INSIGHT · AI GOVERNANCE & COMPLIANCE

Buying an AI estimating tool:
the questions are in the document
that was superseded.

4 November 2026

Australia's national AI guidance for an organisation buying a tool sits in two live documents from the same publisher. The Voluntary AI Safety Standard sets out 10 guardrails, two of which say what a supplier must tell you and what you must record. The Guidance for AI Adoption, published 21 October 2025, sets out 6 essential practices, and every page of the older standard now opens by saying this "updated and simplified guidance for industry evolves the Voluntary AI Safety Standard". Neither says which one a buyer should apply, and the detailed supplier-disclosure clause exists only in the one described as evolved.

Stylised isometric illustration of a document slab passing over a measuring rail, one dimension marked by a glowing caliper, a second slab dimmed behind it

Analysed 4 November 2026 · Voluntary AI Safety Standard, The 10 guardrails and Introduction to the standard, industry.gov.au · Guidance for AI Adoption, ai.gov.au, published 21 October 2025 · OAIC, Guidance on privacy and the use of commercially available AI products · all pages read 4 November 2026.

Key takeaways

Two documents, both live, one said to evolve the other: Every page of the Voluntary AI Safety Standard carries the line: "On 21 October 2025, we published the Guidance for AI Adoption, which outlines 6 essential practices for safe and responsible AI governance. This updated and simplified guidance for industry evolves the Voluntary AI Safety Standard." The standard's own pages returned HTTP 200 on 4 November 2026 and are not marked withdrawn. Read 4 November 2026.
The supplier-disclosure clause is in the older one: Guardrail 8: "Be transparent with other organisations across the AI supply chain about data, models and systems to help them effectively address risks." It states that organisations "must provide information to organisations downstream in the AI supply chain for them to understand the components of the AI system, how it was built and to understand and manage the risk of the use of the AI system." You are downstream of your estimating vendor.
So is the inventory requirement: Guardrail 9: "Keep and maintain records to allow third parties to assess compliance with guardrails." It asks for "an up-to-date, organisation-wide inventory of each AI system" including "capabilities and limitations of the AI system" and "datasets and their providence used for training and testing".
Accuracy is already a legal obligation where personal information is involved: The OAIC states that "Under APP 10, organisations have an obligation to take reasonable steps to ensure the personal information collected, used and disclosed is accurate", and that this risk "is especially high with generative AI, which is probabilistic in nature". That is law, not guidance, and it does not depend on which of the two AI documents you follow.
The two documents moved to different websites: The standard is on industry.gov.au. The guidance that evolved it now answers on ai.gov.au: the old address, industry.gov.au/publications/guidance-ai-adoption, returned HTTP 301 to ai.gov.au on 4 November 2026. A citation to either needs the domain and the read date.
We would record what the vendor answered, and when: Our system-design recommendation: per tool, store what the supplier disclosed against a fixed question list, what they did not answer, who is accountable for the output inside your business, and the date each answer was given.

What a buyer is actually
looking for.

An AI estimating tool reads drawings and returns quantities. Those quantities go into a quote. The quote becomes a price someone is held to. So the buyer's question is not whether the tool is impressive; it is what happens to the number between the model and the contract, and what the supplier is willing to put in writing about it.

The standard distinguishes the two roles: The Voluntary AI Safety Standard's introduction states that the guardrails "explain what developers and deployers of AI systems must do to comply with the guardrails", and that while the first version "applies to both AI deployers and AI developers, it focuses on providing guidance at the organisational and system level for AI deployers". A builder buying a tool is a deployer.
Which puts the disclosure duty on the other side: Guardrail 8 is written for the party upstream of you. It commits organisations to "sharing information and establishing processes to provide sufficient transparency between developers and deployers of AI systems". The buyer's leverage is asking for it before signing, because afterwards it is a request rather than a term.
And leaves the output with you: Guardrail 5: "Enable human control or intervention in an AI system to achieve meaningful human oversight." It asks organisations to "assign accountability for each AI system to someone who shows suitable competence", with "sufficient authority to oversee, intervene and be effective". Nothing in it transfers responsibility for a quote to a supplier.

The two documents,
side by side.

Both were read on 4 November 2026. The comparison is of what each publishes, not of which one governs: neither states that.

The newer guidance does carry the same subjects. What it does not carry, on the pages read, is the clause-level detail: the standard lists what an inventory entry should contain and what a developer should disclose to a deployer, and the guidance names the topics. A buyer drafting a question list gets more from the document that was evolved than from the one that evolved it. The site's reading of the six practices themselves is a separate article.

