Why governance comes early, not last
In most businesses, AI use starts before anyone decides anything. Someone uses a free chatbot to tidy a customer email; someone else pastes a spreadsheet in to get a formula. That is not a disaster, but it means your data rules are being set by default, one person at a time. Governance just means deciding on purpose.
The Australian Government’s National AI Centre publishes Guidance for AI Adoption, which sets out six essential practices and starts with deciding who is accountable. The Office of the Australian Information Commissioner (OAIC) has also published guidance on privacy and the use of commercially available AI products. Neither is written only for large organisations, and both are worth a read. What follows is how we turn that kind of guidance into something a small business will actually keep up.
The five documents worth having
1. An AI use policy (two pages, not twenty)
It should answer, in plain English: which AI tools are approved, and on which accounts (business accounts, not personal ones); what information must never go into AI tools that aren’t approved for it, such as personal information about customers, patients, students or staff, and commercially sensitive material; who checks AI output before it reaches a customer; and how to report a mistake or a worry. Date it, and put a review date on it.
2. An AI register
A spreadsheet is fine. One row per AI tool or automation: what it does, which systems and data it can read or change, who owns it, what the human approval point is, when it was last reviewed. This is the document you will be glad of when a customer, auditor or insurer asks how you use AI.
3. An approval map
For each automation, list the actions and mark which need a person every time, which need a person above a threshold, and which can run on their own. For example: drafting a reply runs on its own; sending it needs a person; sending a quote over a set value needs a manager.
4. A short incident note template
What happened, what the AI produced, who caught it, what was changed. Most incidents will be minor. Writing them down is how you decide when an approval step can loosen and when it can’t.
5. A vendor checklist
Before adopting an AI product: where is data stored and processed, is your data used to train the vendor’s models, can you export and delete your data, what happens when the contract ends, and who at the vendor can see your information.
Privacy obligations to keep in view
If your business is covered by the Privacy Act 1988, the Australian Privacy Principles apply when you put personal information into AI tools, the same as any other system. Health service providers are generally covered regardless of turnover, and in Victoria the Health Records Act 2001 also applies to health information. Training providers handle student information under their own obligations. We are not lawyers and this is not legal advice. For anything you are unsure about, get advice from your lawyer or check the OAIC’s guidance. In practice, the AI-specific questions are usually: is personal information going into a tool that stores or reuses it, would the person reasonably expect that use, and can you meet a request to access or correct what the AI produced about them?
Adoption: the four-week rhythm
Training once and hoping does not work. We use a simple weekly rhythm during the pilot and rollout months of the 90-day plan:
| Week | Focus | What happens |
|---|---|---|
| 1 | Show the real job | A one-hour session on the actual workflow, with real (appropriately handled) examples. What it does well, where it gets things wrong, what the person is still responsible for. |
| 2 | Use it with a buddy | Each user pairs with someone from the pilot group. Questions go into one shared channel, not private messages. |
| 3 | Collect the misses | A 20-minute session reviewing corrections and bad outputs. Fix the prompts, rules or integration where possible, and say what changed. |
| 4 | Decide together | The team says what to keep, what to change and what to drop. Usage numbers and before-and-after timings are shared openly. |
The worries staff actually have
“Is this replacing me?” Answer it directly. If the aim is to absorb growth without hiring, or to get rid of the Thursday-night backlog, say that. If roles will change, say how. Vague reassurance is worse than an honest answer.
“What if it’s wrong and I approved it?” Make the approval step genuinely reviewable: show the source next to the draft, highlight what the AI changed, and keep the volume manageable. A person cannot properly check 300 items an hour.
“I tried it and it was rubbish.” Often true for a generic tool on a specific job. Ask what they tried. Their example is usually a good test case for the pilot.
“I’m already using something else.” Find out what. If it is a personal account of a public tool, move that use onto an approved business account rather than banning it outright. Bans tend to push the use out of sight.
Unapproved AI use
The National AI Centre also covers managing unapproved AI use, and it is a real issue in small businesses: it is free, it is in the browser, and it is quicker than asking. The practical response is to offer an approved alternative that is at least as convenient, explain why the rule exists (usually customer or patient data), and check in periodically without making it a disciplinary matter.
Who should own this?
In a business of under 200 people, one accountable person is enough: often the operations manager, practice manager or a director. Their job is to keep the policy and register current, run the monthly review of incidents and approvals, and sign off new automations. It should take a few hours a month once things are running. If nobody has those hours, that is a planning problem to solve before rolling anything out.
Book a 15-minute call to talk through governance for your business, or see how engagements run.