Treasury released the 2026 Intergenerational Report on 21 September, and one line in its AI chapter describes most Australian businesses exactly: two-thirds have adopted AI in some form, but fewer than 10 per cent describe that adoption as significant. In plain terms, staff are using chatbots, and the business itself has barely changed. This post sets out a 90-day plan for a mid-sized business to move from the first group into the second, based on what the report and the surveys behind it say is actually getting in the way.

This is the work we do at Digital One Agency. Our digital transformation and AI integration service connects AI to the systems a business already runs, and the first step is a free AI audit that tells you where the 90 days should go.

What the Intergenerational Report actually says

The 2026 Intergenerational Report projects the economy out to 2066 and names AI as one of five major transitions. Treasurer Jim Chalmers called advances in AI “the most dramatic change since the last IGR in 2023” in his media release.

The interesting part for a business owner is not the 40-year forecast. It is the diagnosis of right now. The report says early experimentation is well under way, but “deep integration of AI into business processes is in its early stages”. It notes similar results among small and medium enterprises, and it says low trust in AI among workers, consumers and businesses helps explain why measured productivity gains have been very modest so far, despite rapid growth in AI investment.

The report also says adoption is fastest in ICT, financial and professional services, and slower in agriculture, construction, trade and distribution. Transport and manufacturing are expected to take longer “as legacy systems, cost pressures and the complexity of use cases are navigated”.

Run the numbers on those two headline figures and the gap gets sharper. If two-thirds of businesses use AI and fewer than one in ten use it significantly, then of the businesses that have started, roughly 85 per cent have not got past the shallow end. That is the market position most of your competitors are in. It is also the opening.

Why most AI stalls at the chatbot stage

The adoption figure comes from a Reserve Bank survey of 105 medium and large firms in its liaison program, run between June and August 2025 and published as Technology Investment and AI: What Are Firms Telling Us?. Its breakdown is more useful than the headline. Nearly 40 per cent of firms reported minimal use, mostly off-the-shelf tools like ChatGPT for tasks such as summarising emails and research. Around 30 per cent had made more substantive progress, using AI for things like demand forecasting, inventory management and fraud detection.

The firms named their own obstacles:

• finding skilled people such as data engineers
• a lack of digital readiness and uncertain return on investment
• integrating with legacy systems
• the cost of the technology
• regulatory uncertainty

Read that list again and notice what is missing. Nobody said the AI models were not good enough. Every barrier on it sits between the model and the business: the data, the old systems, the people who can connect them, and the confidence to spend money on it. That is an integration problem, and integration problems respond to a plan.

The same RBA work found firms’ IT budgets were being pulled towards cyber risk and upgrades to CRM and ERP platforms. That matters for planning. If you are replacing or upgrading your CRM this year anyway, that project is the cheapest moment to design AI into it rather than bolting it on in 2028.

The trust gap is a design problem, not a training problem

A second data point landed this week. Notion’s Global AI Transformation Study 2026, reported by Technology Decisions on 23 September, found Australia leads the region on tracking AI’s measurable business impact, at 76 per cent. It also found the widest confidence gap in the region: 61 per cent of Australian decision-makers feel highly confident in their company’s AI capabilities, against 26 per cent of knowledge workers. Across Asia-Pacific, 77 per cent of knowledge workers said they would use AI more if they could trust it not to make mistakes on important work.

So Australian leaders are measuring. Their staff do not believe the numbers, or at least do not trust the tools enough to use them where it counts. Treasury names the same trust problem from the other end.

Most businesses answer this with a training session. That rarely works, because the distrust is usually rational. A general chatbot that has never seen your price list, your policies or your customer history will make mistakes on important work, and your staff know it. Trust comes from systems that are wired to your real data, that show where an answer came from, and that hand off to a person when they are unsure. You build that; you cannot lecture it into existence.

Days 1 to 30: inventory, data rules and one owner

The first month produces decisions, not software.

List what is already in use. Ask every team which AI tools they use, paid or free, and what they paste into them. In most mid-sized businesses this turns up several tools nobody approved and at least one team pasting customer details into a free account. You cannot integrate what you have not found.

Map the five processes that cost the most time. Not the most interesting ones. Look for work that is repetitive, rule-heavy and touches a system of record: enquiry triage, quote preparation, invoice and document handling, rostering, support tickets. Write down how many hours a week each takes and who does it. That becomes your baseline.

Write the data rules. Decide which data can go to which tools. Customer personal information, commercial terms and health or financial details need a clear answer before anything is connected. If some of that data should never leave your environment, that is a case for private AI rather than a public tool.

