The most reliable way for a small business to use AI is on narrow, repetitive tasks with a person checking the output: summarising long messages or documents, sorting incoming enquiries into categories, drafting replies that staff approve before sending, and answering staff questions from your own documents. These tasks have clear inputs, a clear definition of “good”, and a cheap way to catch mistakes. A general-purpose chatbot facing customers has none of those, which is why so many end up switched off.
Why is the chatbot a trap?
A customer-facing chatbot has to handle any question, in any wording, with no human in the loop, and it speaks for your business every time. That is the hardest possible setting for a language model. It can state things confidently that are wrong, invent policies or prices, and frustrate people who just wanted a phone number.
There are good chatbots, but they are usually tightly limited, well tested and backed by a clear hand-off to a person. For most small businesses, the same budget does more elsewhere. Start where mistakes are cheap and visible to your staff, not your customers.
What AI jobs actually work for a small business?
Summarising. Long email threads, call transcripts, meeting notes, tender documents, customer feedback. A model produces a short summary with the key facts and actions, and a person reads it before acting. The time saved is real, and an error is easy to spot because the original is right there.
Classifying enquiries. A new enquiry arrives through your website form. A model reads it and tags it: quote request, existing customer, complaint, job application, spam. It can also pull out structured details, such as suburb, service type or urgency, and add them to your CRM. Routing then happens automatically, and staff can correct the odd wrong tag. If your forms are not reliably reaching your CRM yet, fix that first: see connecting your website forms to your CRM properly.
Drafting replies with human review. For common enquiries, a model drafts a response using your templates, prices and tone, and a staff member edits and sends it. The draft never goes out on its own. This works well for quote follow-ups, booking confirmations with specific details, and answers to the same ten questions you get every week.
Searching your own documents. Often called retrieval, or RAG. Your procedures, product specs, past proposals or policies are indexed, and staff ask questions in plain language. The system finds the relevant passages and the model answers using only those, with links back to the source. This is far more dependable than asking a general model, because it is grounded in your material and staff can check the source.
Tidying data. Standardising addresses, splitting full names, extracting details from emailed PDFs into a spreadsheet or app. Dull work that a model does quickly, with spot checks.
Where is AI a bad fit?
- Anything that must be exactly right every time without review: prices on invoices, tax calculations, legal or medical advice, compliance decisions.
- Decisions about people such as hiring, credit or eligibility, where errors and bias have real consequences and may carry legal obligations.
- Tasks a simple rule already handles. If “send enquiries with ‘invoice’ in the subject to accounts” works, you do not need a model.
- Very low volume. If you get five enquiries a week, the setup and running cost may outweigh the time saved.
- Problems that are really process problems. AI will not fix a booking system nobody updates or a form that sends to the wrong inbox.
What about privacy when you send customer data to an AI API?
Every one of the uses above involves sending text to an AI provider, and that text often contains personal information: names, phone numbers, addresses, health or financial details. Under the Privacy Act 1988 and the Australian Privacy Principles, sending personal information to a provider overseas can be a cross-border disclosure, and you remain responsible for how it is handled. Many AI providers process data outside Australia.
This is not legal advice, and the details depend on your business and the data involved. Practical points to raise with your adviser and check in each provider’s terms:
- Use the business API, not a consumer chat app. Business and API terms from major providers generally say customer data is not used to train models by default, but check the current terms for your plan.
- Find out where data is processed and how long it is kept. Some providers offer regional processing or shorter retention on certain plans.
- Send the minimum. Strip details the task does not need before the text leaves your system.
- Update your privacy policy so it accurately describes the service providers you use and where data may go.
- Be careful with sensitive information. Health information in particular carries stricter rules. Clinics should read the OAIC’s guidance before sending patient details anywhere.
Our article on the Privacy Act and your website forms covers the wider basics, and the OAIC publishes guidance on using commercially available AI products.
How should you start?
- Pick one task that is repetitive, text-heavy and currently done by a person. Write down what a good output looks like.
- Gather twenty real examples with the answer a staff member would give. These become your test set.
- Build the smallest version, usually a script or workflow connected to your existing tools through their APIs, not a new platform.
- Keep a person in the loop. Outputs are suggestions until you have seen enough of them to trust a narrower slice.
- Log inputs, outputs and corrections, so you can see accuracy over time and improve the prompts.
- Set a budget alert on the AI provider account. Usage-based billing can surprise you.
Be sceptical of anyone promising AI will run a department. Narrow, checked automation is less exciting, and it is the kind that is still running a year later.
Much of this work is really integration: getting data from your form, inbox or CRM to the model and back again. Our REST and GraphQL APIs page and HubSpot page explain how that side usually fits together. If you are weighing whether AI tools could help you be found online, that is a different question, covered in getting your business recommended by AI answers.
If you only do three things
Choose one narrow, repetitive task instead of a chatbot, keep a person reviewing the output, and check what customer data you would be sending and where it goes before you send any. That combination gets real value from AI with very little risk.
If you have a task in mind and want it built properly, our AI automation and integration service handles small, defined jobs like these, connected to the tools you already use. From $900. Most AI automation jobs land between $900 and $4,000.
