"Agent washing" is the habit of labelling ordinary software, scripts or outsourced labour as AI agents. If you buy agentic delivery, you need a way to tell the real thing from the label. This article gives a checklist of questions and the evidence to ask for, and answers each one for us where the answer is standing policy.
This is not legal advice. Where we touch consumer law, we say what has been reported and leave the interpretation to you and your lawyers.
What has been said about agent washing
Gartner is reported to have forecast, in June 2025, that over 40% of agentic AI projects will be cancelled by the end of 2027, citing cost, unclear value and weak risk controls. The same report is said to estimate that only about 130 of the thousands of vendors claiming agentic products are real, and to use the term agent washing for the rest. We have this from a secondary report (CRN), not from Gartner directly, so treat the figures as reported, not confirmed.
The point does not depend on the exact numbers. A label is not evidence, and the cost of finding that out late is high.
The consumer law angle, carefully
Australian Consumer Law prohibits misleading or deceptive conduct. Law-firm commentary reports that the ACCC identified AI washing as a consumer risk in December 2025 and flagged systems wrongly promoted as agentic (MinterEllison, MK Law). We could not confirm that at a primary ACCC source, and we found no ACCC case against a named company. Commentary also points to the ACCC's greenwashing guidance as a practical benchmark: have evidence, and avoid broad claims with no qualification.
Overseas, regulators have acted. The US Securities and Exchange Commission announced charges on 18 March 2024 against two investment advisers over unsubstantiated AI claims (O'Melveny summary), and the US Federal Trade Commission settled with DoNotPay over a "robot lawyer" claim that was not tested (ABA Journal). These are not Australian precedents. They do show what unsupported claims look like to a regulator: claims with no testing behind them.
For a buyer, the useful reading is practical. If a supplier's claim matters to your decision, ask for the evidence behind it, and keep the answer.
The checklist
Who and what
- Who is accountable? Ask for a named person, in the contract. *Our answer:* a named human lead owns each engagement and signs the deliverables.
- What exactly is the agent? Ask which tasks a software agent performs, which a person performs, and which models and tools are involved. A supplier who cannot separate these may not know.
- What will you not hand to an agent? Real delivery has limits. The Australian Signals Directorate's joint guidance of 1 May 2026 is reported to advise keeping agents to low-risk tasks until practices mature. A supplier with no exclusions is a warning sign.
Disclosure and approval
- Will AI use be disclosed on each deliverable? *Our answer:* yes, on every deliverable, stating what the agents did.
- Will you seek approval before using AI on my work? For Commonwealth buyers, the DTA AI Model Clauses are reported to require notice and prior written approval when a supplier plans to use AI to deliver services. *Our answer:* we seek prior written approval before using AI on your work.
Evidence
- Show me a sample deliverable. Ask for a real or clearly labelled synthetic example with the disclosure line and review record. Slides are not evidence.
- Show me how you review. Ask who checks agent output, against what, and where the record sits.
- How do you test? For code, ask what must pass before handover. For analysis, ask how claims are tied to sources.
- Where does my data go? Ask per component: model provider, region, who has access. If the answer is a slogan such as "sovereign" or "secure", ask for the documents. We make no certification claims in this article.
- Can I see the method behind any number? Treat "ten times faster" or "percent saved" with no published method as unsupported.
Commercial terms
- Is the scope fixed? *Our answer:* scope and acceptance criteria are fixed up front, so you can tell whether the work was delivered.
- What are the acceptance criteria, and who signs? They should be testable. If they are vague, the label is doing the work.
- What do I get at handover? Ask whether outputs come in standard formats and whether anything depends on the supplier's own runtime.
- What happens if an agent is wrong? The honest answer is that it can be, and that review and acceptance are the controls. Be cautious of "zero errors" or "guaranteed ROI".
Words that should prompt a follow-up question
None of these is proof of washing. Each should trigger a request for evidence.
- "Fully automated" or "hands-free"
- "Replaces your team"
- "Guaranteed ROI" or "zero hallucinations"
- "Enterprise-grade" with no definition
- Compliance claims with no document: IRAP, PSPF, Essential Eight, ISO
- A client list you cannot verify
Using the checklist
Send the questions in writing before you shortlist, and compare answers side by side. The strongest signal is rarely a single answer. It is whether the supplier replies in specifics, names a person, and offers a small fixed-scope pilot with acceptance criteria before asking you to commit to a large one.
Run it on us too. We would rather answer these questions before a contract than after.
How this article was written
Drafted with AI agents and reviewed by xShopper. Facts are linked to their sources; anything we could not confirm at a primary source is marked as such.