To run a small AI pilot safely, pick one repetitive task, write down how you'll judge success before you start, limit the pilot to the data that task needs, use a business-grade AI service whose terms say it won't train on your data, keep a person approving every output, run it alongside the current process, and set a fixed end date with a decision at the end: extend, change or stop. Write all of that on one page before anything is switched on. That page is what your IT, security and finance people need to say yes, and it's what stops a small experiment turning into an open-ended project.
Why "do something with AI" is a hard brief
A brief like "do something with AI" has no edges: no task, no measure of success, no data rules and no budget. So IT worries about data leaving the business, finance worries about a subscription nobody can switch off, and the team worries about being blamed for its mistakes. A pilot answers those worries by being small and specific: one task, one team, a few weeks, a person checking every result, and a written decision at the end.
Step 1: Choose one task
The best first task is boring: daily work where most of the effort is reading something and typing it somewhere else. Good candidates tick most of these:
- High volume and repetitive: supplier invoices, delivery dockets, timesheets, purchase orders, customer order emails.
- Messy input: PDFs, scanned paperwork, emails or free text, which is why it's still done by hand.
- A clear right answer: a person can check the result quickly against the original document.
- Low harm if it's wrong: a mistake is caught at the check step and costs a few minutes, not a payment to the wrong bank account.
Often the obvious candidate is information typed in again: the order arrives as a PDF, someone keys it into the sales system, then someone else keys the same details into the job or stock system. That's double data entry (the same information typed into two systems), and our manual data entry page covers the wider problem. AI that reads the document and prepares the entry for a person to approve is a sensible first pilot.
Avoid anything customer-facing, anything that decides something about a person (hiring, credit, pay), and anything nobody can easily check. And if the real issue is two systems that don't talk to each other, a plain connection between them may remove the typing without any AI at all.
Step 2: Define success before you start
Measure the work the old way first, so you have a baseline. Keep to a few measures your team can record:
- Time per item: how long a person takes to process one invoice or order today, timed on a normal week.
- Error rate, checked by a person: of the items the AI prepares, how many did the reviewer have to correct, and what kind of corrections were they?
- Items it couldn't handle: documents it skipped or flagged as unsure.
- Reviewer effort: whether checking the AI's work is actually quicker than doing the work.
Write down what result would count as worth continuing. Your business sets those numbers from its own measurements, not a vendor's brochure.
Step 3: Decide what data the pilot will touch, and how it's protected
This is the section your IT and security people will read first. Be specific.
List the data. For example: "invoices from three named suppliers: supplier names, ABNs, amounts and line items, no customer or staff personal information." If a document might contain personal information (names, addresses, bank or health details), say so and decide whether it's in or out.
Use a business-grade service, not a free consumer tool. Major providers' business and enterprise plans state that your inputs and outputs aren't used to train their models by default. OpenAI states this for its business plans and API on its business data page, and Microsoft states that Microsoft 365 Copilot prompts and responses aren't used to train foundation models in its enterprise data protection documentation. Free and personal plans can differ and terms change, so check the current terms for your exact plan and keep a copy.
Check the privacy rules in plain terms. The Australian Privacy Principles in the Privacy Act set out how organisations collect, use, store and disclose personal information. Smaller businesses under a turnover threshold are generally exempt, though some are covered regardless; the OAIC's small business guidance explains who. The OAIC's guidance on commercially available AI products says, in short: use personal information only for the purpose it was collected for, keep it secure, understand whether the provider can access what you put in, have a person verify AI output, be open about your AI use, and as best practice don't enter personal information into publicly available AI tools.
The easiest pilot to approve is one that needs little or no personal information. We're not lawyers, so if yours does involve it, get advice. Also ask where the service stores data and for how long, and turn off history or sharing features the pilot doesn't need.
Step 4: Keep a person approving every output
The AI does the first pass; a person approves. Nothing it prepares reaches your accounts, job system, customers or suppliers unchecked. The federal government's Guidance for AI Adoption makes the same point: someone accountable, and people kept in control.