	Voluntary AI Safety Standard	Guidance for AI Adoption
Structure	10 guardrails	6 essential practices
Where it answers	industry.gov.au	ai.gov.au (the industry.gov.au address 301s there)
Date it carries	No publication date on the pages read	"On 21 October 2025, we published..."
Status shown	Live, not marked withdrawn; each page opens with the line that the newer guidance "evolves" it	Presented as the current guidance
Supplier disclosure	Guardrail 8, with sub-clauses on developer-to-deployer transparency	"supply chain accountability" listed among organisation-wide practices
Records	Guardrail 9, with a specified inventory content list	"AI registers" listed among organisation-wide practices
Third-party systems	Addressed through guardrails 8 and 9	"procurement, development or deployment of third-party AI systems"; risk assessment covers "systems developed or procured from third party suppliers"

The six essential practices

What Guardrail 9 asks you
to be able to produce.

Guardrail 9 states that organisations "must maintain records to demonstrate that they have implemented and are complying with the guardrails, this includes maintaining an AI inventory and consistent AI system documentation", and adds that "These records may be required to input into any future conformity assessments mandated for use in high-risk settings." The inventory content it names, quoted from the page:

Guardrail 4 sits behind this: "Test AI models and systems to evaluate model performance and monitor the system once deployed", performed "according to the clearly defined acceptance criteria that considers the prior risk and impact assessment". Acceptance criteria are the buyer's to define. A tool cannot pass a test that was never written down.

people accountable: Named, not a team. Guardrail 5 asks for someone with "suitable competence" and "sufficient authority to oversee, intervene and be effective".
purpose and business goals: What the tool was brought in to do. A takeoff tool approved for internal checking is a different record from one approved to produce a figure that leaves the business.
capabilities and limitations of the AI system: The limitations half is the one a vendor is least likely to volunteer and the one a buyer most needs before a number reaches a quote.
technical requirements and components: Including what it runs on and what it connects to.
datasets and their providence used for training and testing: Quoted as published. For an estimating tool this is the question of what drawings it learned from and what it was measured against.

Where privacy law
enters an estimating workflow.

Estimating looks like a quantities problem, not a privacy one. It stops looking that way as soon as the documents fed to the tool carry names: a subcontractor schedule, a site contact list inside a drawing set, a scanned contract. The OAIC's guidance on commercially available AI products addresses that directly.

Nothing here determines whether any particular workflow handles personal information, whether APP 10 applies to any business, or what reasonable steps would be. Those are matters for the organisation and its advisers. The point is narrower: an estimating tool is not outside privacy law merely because its subject is quantities.

Due diligence before adoption: The OAIC states that "When looking to adopt a commercially available product, organisations should conduct due diligence to ensure the product is suitable to its intended uses", including "considering whether the product has been tested for such uses, how human oversight can be embedded into processes, the potential privacy and security risks".
Accuracy is an obligation, not an aspiration: "Under APP 10, organisations have an obligation to take reasonable steps to ensure the personal information collected, used and disclosed is accurate." The OAIC notes AI systems "are known to produce inaccurate or false results" and that the risk "is especially high with generative AI, which is probabilistic in nature and does not 'understand' the data it handles or generates".
Where the system runs matters: The OAIC observes that "Deploying a system locally or on premises is likely to be more privacy-preserving as it limits the risks of third party access to the data." For a buyer that is a procurement question with a deployment answer, asked before the contract rather than after.

What we would put in a system

The point of contact is the vendor conversation, and then every quote the tool touches afterwards.

Where agents help is narrow. A document intelligence agent can read a supplier's material against a fixed question list and draft which of the Guardrail 9 inventory items are answered and which are not. A compliance agent can draft the list of quotes issued while a given tool version was in use. A person decides whether the tool may be used, on what, and who signs the number. No agent decides whether a quote is correct, whether a price may be relied on, or whether any obligation is met. Those are matters for the business and its advisers.

Field	Why it is load-bearing
Tool, supplier and version, with the date recorded	Guardrail 9 asks for an inventory of each AI system. A version is what makes a later answer comparable to an earlier one.
Stated capabilities and stated limitations, in the supplier's words	Guardrail 9 names both. Recording the supplier's own wording is what makes the claim checkable later.
What the supplier disclosed, and what they did not answer	Guardrail 8 puts the disclosure duty upstream. An unanswered question is a finding, and only a record shows it was asked.
Acceptance criteria, written before the trial	Guardrail 4 asks for testing against "clearly defined acceptance criteria". Criteria written after a result are not criteria.
The named person accountable for the output	Guardrail 5. The quote carries a business's name; the record should carry a person's.
Whether outputs may leave the business unreviewed, and who decided	This is the boundary that turns a drafting aid into a priced commitment.
Whether the documents fed in carry personal information, and where the system runs	The OAIC's due diligence and deployment-location points, as a field rather than a recollection.
Which quotes relied on the tool, and under which version	If a limitation surfaces later, this is the only field that scopes the problem.

Quoting agent · How we deploy AI

Questions worth asking
of your own tooling

For a builder or construction company: if an AI tool produced a quantity that went into a quote last month, can you name the tool version and the person who approved the number?

Guardrail 9 asks for an inventory of each AI system and Guardrail 5 for a named accountable person. If the answer comes from asking an estimator what they remember, neither exists as a record.

For a head of estimating: what are the acceptance criteria the tool was measured against, and were they written before or after the trial?

