How to build an AI roadmap for an Australian business or organisation

Summary

This article gives you an AI roadmap for the next six to twelve months, in four stages, with what to decide and what to measure at each. It is one of the Putting AI to Work series for Australian businesses and organisations, and the other articles in the series are mapped below. It is written for small, medium and large organisations alike.

Skip to the AI Roadmap Template ↓

Background

An AI roadmap for a business or organisation is a short plan for six to twelve months, as long as it needs to be and no longer, in four stages. Decide what you want AI to improve and measure where you start. Let everyone experiment safely and bank the quick wins. Run one shared process with AI and a named owner. Then repeat it and connect systems only where the next project needs it. Each stage ends with a decision the owner makes on the numbers.

The federal guidance will not write that plan for you. The Guidance for AI Adoption comes in two versions. The short one, called Foundations, is the one the National AI Centre wrote for small and medium enterprises (SMEs) and not-for-profits, and it sets out six essential practices and 30 actions. The long one, Implementation Practices, runs to 50 pages. Not one of the 30 Foundations actions mentions a strategy, a business goal or a benefit to measure.¹² The Voluntary AI Safety Standard it replaced asked for “an AI Strategy”.³

The Centre (the National AI Centre, throughout this guide) has planning pages that fill part of the gap. They tell you to time a task before AI and again with it.⁴ No Australian government survey then records whether the business’s AI paid off. The Australian Bureau of Statistics (ABS) counts whether a business uses AI, and says its question “is not designed to measure intensity or extent of use”. The Centre’s adoption tracker (a survey of SME owners) asks whether they agree AI “can help”.⁵⁶ So the baseline you take in month one (the timing of each process before AI touches it) is the only evidence your roadmap will have.

The policy that sits under the roadmap is in How to write an AI policy, and the policy itself is in the AI policy template. Choosing the software and connecting it to your systems is in Choosing AI software and connecting it to your systems, and whether a role should go to a person or to AI is in Should your next hire be a person or AI?.

Checked against the National AI Centre’s Guidance for AI Adoption and its planning pages, the Voluntary AI Safety Standard, the ABS, the National AI Plan, the Productivity Commission, the Fair Work Ombudsman, ASIC, the Federal Court and business.gov.au on 18 October 2026; the sign-off section was re-checked against the Office of Parliamentary Counsel and the Digital Transformation Agency on 5 November 2026 (notes 30 and 31). The counts of actions and practices are ours, made from the documents as published. This covers Australian businesses and organisations of any size that use AI tools bought from a vendor, not businesses that build them. It is not legal advice; where a law applies, we name it so you can check it. We re-check it quarterly.

Methodology

A long article made quickly with AI can be slop, and the risk is real. We would rather say how this one was made than have you guess.

1. The questions came first. The issues this article addresses are ones the author has worked through with clients facing them, in his consulting practice.
2. AI did the research and the first draft. Claude Code, running Claude Fable 5.1, read the primary sources and drafted the article to the house standard: every claim tied to a source, and every source opened by a person before the article leaves draft. Where an article says "we read" a document, or all 121 of something, the model read every one of them in full and did the counting; the author checks the method and a sample, not every page.
3. The author did the first review. The author read the draft end to end, cut what did not hold, and edited the body.
4. Other AI tools checked it. A second and third model read each article for errors and for one model's blind spots. This article has not had that check yet.
5. Specialists review it last. Reviewers across technology, law and human resources read the final draft. This article has not had its specialist review yet, and no law firm has reviewed it.

Why use AI at all, given the risk? Because at the rate AI is moving it would be remiss not to use the top frontier models to synthesise the current state of play, so that a business decides from the most up-to-date information there is. Feedback and corrections are welcome, and so are volunteers, law firms among them, who want to make this the best resource available to Australian businesses: write to the editor.

The federal guidance tells you what to govern, not what to aim for

The Guidance for AI Adoption comes in two versions. The Implementation practices describe Foundations as “a high-level, accessible guide … for organisations such as small and medium-sized enterprises and not-for-profits”.² Foundations sets out six essential practices and 30 numbered actions under them: 13 marked “Getting started” and 17 “Next steps”. It says “You don’t need to implement everything at once. Get started across all six essential practices”.¹ Every one of the 30 is about governance: who is accountable, who is affected, what the risks are, what you tell people, how you test, how a person stays in control. None asks what the AI is for.

