AI integration for small business: where it pays off and what it costs

AI integration means connecting an AI model to the software a business runs on, so that work a person does by reading and retyping happens faster, with a person still checking what matters. For a small business it pays off in a few specific places: reading documents such as receipts, invoices and forms into structured data; sorting and routing enquiries; drafting replies and summaries for someone to edit; and answering questions from the business's own content. With us, AI features start from $3,000 ex GST, priced on top of a custom software or web app package and fixed in a written scope, plus the AI provider's charge for each request the system makes. This guide covers where AI earns its place, where it does not, what it costs, how to handle customer data, and two products our co-founder Anis built that show the approach.

The pattern is the same every time: a lot of similar information arrives, a person reads it and types the important parts somewhere else. That is the work AI does well.

Reading documents. Receipts, supplier invoices, application forms, delivery dockets and certificates can be read into structured fields: the supplier, the date, the amount, the GST, the reference number. Each result carries a confidence score, so a person looks closely at the uncertain ones and confirms the rest quickly.

Sorting enquiries. Incoming emails and form submissions can be classified by type and urgency, routed to the right person and used to start a record, so nobody spends the first hour of the day triaging an inbox. Our own contact form works this way for spam: its risk score includes an AI classification, and a submission the checks flag is held for a founder to review, never deleted because of its score.

Drafting. A first draft of a reply, a job summary, a quote description or meeting notes, for a person to edit and send. The time saved is in starting from something rather than a blank page.

Answering from your own content. A chat assistant can answer questions from your policies, product information or help articles. It needs limits: the assistant on this site runs behind per-visitor limits and daily caps on AI use, and the prices it gives come from our published ranges, never from the model. Our security page describes those limits.

A quick test for any idea: does it happen often, is each case much like the last, can a person check the result in seconds, and is a mistake cheap to catch? If the answer to all four is yes, AI is worth a look.

• Anything that must be exact. Totals, tax, prices, safety limits and dates are for code, not for a model. A model can choose; deterministic code should calculate.
• Decisions about people without review. Approving a refund, rejecting an application or changing someone's account needs a person, at least until the system has earned trust.
• Rare tasks. If something happens twice a month, a good form or checklist is cheaper than an AI feature.
• Information that does not exist yet. An assistant cannot answer from documents nobody has written. Write them first.
• Problems a rule solves. "Enquiries from these postcodes go to this person" needs no model at all.

Saying no to AI where it does not fit is part of the job. The discovery workshop that starts every project over $10,000 is where we check whether an AI feature is worth its cost, and sometimes the honest answer is a better form.

With us, an AI feature starts from $3,000 ex GST, priced on top of the package it sits in: a web app or custom software system at $15,000 to $40,000, or a mobile app at $20,000 to $50,000. Like the rest of the project, it is fixed in the written scope. What moves an AI feature up from $3,000 is the amount of work around the model: how many kinds of document it reads, how carefully results must be checked and the screens that checking needs, how many systems the result goes to, and, for an assistant, how much content it answers from.

There is also a running cost. The AI provider charges for each request the system makes, and the amount depends on the model and the volume. A well-built feature controls that cost from the start:

• A cap before the spend. A billing or usage check runs before each request, so a runaway script or an abusive user is stopped before it runs up a bill.
• The smallest model that does the job. A larger model is used only where it makes a real difference.
• A refund on failure. If a request fails or times out, the user is not charged for it.

Hosting, monitoring, updates and fixes after launch are covered by our care plan at $300, $700 or $1,500 a month ex GST. Every price is on our pricing page.

Carefully, and in writing. The Office of the Australian Information Commissioner's guidance on privacy and the use of commercially available AI products, published on 21 October 2024 and updated on 17 January 2025, says privacy obligations apply to any personal information put into an AI system, and to output that contains personal information. As a matter of best practice, it recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools.

In practice, that leads to six rules worth building by. Our own site follows them: its privacy policy names each AI provider it uses and what each one receives.

1. Send only what the task needs. A receipt reader needs the receipt, not the customer's whole history.
2. Use the provider's business service, not a consumer chat app, and read its terms on what it keeps and for how long.
3. Name the provider in your privacy policy and say what it receives.
4. Keep a person in the loop for anything that affects a customer.
5. Treat what users type as data, never as instructions, so nobody can talk the system into doing something it should not.
6. Cap usage and keep a record of what the AI was asked and what it returned.

This is how we build, not legal advice. For your own obligations, the OAIC guidance and its checklists are the place to start, and a lawyer is the right person for anything specific to your business.

Two products our co-founder Anis built and runs himself, before ELORA DIGITAL opened, show the approach. Both belong to him, not to ELORA DIGITAL, and neither is client work.

Reckonly: AI reads, a person confirms. A sole trader photographs a receipt, and AI OCR built on Claude extracts the vendor, ABN, total, GST and date, with a confidence score so the uncertain receipts stand out. The user checks and confirms each one, and only then does it go to Xero, MYOB AccountRight or Zoho Books, as a draft. A billing check runs before every AI call, so the spend is capped before it happens.

Scentena: the AI chooses, the engine decides. Scentena turns a plain-English brief into a perfume formula. The model may only choose from a fixed list of real materials in the database; a deterministic engine then applies the IFRA safety limits and makes the formula add up to 100 percent, so the model never sets a percentage. An AI request reserves the user's credit and refunds it if the call fails, with a 60-second timeout.

No user, revenue or accuracy figures are published for either product, and none are quoted here. They show the shape we bring to client work: the model proposes from real options, code checks the result against the rules, and a person sees it before it matters. Our work has the full case studies.

Write down the task: what comes in, what a person does with it today, how often it happens and what a mistake costs. Then tell us about it. We are a small studio in Box Hill North, Melbourne, working with businesses across Australia, and a founder replies with an estimate range within one business day. If a form or a simple rule would do the job better than AI, we will say so.