An AI Vendor Due Diligence Checklist

AI vendor due diligence means checking an AI product before adopting it: what it does with your data, whether it trains on it, how it is secured, what access it needs, what testing backs it, and how you leave. Australia's privacy regulator expects this check, and says it should not be "set and forget".

This is general information, not legal advice.

What is AI vendor due diligence?

AI vendor due diligence is the set of checks a business runs before buying or connecting an AI product: whether it suits the intended use, how it handles data, how it is secured, what evidence supports its claims and how the business can leave. It extends normal software procurement with questions specific to AI.

Those extra questions exist because AI products behave differently from ordinary software. They may keep what you enter, use it to improve models, produce confident errors, or act inside connected systems. A standard security questionnaire misses most of that.

For a growing business, due diligence should be proportionate. A note-taking tool used on internal meetings needs a short check. An agent that will read the CRM and draft customer emails needs a thorough one.

What does Australian guidance expect?

The Office of the Australian Information Commissioner expects organisations to conduct due diligence to make sure an AI product suits its intended uses, and says this should not be "set and forget". The National AI Centre's guidance adds: ask the supplier for proof the system has been properly tested, and make clear who is responsible across the supply chain.

According to the OAIC, due diligence should consider whether the product has been tested for your uses, how human oversight can be embedded, the privacy and security risks, and who will have access to personal information. The OAIC lists specific questions, several of which appear in the checklist below.

According to the National AI Centre's Foundations guidance, a business buying an AI system should "ask the developer or supplier to show proof that it's been properly tested". The guidance also asks businesses to clarify supply chain accountabilities, so they know who to talk to when something goes wrong.

The privacy obligations behind these checks are covered in the guide to AI and privacy in Australia.

What questions should you ask an AI vendor?

Ask questions in seven areas: fit for purpose, data use, training, security and access, proof of testing, human oversight, and exit. Each area has two or three questions worth answering before you sign. Get the answers in writing, ideally in the contract or the vendor's published terms, not only in a sales call.

Several of these come straight from the OAIC's list, including what data the product has been trained and tested on, what its data flows are, and whether third parties can access what you input or generate.

Exit questions get skipped most often. Yet if the knowledge, prompts and workflows your team builds live only inside one vendor's product, leaving means starting again.

How do you check how a vendor uses your data?

Read the terms for the specific plan you will buy, not the product's general reputation. Consumer and business plans of the same tool often differ on training, retention and admin control. Check whether training on your data is off by default, how long data is kept, and whether an administrator can enforce the settings.

Anthropic's terms show why the plan matters. Its August 2025 consumer terms update lets Free, Pro and Max users choose whether chats are used to improve Claude, with retention up to five years if allowed. The changes "do not apply to services under our Commercial Terms", which cover Team, Enterprise and the API.

The same pattern runs across most AI vendors. So the due diligence question is never "is this tool safe?" but "is this plan, configured this way, suitable for this data?" The guide to whether Claude is safe for business works through one vendor in detail.

Check connections separately. A product that connects to your CRM, inbox or drive may reach far more data than anyone types into it.

How do you test a vendor's claims?

Ask for evidence, then run your own trial on real work before committing. Give the product three or four tasks the team actually does, with the information it would really have, and compare the output against what a person would produce. Watch especially for confident errors, because plausible mistakes are the ones nobody catches.

Vendor claims deserve scepticism in this market. Gartner estimates that only about 130 of the thousands of vendors marketing "agentic" products are real, a practice it calls "agent washing". The label on the product says little about what it does.

A good trial is short and specific:

1. Choose real tasks. Pick recurring work, not a showcase example the vendor suggests.
2. Use real context. Give it the documents and data it would have in practice, within your data rules.
3. Score the output. Note accuracy, rewrites needed and time saved.
4. Test the edges. Ask it about things it should not know, and see whether it says so or invents an answer.

Before buying anything, check whether a tool you already pay for can do the job. The guide to build, buy or switch on covers that decision.

What should you check after you sign?

Due diligence continues after signing. Connect the product read only first and add write access one tool and one action at a time. Record it in your AI register with an owner. Review it quarterly: what it can access, whether usage justifies the cost, and whether the vendor's terms have changed.

Vendors change their terms, models and features often, and a product that passed due diligence last year may behave differently now. Set a calendar reminder to reread the terms for the plan you use.

Also watch for permissions that grow. An integration approved to read the CRM may later be given rights to update it, one convenient click at a time. Each new permission deserves the same question as the first: is this needed, and who approved it?

Skein's strategy work is a fixed-scope diagnostic that ends in a ranked shortlist of three to five opportunities, and if a tool the client already pays for can do the job, it is switched on rather than built; the AI strategy page shows the process. That suits a business weighing several AI products without a clear way to compare them. A business buying one low-risk tool for internal notes can run this checklist itself in an afternoon.

Frequently asked questions

What is AI vendor due diligence?

AI vendor due diligence is checking an AI product before adopting it. It covers whether the product suits the intended use, how it handles and stores data, whether it trains on that data, how it is secured, what testing supports its claims, where people can oversee it and how the business can leave.

What questions should I ask an AI vendor?

Ask what the product has been tested for, where your data goes, who can access it, whether it trains on your data and whether that can be turned off, what security and admin controls exist, what evidence supports its claims, and how you can export your data and leave.

Does Australian law require AI vendor due diligence?

The privacy regulator expects organisations to conduct due diligence before adopting AI products that will handle personal information, and says it should be ongoing rather than set and forget. The Privacy Act applies to AI use involving personal information. Take advice on how this applies to your business.

Are business AI plans safer than consumer plans?

Often, because business plans usually offer different data terms, such as no training on your content by default, plus admin controls and single sign-on. However, terms vary by vendor and change over time. Always check the specific plan's terms rather than assuming every business plan is the same.

How often should AI vendors be reviewed?

Review each AI vendor at least quarterly. Check what the product can access, whether usage justifies the cost and whether the vendor's terms, models or features have changed. Review sooner if the vendor announces a significant change or you plan to give it access to new systems or data.

Related reading