10 Questions to Ask Before You Buy Any AI Tool

The procurement playbooks circulating in 2026 read like compliance audits. Forty-seven questions about model card lineage, sub-processor chains, ISO 42001 alignment. Useful if you're an ASX 200 procurement officer. Useless if you're a 30-person business trying to decide whether to buy a sales AI tool by Friday.

Australian SMEs need a shorter list. Ten questions, asked in one call, that tell you whether the vendor knows what they're doing and whether the tool will still make sense in eighteen months.

These are the ten I'd ask. Each comes with a note on why it matters and what a good answer sounds like.

1. Where does our data physically sit, and which jurisdictions can reach it?

Data residency is a contract question, not a marketing one. Under Australian Privacy Principle 8, your organisation stays accountable for personal information once it leaves Australia, even when a third-party AI vendor is processing it. The Privacy and Other Legislation Amendment Act 2024 also adds new transparency obligations around automated decisions that come into effect 10 December 2026.

Good answer: a specific region (ap-southeast-2, eu-west-2), a named cloud provider, and a clear statement on Australian Privacy Act compliance. Bad answer: "We use AWS." That tells you nothing.

2. What model are you actually using under the hood, and do you swap it?

This is the buyer trap that surfaced hardest in 2026. Many AI wrappers route prompts through whichever LLM is cheapest that day — GPT-4, Claude, an open-source model, sometimes mid-conversation. Smart routing can cut vendor costs by 30 to 85 percent according to MindStudio's 2026 router comparison, but you're paying for output quality, not their margins.

Good answer: a named model, a versioning policy, and notice before any swap. Bad answer: vague language about "best-in-class models" or "our proprietary AI."

3. What was the model trained on, and is our data part of training?

Two separate questions. Training data of the underlying model affects accuracy and bias. Whether your data joins the training set affects everything else.

Good answer: a clear "no, your prompts and outputs are not used for training" written into the contract — not just the FAQ page. Bad answer: anything ambiguous or buried in a terms of service link.

4. Who else has access to our data, and through what sub-processors?

Most AI vendors are themselves built on top of other vendors. Your data may pass through three or four organisations before it returns to you. Each one is a potential breach point.

Good answer: a sub-processor list, willingness to notify you when it changes, and SOC 2 Type 2 reports for the primary vendor. SOC 2 Type 2 is now the procurement floor — vendor risk teams rarely open a commercial conversation without one on the table.

5. If we leave in two years, how do we get our data and our work out?

Export rights are easy to promise and hard to deliver. Ask for a specific format (JSON, CSV, open standard — not proprietary), a defined timeline (48 hours, not "reasonable period"), and confirm it survives contract termination including termination by the vendor.

Good answer: a sample export they'll show you during evaluation. Bad answer: "We support data export." Make them prove it.

6. How does this connect to the tools we already run?

A standalone AI tool that doesn't talk to your CRM, your inbox, or your project management software creates more work than it saves. You'll end up copying data between systems and resenting the purchase by month three.

Good answer: native integrations with the tools you actually use, plus a documented API. Bad answer: "We have an open API" with no specifics, or a Zapier integration as the only option.

7. What happens when the AI gets it wrong, and who catches it?

AI fails. The question is whether the system is designed to fail loudly or quietly. A confident-sounding wrong answer that ships to a customer is worse than no answer at all.

Good answer: built-in confidence scoring, human-in-the-loop review for high-stakes outputs, and clear escalation paths. This is also why we build AI that augments your team rather than replaces them — the human stays in the loop on purpose.

8. How does pricing actually work at scale?

Most AI tools price per seat, per call, per token, or some combination. The number you see on the pricing page is rarely what you pay in month six.

Good answer: a clear pricing model, an estimate based on your actual expected usage, and a written commitment on how price increases work. Bad answer: "Contact sales for enterprise pricing" with no benchmarks.

9. What does support look like after we've signed?

The sales conversation is the best service you will ever get from this vendor. Ask what comes after. Implementation support, response times on bugs, who answers when something breaks at 9pm on a Tuesday.

Good answer: a named point of contact, defined response SLAs by severity, and a real human on a real channel. Bonus marks if support hours actually cover Australian business hours rather than US Pacific time only. Bad answer: a help centre URL and a chatbot.

10. If you were in my seat, would you buy this today?

The gut-check question. Ask it at the end of the call, after you've worked through the other nine. You're not looking for a sales pitch — you're looking at how they answer.

Good answer: a thoughtful "yes, because" or "not yet, here's what would have to change first." Bad answer: enthusiastic agreement with no nuance. Anyone who can't think of a reason not to buy their own product is either lying or doesn't understand it.

---

Ten questions, thirty minutes, one call. If a vendor can't answer most of them clearly, the tool isn't ready for you yet — or the vendor isn't ready to sell to a business that takes the purchase seriously.

The point of this checklist isn't to catch out vendors. It's to find the ones who've thought hard about the same questions you're now asking. Those are the ones worth signing with.

At Kursol we build AI that augments your team, not replaces them. The procurement question we care most about is question seven — what happens when the AI is wrong, and is there a human there to catch it.

FAQ

Do I need SOC 2 Type 2 from every AI vendor?

For anything that touches personal information regulated under the Privacy Act, yes — it's the procurement floor in 2026. For a tool that only processes public data or your own internal notes, you can be more flexible, but ask anyway. It tells you whether the vendor takes security seriously.

What if the vendor uses OpenAI or Anthropic under the hood?

That's fine — most AI tools do. What matters is that they tell you which model, that they have a contract in place that prevents your data being used for training, and that they notify you before swapping to a different model mid-stream.

How long should AI vendor evaluation actually take?

For most SME purchases, one 30-minute call with these ten questions plus a two-week pilot is enough. If a vendor asks for a 90-day enterprise procurement process, they're either selling to the wrong customer or padding their sales cycle.