform and/or via the application", and requires genuine consideration "at an aggregate level". Clause 6.3 says the platform feedback forum "is not a decision-making body". Clause 10.2 lets a first-stage dispute run the same way.

Two parties pushed back on this. One asked for a human review pathway for pay disputes rather than automated determination alone; another asked the dispute procedure provide for timeframes and human involvement. The Full Bench made no substantive change and parked such proposals for the review. That is not a finding that human review is unnecessary.

Here is the part HR cannot lift. Section 536KK requires a minimum standards order to include "a term that provides a procedure for settling disputes about any matters arising under the order". That is the whole section: no independence requirement, no representation requirement. Section 186(6) holds an enterprise agreement to a higher bar, requiring a term that allows the Commission or another independent person to settle disputes, and that allows representation. Digital intake is portable. Losing the independent settler is not.

The governance line

Keep the consideration standard straight. The order permits consideration "at an aggregate level". Your model term requires prompt and genuine consideration of matters raised. That is a per-matter human act, and a record that reads as machine-assembled invites the argument that nobody considered anything.

Privacy sits alongside, not inside, this obligation. If the tool observes how people work, how it handles that data is a separate duty the pack does not discharge. Fairness deserves its own line: if a system ranks or allocates, the group most likely affected unevenly belongs in the expected-effects section, not a complaint.

AI is useful here. It can draft the four required items from your change documentation and flag anything withheld without a stated basis. It cannot decide what is commercially sensitive, and it cannot perform the consideration. TheAICommand works to the Verified Draft Method: de-identify the inputs, ground the model in your own source material, keep a person at the decision point, verify against the primary source, and log what happened.

None of this is legal advice. The instrument that binds you is yours, not the one quoted here.

What never to withhold or automate

  • Never treat the carve-out as covering the four required items. They are required in writing whatever sits behind the interface.
  • Never withhold a category. "The system is proprietary" is not an assessment. Item by item, with a basis.
  • Never leave a withholding unattributed. A named person, a date and one sentence turn a refusal into a decision.
  • Never let the vendor set your disclosure line. The confidentiality interest is theirs. The obligation is yours.
  • Never automate the genuine consideration of feedback. The thinking is the obligation.
  • Never paste identifiable employee data into a general tool. Use [EMPLOYEE_NAME] and [TEAM].
  • Never present the order as binding your workforce. It illustrates a line, not a source of duty.

Bottom line

The first industrial instrument written for an app-managed workforce mentions algorithms only to exclude them from disclosure, and leaves a full consultation obligation beside them. That is the shape of the rule HR already lives under: the model term names four things you must provide in writing, then separately says you need not disclose confidential or commercially sensitive information. The second line does not shrink the first.

Do this Monday

  • Mark the two lines in your own consultation clause, so the written-information limb and the confidentiality carve-out stop reading as one sentence.
  • Take the AI change closest to go-live and draft the four required items first, before anyone raises what cannot be shared.
  • List the withheld items individually, with a named decider and a date. No categories, no unattributed refusals.
  • Write the "what travels instead" column: decision logic, the inputs, what is measured, the override route, the review route.
  • Set the review date in the pack, and diarise what was raised and what changed as a result.
  • Check the adjacent duties separately. Privacy, and in NSW the digital work system provisions, are different sources.

References

  1. Fair Work Commission, Interim On-Demand Delivery Employee-like Worker Minimum Standards Order [MS900103], PR813107, made 11 August 2026, clauses 1.2, 5.1 to 5.5, 6.3, 6.8(b) and 10.2. https://www.fwc.gov.au/documents/sites/ms2024-1/ms2024-3-order-mso-11-08-2026.pdf
  2. Fair Work Commission Full Bench, [2026] FWCFB 211, 11 August 2026, at [12], [17], [27], [28]. https://www.fwc.gov.au/documents/decisionssigned/pdf/2026fwcfb211.pdf
  3. Fair Work Ombudsman, New minimum standards for on-demand delivery workers, 13 August 2026. https://www.fairwork.gov.au/about-us/workplace-laws/fair-work-commission-orders/minimum-standards-order-on-demand-delivery-workers
  4. Fair Work (Model Terms) Determination 2025, F2025L00202, Schedule 2, in force 25 February 2025. https://www.legislation.gov.au/F2025L00202/2025-02-25/2025-02-25/text/original/epub/OEBPS/document_1/document_1.html
  5. Fair Work Act 2009 (Cth), Compilation No. 73, C2026C00355, in force 7 July 2026, ss 186(6), 536KK and 536KL. https://www.legislation.gov.au/C2009A00028/latest/text

TheAICommand. Intelligence, At Your Command.

Frequently asked questions

Does the new minimum standards order apply to our employees?
No. It binds digital labour platform operators and protects employee-like workers, who are independent contractors engaged through a platform under Part 3A-2 of the Fair Work Act. It commenced on 17 August 2026 and creates no obligation for an ordinary employer. Its value to HR is as an illustration of where a disclosure carve-out falls, not as a source of duty.
Can we refuse to share how an AI tool makes decisions during consultation?
The model consultation term says the employer is not required to disclose confidential or commercially sensitive information. Whether a specific item meets that description is a question of fact, not a standing permission. Decide it item by item, record who decided and why, and keep providing the four things the term requires in writing regardless of what you withhold.
What must we provide in writing under the model consultation term?
Schedule 2 of the Fair Work (Model Terms) Determination 2025 requires the employer to provide in writing all relevant information about the change including the nature of the change proposed, the reasons or justification for the change, information about the expected effects of the change on the employees, and any other matters likely to affect the employees. The carve-out sits separately and does not reduce those four.
Did the Fair Work Commission decide that platforms can keep their algorithms secret?
The order says the operator need not disclose information about its platform such as any algorithms used therein, in clause 5.5 and again in clause 6.8(b). The decision gives no reasons for it. Those clauses sat in the draft order the parties had consented to, and the Full Bench declined to make any substantive alteration. The order is interim and a review is foreshadowed.
Can we run grievances through an app the way the order allows?
Partly. The order lets a first-stage dispute be pursued by email, a digital or online platform or the application. An enterprise agreement is held to a different standard. Section 186(6) of the Fair Work Act requires a dispute term allowing the Commission or another independent person to settle disputes, and allowing representation. Digital intake is fine. Removing the independent settler is not.
ConsultationEmployee RelationsFair WorkEnterprise AgreementsAI GovernanceAlgorithmic ManagementChange ManagementWorkforce
← Back to HR & AI