We searched each document for the words that would carry a goal or a benefit, and read each hit in context:

Document	Asks for a strategy or a goal?	Asks you to measure a benefit?	Asks for a baseline before you start?
Voluntary AI Safety Standard, September 20243	Yes. Guardrail 1 lists “an overall owner for AI use” and “an AI Strategy”; 1.2.2 asks for the “strategic intent to deploy and use AI systems in line with the organisation’s strategy and values”	In part. 1.2.7 asks for a process “that supports mapping from business targets to system performance”	No
Guidance for AI Adoption: Foundations, October 2025, the SME version1	No. The words strategy, goal, objective, benefit, value and success do not appear in any of the 30 actions	No	No
Guidance for AI Adoption: Implementation practices, October 20252	Yes, in 2 of 73 practices: 1.4.1 (the “strategic intent to develop and deploy AI systems in line with organisational strategy and values”) and 4.1.1 (the register records “purpose and business goals”)	In 2 of 73: 2.1.5 (identify “the potential benefits and harms”) and 5.2.2 (a deployment process “that maps business targets to system performance metrics”). The word “productivity” does not appear	No
National AI Centre, Planning for AI pages, 202674	Yes. “A clear AI strategy … should define: what problem you’re trying to solve, how you want to improve it, how you’ll measure success or value”	Yes, with a method	Yes. “Measure how long the task takes today”
National AI Centre adoption tracker6	Not applicable (a survey)	It asks “Do you agree that AI can help you achieve the following business outcomes?”, a question about expectation	No
ABS Characteristics of Australian Business, 2024-255	Not applicable	No. The question “is not designed to measure intensity or extent of use within the business”	No
National AI Plan, December 20258	Not applicable	The Plan will “monitor adoption, skills development and responsible AI practices”	No

Three things follow for an owner. First, the strategy is your job. The 2024 Standard asked for one, the 2025 version written for small and medium businesses does not, and only the Centre’s web pages put it back. Second, if you follow Foundations to the letter you will have a well-governed AI program and no way to say whether it was worth doing, because none of its 30 actions asks you to measure a benefit.

Third, no public number will benchmark you. The tracker’s methodology page says the survey covers “business outcomes that SMEs are achieving through AI adoption”, but the question it publishes asks whether owners agree AI “can help” achieve them, which records belief, not outcome.⁶ The ABS survey is biennial, and the ABS says it “will include additional content in the 2026–27 BCS to provide further insights into how businesses are adopting and using AI”.⁹

The Implementation practices are the long version: 73 numbered practices, of which 10 are marked as for developers only and 3 more carry a developer-only part (the guidance marks practices “(DEP) or (DEV)” where they apply to one side).² A business that buys its AI is left with at least 63, written for “governance professionals and technical experts”. A roadmap for a business of a few dozen people starts from the 13 Foundations “Getting started” actions, and the example below covers all of them by the end of month six.

A roadmap has four stages, and each ends with a decision

The stages below are ours. They follow the Centre’s own sequence for a single process (map it, identify the opportunities, redesign the workflow)⁷ and its advice to start small: “Choose a low-risk task where AI can assist, not decide”, “Agree on boundaries and what ‘good’ means before testing”, and “Keep early experiments time-boxed and easy to review”.¹⁰

The stages are a loop, not a project with an end. AI is improving at a pace no plan can fix for a year, and what it can do today is already more than most businesses have absorbed. So the roadmap is run in short stages, with a decision on the numbers at the end of each, and the twelve-month review rewrites the plan from evidence rather than closing it. The rebuild, parallel run and cutover of stage 2 then goes on process by process for as long as there are processes worth moving.

Stage 0: decide and measure (month 1). The owner decides what AI is meant to improve, what is out of scope, which tool and plan the business uses (the one general-purpose AI assistant staff will work in, such as Claude, ChatGPT, Copilot or Gemini, on a business plan rather than personal or free accounts), and who owns AI overall. Foundations asks for exactly that last one, a senior leader who owns AI governance overall.¹ The policy goes in now, before anyone is trained, and the candidate processes are timed before anything changes.

Stage 1: everyone experimenting, safely (from month 2). Everyone is trained first: a session on the policy and the approved setup, with each person trying the tool on one of their own tasks while someone who uses it sits with them, and the training open to anyone again later. Then staff use the approved tool on their own work, inside the policy. The approved assistant is connected, read-only, to the tools staff already work in, email, calendar, documents and chat, which for most Australian businesses means Microsoft 365 or Google Workspace. An assistant that can read the calendar and the shared drive is where most of the quick wins come from. Write access to any core system waits for the six rules in Choosing AI software and connecting it to your systems. The quick wins are written down so anyone can repeat them. One shared process is chosen as the first project, and its owner is named.

Stage 2: one shared process (months 4 to 6). A process that several people touch is mapped and handed to one or two people, carved off from their usual work, who rebuild it with AI in it. The rebuilt process runs in parallel with the current one, on the same inputs, until its outputs match. Then the business cuts over, and a named person is accountable for it from that day. It is measured on its own numbers against the month-one baseline. The six-month review decides whether stage 3 is funded, and the same one or two people take the next process. The pattern is the one Salim Ismail’s The Organizational Singularity calls parallel-run-then-deprecate: a separate “edge twin” of three to five people plus AI agents (AI tools that carry out steps on their own rather than drafting for a person) rebuilds a workflow, runs it beside the original and replaces it once it outperforms, then takes the next, easiest first.¹¹ In a business of a few dozen people the twin is one or two people, and the bar we set is that the outputs match before anything is switched off.