Make the check practical: the reviewer sees the original document next to what the AI extracted, with anything it was unsure about highlighted. If checking takes as long as doing the work, that's a result too, learned cheaply.
Step 5: Run it alongside the current process
The old process keeps running, and it's still what goes into your systems. The AI's output is compared against it, so a bad week costs only reviewer time and you get a fair comparison for step 2. Once results are steady, reviewed AI output might replace manual entry for part of the work, such as one supplier.
Step 6: Set a fixed timeframe and a fixed cost
A few weeks is usually enough for a repetitive task with real volume. Put the end date and a review meeting in the calendar on day one. Cap the spend with a cancellable monthly plan or usage limits on a pay-as-you-go service, write the maximum cost on the pilot page so finance knows the worst case, and avoid annual commitments until the pilot has earned one.
Step 7: Document it for IT, security and finance
Keep it to one or two pages covering:
- The task, the team and who has access.
- The exact product and plan, with links to its data-use and privacy terms.
- What data goes in, what's excluded, and any personal information involved.
- Where data is stored and for how long.
- How the approval step works, and how AI-prepared items are marked.
- Start date, end date, maximum cost and who can switch it off.
- How you'll remove data and close accounts if you stop.
Our guide to briefing an integration project uses the same one-page approach for systems work, and the format carries over well.
Step 8: Decide at the end: extend, change or stop
At the review, compare the measures with the baseline and decide:
- Extend: it worked. Widen it carefully and decide whether it needs connecting to your systems, so the approved result flows straight in instead of being copied across.
- Change: promising but not there yet. Adjust the task or check step and run another fixed period.
- Stop: it didn't help. Cancel, remove the data, note what you learned and move on. That's a far better outcome than a stalled project.
Red flags to watch for
- A vendor who wants all your data up front. "Connect everything and we'll find the value" isn't a pilot.
- No human check. Any setup where AI output goes straight into your accounts, payroll or customer communications.
- Open-ended costs. Per-use pricing with no cap, or an annual contract before you've seen results.
- Vague data terms. If the vendor can't tell you in writing whether your data trains their model, where it's stored and how to delete it, walk away.
- Promised savings before they've seen your work. Your pilot measures that.
- A pilot with no end date. If nobody has to decide anything, it will drift.
When it's worth getting help
Plenty of pilots need no outside help: a business-grade AI tool, one task, a spreadsheet of results and a patient reviewer. Help is worth it when the task touches several systems, when approved output needs to flow into your job, stock or finance system, or when you're unsure where to start. Our AI consultant for business and AI workflow automation pages explain how we approach practical AI work, always with a person approving.
If the bigger question is where your processes and systems are losing time in the first place, our $5,000 Process & Systems Audit maps how work moves through your business and gives you a written, prioritised list of what to fix, which often shows the best first pilot. The business process improvement guide is a good place to start if you'd rather work through it yourself, and practical AI uses for Australian SMEs has more task ideas.
Frequently asked questions
How long should an AI pilot run?
For a daily task, usually a few weeks. Set the end date before you start and hold the review on that date, even if results are unclear.
Is it safe to put company data into AI tools?
It depends on the plan. Business and enterprise plans of the major services state that your data isn't used for training by default; free plans may differ. Check the current terms, keep personal information out where you can, and follow the OAIC's guidance.
Do the Australian Privacy Principles apply to my AI pilot?
If your business is covered by the Privacy Act and the pilot handles personal information, yes. Many smaller businesses are exempt, but some are covered regardless, so check the OAIC's guidance. The easiest path is a pilot task that doesn't need personal information.
What should an AI pilot cost?
Keep it to a cancellable subscription or a capped pay-as-you-go service, plus your team's time to review results. If connecting it to your systems is needed later, that's a separate piece of work, which we scope and give a fixed quote for after a free initial chat.
To talk through which task to pilot first, or how to connect the result to your systems once it works, book a free initial chat.