Guardrail 4 asks for testing "according to the clearly defined acceptance criteria that considers the prior risk and impact assessment". Criteria written after a result describe the result.

For a procurement or commercial manager: what did the supplier decline to answer?

Guardrail 8 places the transparency duty on the organisation upstream of you. The unanswered questions are the finding, and they only exist if someone wrote down that they were asked.

For a trade or specialist contractor: does anything you feed the tool contain a person's name?

Subcontractor schedules, contact blocks inside drawing sets and scanned contracts often do. The OAIC's guidance on commercially available AI products applies to the handling of personal information whatever the tool is nominally for.

For a director: which of the two national documents does your AI policy cite, and on what date was that page read?

Both are live, the newer one is described as evolving the older, and they answer on different websites. A policy citing either without a read date cannot be checked against what the page says now.

Questions people ask
about AI estimating tools

Is AI construction estimating regulated in Australia?

The documents read here are guidance, not regulation. The Voluntary AI Safety Standard is voluntary on its face. Obligations that do apply come from elsewhere, for example the Privacy Act's APP 10 accuracy obligation where personal information is handled. Whether any particular obligation applies to a business is a matter for that business and its advisers.

What should I ask an AI estimating vendor?

The content list in Guardrail 9 is the closest thing to a published question list: people accountable, purpose and business goals, capabilities and limitations, technical requirements and components, and datasets and their provenance used for training and testing. Guardrail 8 is the clause that says a supplier should be telling you.

Can an AI tool produce a quote my business is bound by?

Nothing in the documents read transfers responsibility for an output to a supplier. Guardrail 5 asks for a named person with authority to oversee and intervene. What a quote commits a business to is a contract question for the business and its advisers.

Does the Voluntary AI Safety Standard still apply?

Its pages were live and not marked withdrawn on 4 November 2026, and each opens by stating that the Guidance for AI Adoption of 21 October 2025 "evolves" it. Neither document says which one an organisation should apply. We report both rather than choosing between them.

Where is the Guidance for AI Adoption published?

On ai.gov.au as read on 4 November 2026. The industry.gov.au address for it returned HTTP 301 to that location on the same day.

What this analysis
does and does not show.

As read on 4 November 2026

Documents: Voluntary AI Safety Standard (10 guardrails) on industry.gov.au; Guidance for AI Adoption (6 essential practices, 21 October 2025) on ai.gov.au
Relationship as stated: Each standard page opens: the newer guidance "evolves the Voluntary AI Safety Standard"
Standard's status: HTTP 200, not marked withdrawn
Old guidance address: industry.gov.au/publications/guidance-ai-adoption returned HTTP 301 to ai.gov.au
Supplier disclosure: Guardrail 8, developer-to-deployer transparency
Records: Guardrail 9, AI inventory with a named content list
Accuracy obligation: OAIC: APP 10 reasonable steps to ensure accuracy of personal information

Both AI documents are guidance published by the National AI Centre and the department. Neither is legislation, and neither states which of the two an organisation should apply. We report what each says and do not resolve the relationship.
We did not download the PDF copies of either document. Every quotation is from the web pages named, as read on 4 November 2026.
The comparison in section 02 is of what each document publishes on the pages read. The newer guidance may address the same matters in more detail elsewhere on its site or in its downloadable form, which we did not read.
Nothing here determines whether any AI tool is suitable for any purpose, whether any quote may be relied on, whether any business handles personal information, or whether any obligation under the Privacy Act applies. Those are matters for the business and its advisers.
No tool, vendor or product is named, tested or assessed here, and no claim is made or repeated about any tool's accuracy.
This reading is about buying and deploying a tool. What to record once an AI tool has touched a decision is a separate article, and what a licensed person may sign is another.

What to watch next

Only what the publishers have themselves stated.

Guardrail 9 points at conformity assessments that do not yet exist: It states the records "may be required to input into any future conformity assessments mandated for use in high-risk settings". No such mandate is cited on the pages read, and no date is given.
The two documents may converge, or one may be retired: The standard is described as evolved but remains published. Whether its clause-level detail survives into the newer guidance is a question a reader can only answer by rereading both. That is the case for recording which one a policy cites, and when it was read.

Bring us one tool
and the questions you asked it.

Tell us which AI tool touches your quantities or your pricing, and what the supplier told you. We will show you which Guardrail 9 inventory items your record would already answer, which ones it would not, and what a per-quote use record would have held.

This article is general information about published Australian Government AI guidance and OAIC privacy guidance, and about how records of tool selection are structured. It is not legal, privacy or procurement advice. It does not assess any product, does not determine whether any obligation applies to any business, and names no vendor. All pages are as read on 4 November 2026 and can change.

Send us the tool · How we deploy AI

Sources

Suggested citation: Voluntary AI Safety Standard (10 guardrails) and Guidance for AI Adoption, with OAIC guidance on commercially available AI products, as read 4 November 2026. No figure in this reading is derived by us.

Buying an AI estimating tool: the questions are in the document that was superseded.