Generative AI fundamentals

Why it matters

Writing, summarising and reformatting information takes up a surprising share of a small business owner’s week: quotes, follow-up emails, policy documents, job ads, social posts, meeting notes. Generative AI does the first draft of all of these in seconds. Used well, that is hours back every week and a higher floor on quality. Used badly, it produces confident errors in your name, leaks customer information, or quietly adds a monthly subscription nobody is using. The difference is understanding, not technical skill.

What a large language model actually does

A large language model (LLM) is a very large statistical model trained on enormous amounts of text. During training it learns to predict the next word, given everything before it. Do that across most of the written internet and the result is a system that has absorbed grammar, facts, styles, reasoning patterns and a great deal of nonsense, all mixed together. When you type a prompt, the model continues it with the most plausible response, one token at a time.

This explains both the strengths and the failures. The model is extraordinary at producing text that sounds right, in any style, on any topic. It has no built-in notion of truth; "plausible" and "true" usually coincide, but not always. It does not look things up unless the tool around it is built to do so, it does not know anything about your business unless you tell it, and it has a training cut-off date after which it knows nothing. Image, audio and video generators work on the same principle with different kinds of data.

Hallucination, and why it is not a bug you can wait out

Hallucination is the industry term for a model stating something false with complete confidence: an invented case citation, a made-up statistic, a plausible-sounding Fair Work rule that does not exist. It happens because the model is generating likely text, not retrieving facts. Newer models hallucinate less, and tools that give the model your documents to work from (often called retrieval-augmented generation) reduce it further, but no current system eliminates it.

The practical response is design, not hope. Give the model the source material and ask it to work only from that. Ask it to quote or cite where it found each claim. Use it for tasks where a wrong answer is easy to spot or cheap to fix. And never let generated text reach a customer, a regulator or a court without a person who knows the subject reading it first.

Prompting: giving the model what a good brief would give a person

A prompt is just the instruction and context you give the model. The tools are forgiving, but the quality of output tracks the quality of the brief, exactly as it would with a contractor. State the role and audience, the task, the format you want, any constraints, and, crucially, the source material. "Write a reply to this customer" produces something generic; "You are the owner of a Perth plumbing business. Reply to the email below in a warm, brief tone, confirm Thursday 8 am, and mention our 12-month workmanship guarantee" produces something usable.

Iterate rather than perfect. Ask for three options, then ask for changes. Tell the model what was wrong with its last attempt. Once a prompt works for a repeated task, save it as a template so staff get consistent results. Many tools let you store these as custom instructions or reusable assistants.

Privacy, cost and the fine print

Everything you type into a generative tool goes somewhere. On many free consumer plans, your input may be stored, reviewed by the provider and used to train future models. Business and enterprise plans usually offer the opposite: no training on your data, retention controls and a contract. If staff paste customer names, medical details or financial records into a consumer chat tool, you have disclosed personal information to a third party, and the Australian Privacy Principles about use, disclosure and overseas transfer are in play. Decide which tools are approved, on which plan, before this happens by accident.

Cost is usually a subscription per user per month, or a metered charge per request for tools built on top of a model provider. Neither is large on its own; the risk is sprawl, where five tools do similar things and nobody owns the bill. Start with one general assistant on a business plan and the AI features already inside tools you pay for, then add specialist tools only when a specific job justifies them.

How to introduce generative AI safely

1. Choose one approved assistant on a business plan Confirm it does not train on your data and where data is stored.
2. Write a one-page usage rule What may be pasted in, what may not, and what always needs a human check.
3. Pick three routine writing tasks Quotes, follow-ups, job ads or meeting summaries are typical starting points.
4. Build a prompt template for each Role, task, source, format, constraints. Save and share them.
5. Review every output for a month Note what the model gets wrong and refine the templates.
6. Measure time saved and errors caught Compare with your pre-AI baseline before expanding.
7. Train the team A one-hour session on prompting and the limits prevents most problems.

Real-world example

A Wollongong recruitment agency drafts faster without losing its voice

Escarpment Recruitment places trades and warehouse staff for employers in the Illawarra. Its three consultants spend hours every week writing job ads, candidate summaries and client updates. The director, Luke, trials a general assistant on a business plan that does not train on their data. He writes three prompt templates: one for job ads from a client brief, one for candidate summaries from interview notes, and one for weekly client updates.

The job-ad template includes the agency’s tone guide and a reminder to avoid any wording that could breach anti-discrimination law. Early on, the assistant invents a required licence that the client never mentioned; the consultant catches it in review and the template gains a line telling the model to flag anything not in the brief. Candidate names and personal details stay in the agency’s CRM; the summary prompt uses initials only. After two months the team estimates it has saved several hours a week per consultant, and the ads are more consistent than before.

Common mistakes

• Trusting a fact because it reads well Check every number, name, date and rule against a source. The model is confident regardless.
• Using a consumer plan for business data Free tiers often train on your input and offer no contract. Pay for a business plan with data controls.
• One-line prompts, then blaming the tool A thin brief gets a generic answer. Give role, task, source, format and constraints.
• Letting generated text go out unread The person who sends it owns it. Build the review step into the workflow.
• Subscription sprawl Several overlapping tools, each billed monthly. Approve one general assistant and add specialists only for proven needs.

KPIs to track

• Draft time per document Minutes from brief to a ready-to-send draft, including review. Compare with your pre-AI baseline.
• Review corrections per ten outputs How often the reviewer changes something material. Falling means prompts and context are improving.
• Privacy rule breaches Instances of personal information entering an unapproved tool. The target is zero, and staff should feel safe reporting one.
• Cost per user per month Total generative AI spend divided by active users. Watch for licences nobody uses.

Frequently asked questions

Which generative AI tool should I use?

Start with the AI features already inside tools you pay for, plus one general assistant from a major provider on a business plan. The leading assistants are more alike than different for routine business writing; data terms and integration with your existing tools matter more than the brand.

Can I stop the model making things up?

Not entirely. You can reduce it sharply by giving the model the source material, asking it to work only from that and to cite where it found things, and using tools that retrieve from your documents. A human check remains necessary for anything that matters.

Who owns content the AI writes for me?

Terms vary by provider, and Australian copyright law on AI-generated works is still developing. Most business plans assign output to you, but material that merely reproduces someone else’s work can still infringe. Edit substantially and get advice before relying on generated content commercially.

Is it legal to put customer details into an AI tool?

It can be, if the tool’s terms and your privacy policy cover that use and disclosure, and the data is protected appropriately. Many businesses simply avoid it by using initials or anonymised data. The OAIC has published guidance on privacy and generative AI worth reading.

Key takeaways

• Large language models predict plausible text; plausible usually means true, but not always.
• Hallucination is a design constraint, not a passing bug: supply sources, ask for citations and keep a human check.
• Good prompts look like good briefs: role, task, source material, format and constraints.
• Use business plans that do not train on your data, and write down what staff may paste in.
• Start with routine writing tasks, measure the time saved, and expand only on evidence.