Check the automated decision rules. From 10 December 2026, businesses covered by the Privacy Act must explain in their privacy policy when a computer program makes, or substantially helps make, decisions that significantly affect people. If any of your five processes decides something about a customer, design for that now. We covered the detail in our post on the automated decision-making obligation, and the underlying rules sit in the OAIC’s APP 1 guidelines.

Name one owner. Someone senior who can make a call on a process change without a committee. Projects without one stall in month two.

Days 31 to 60: three integrations, not thirty

Pick three processes from your list and connect AI to the systems where the work actually lives. Three is enough to prove the pattern and small enough to finish. Good first candidates share three traits: the input is already digital, a person checks the output today anyway, and the result lands in a system you already own.

Typical first integrations look like this:

• Inbound enquiries into the CRM. Emails and web forms are read, classified, matched to an existing customer and given a draft reply, with a person approving before anything is sent.
• Documents into the finance or operations system. Invoices, purchase orders or applications are read and the fields extracted into the ERP or accounting package, with exceptions queued for a human.
• Answers from your own knowledge. An internal assistant that answers staff questions from your policies, product data and past jobs, and cites the document it used.

Each one attacks a barrier from the RBA list. The integration work deals with legacy systems. The human approval step deals with trust. Citing sources deals with the 77 per cent of workers who want to know the answer is right before they rely on it.

Build security in from the first line of code, not after launch. An AI connected to your CRM can be manipulated through the content it reads, which is why we treat secure AI system development as part of every integration rather than an extra.

Days 61 to 90: measure against the baseline, then decide

By day 60 the three integrations should be running with real staff on real work. The last month is about evidence.

Measure the same things you wrote down in month one: hours per week, turnaround time, error or rework rate. Add one number the Notion study suggests most businesses skip: how often staff actually use the tool when they are not being watched. Usage is the honest measure of trust.

Then make three decisions. Which of the three integrations earns a wider rollout. Which one gets switched off. And which two processes from your original list go next. At that point you have a working pattern, a baseline, and a business case built from your own numbers rather than a vendor’s slide.

That is the difference between the two-thirds and the one in ten. It is not a bigger budget. It is AI connected to the systems that run the business, with a measurement that proves it.

What this costs and how we run it

We do not publish a fixed price for integration work, because the cost is driven by things we cannot see from outside: how many systems need connecting, whether they have usable APIs, how clean the data is, and how much of it has to stay inside your own environment. What we can tell you is how the cost gets fixed. The free AI audit maps your processes and systems, and from that we give you a costed plan before you commit to any build.

For businesses that want help with the first 30 days itself, our AI consulting services engagement is built around a costed plan and a first working system in 30 days. When the plan calls for something custom, our AI development solutions team builds and runs it in production.

Why an agency rather than doing it yourself? Mostly the skills gap the RBA survey names. A mid-sized business rarely has a data engineer, an integration developer and someone who understands AI security sitting idle. Why not a large consultancy? Because the first 90 days need people who write code against your CRM, not a strategy deck about it.

Start with the audit

If your business is in the two-thirds that has tried AI but not in the one in ten that has integrated it, the next step is concrete. Request the free AI audit and we will review your current tools, your highest-cost processes and the systems they touch, then send you a short written list of the first three integrations we would build and what each would take. If you would rather talk it through first, book a meeting and bring your list of processes.

Frequently asked questions

What does the 2026 Intergenerational Report say about business AI adoption?

It says two-thirds of Australian businesses report adopting AI in some form, but fewer than 10 per cent describe that adoption as significant, with similar results among small and medium enterprises. It also says low trust in AI helps explain why measured productivity gains have been very modest so far.

What is the difference between using AI and integrating AI?

Using AI usually means staff working in a standalone chatbot and copying results back into other systems. Integrating AI means connecting it to your CRM, ERP, document stores and workflows, so it works on your real data and its output lands where the work happens, with a person approving where it matters.

Why do AI projects stall in mid-sized businesses?

The firms in the Reserve Bank's survey named skills shortages, low digital readiness, uncertain return on investment, legacy system integration, cost and regulatory uncertainty. None of those is about the AI model itself, which is why a plan focused on data, systems and measurement works better than buying another tool.

Do we need to worry about privacy law when integrating AI?

Yes. From 10 December 2026, businesses covered by the Privacy Act must explain in their privacy policy when a computer program makes, or substantially helps make, decisions that significantly affect individuals. If an AI integration decides anything about customers, design the disclosure and a human review step in from the start.

How long does a first AI integration take?

The plan above allows 30 days for inventory and data rules, 30 days to build three integrations and 30 days to measure them. Actual timing depends on how many systems are involved and whether they have usable APIs, which the free AI audit establishes before any build starts.