Stage 3: repeat, and connect what needs connecting (months 7 to 12). The stage 2 pattern (rebuild, run in parallel, cut over when the outputs match) is repeated on the next two or three candidates. Systems get connected only where the next project needs it, because connecting a system that no project uses is cost without return. The twelve-month review rewrites the roadmap from evidence.

A business of under 20 staff can compress this. Stage 0 is an afternoon, stage 1 is a month, and the first shared process may be the owner’s own inbox or the month-end paperwork.

An example roadmap, month by month

The composite is a services business of a few dozen people: an office team, a small finance function, a customer service desk, and an accounting system, an email suite and a shared drive it already pays for. The numbers in the last column refer to the Foundations actions each step carries out.¹

When	What happens	Decisions the owner makes	What gets measured	Foundations actions covered
Month 1	Name the AI owner. Adopt the AI policy. Pick the approved tool and plan. Start the AI register. Screen each proposed use for risk. List three to five candidate processes and time them for two weeks	What AI should improve this year; what is out of scope (for example, anything that decides about a person); the tool and plan; the budget for stage 1; what you will tell staff about freed time	Baseline per candidate: runs a week, minutes a run, share of runs needing rework; a team readiness check	1.1, 1.2, 3.1, 4.1
Months 2 to 3	Train everyone on the approved setup. Connect the assistant, read-only, to email, calendar, documents and chat. Each team tries the quick wins and writes down the ones that work. Turn on the account security and admin settings. Tell customers where they will meet AI. A person checks anything that leaves the business. Ask the vendor for its test evidence. Score the candidates and choose the first project	Which quick wins become standard; the first project; its process owner; whether the change will significantly affect anyone’s job (if so, consultation starts)	Who uses the tool weekly; quick wins written down; before and after on two or three tasks	4.2, 5.1, 5.4, 6.1, 1.4
Months 4 to 6	Map the first process with the people who do it. Carve off one or two of them to rebuild it with AI’s job and the person’s job written down, and pick the oversight pattern. Run the rebuilt process in parallel with the current one until the outputs match, then cut over. Assess its risks and who it affects; set up a way to raise problems. Write down what the vendor is responsible for, and what the system cannot do. Test against acceptance criteria before go-live. Build in a way to pause it, train the people who oversee it, and keep the old way available. Monitor after go-live	Build, buy or use what you have; go-live, signed by the process owner; the oversight pattern and its limits; at month 6, fund stage 3 or stop	The process’s own numbers against baseline (turnaround, first-time accuracy, volume per person); share of AI outputs a person changed; what the freed time went to	1.3, 1.5, 2.1, 2.2, 2.3, 3.3, 3.4, 4.3, 5.2, 5.3, 6.2, 6.3, 6.4
Months 7 to 9	Repeat the pattern on the next two candidates. Write down how an AI incident is reported and reviewed. Read the feedback for patterns across processes	The next projects, in order; whether any needs a connection to a core system (if so, the rules come first)	The same process numbers for each new project; incidents and near misses	3.5, 2.4, 3.2
Months 10 to 12	Connect systems where the projects need it, and record what each vendor tells you about the components. Turn the policy into a short governance framework. If anything now affects customers or staff decisions, set up a way to explain its outcomes and stress-test it	What year two funds; what stops; whether any use needs independent testing	Year totals against baseline; staff readiness check re-run; capacity: did volume grow without a matching hire?	1.6, 4.4, 4.5, 5.5, 5.6 if screening flagged it

By the end of month six the example has carried out all 13 “Getting started” actions and 9 of the 17 “Next steps”. The rest fall in months seven to twelve, and one (independent testing, 5.6) applies only if the risk screen flagged a use. The guidance itself says to begin with the getting-started actions and “move to the ‘next steps’ actions” as AI use expands.¹ Nothing in the table needs a consultant. Two of its rows need care: the consultation decision in months two to three, and the sign-off in months four to six, both covered below.

How the roadmap is run. One agenda item a month, at a meeting you already hold: what moved between stages, the numbers, the decisions due, and the next project. One rule keeps a small team’s roadmap short. Nothing new joins the roadmap until something on it finishes or stops. Staff remain free to try things inside the policy, and that is where the next candidates come from.

Quick wins are high fit and low impact, and need nothing new

The Centre’s opportunity matrix is the clearest test we found. Score a task for AI fit, then ask what happens if the output is wrong. High fit and high impact: “prioritise for AI rollout”, with care. High fit and low impact: “treat as a quick win for AI use”. Low fit and high impact: “fix the process first, then reconsider AI”.¹² Its rule for the last group is the one to keep: “If a simpler solution can deliver 80% of the benefit at 20% of the cost or complexity, start there.”¹²

In practice a quick win is a task one person does often, where AI drafts and the person decides, using a tool the business already pays for, and where a bad output costs a minute to fix. Drafting a reply that a person reads and sends, turning meeting notes into a task list, reading receipts or invoices into a sheet that a person checks, and summarising a report you already produce.

The roadmap’s job is to write each one down once it works, so the next person does not have to rediscover it, and to put it in the AI register (the business’s own list of its AI uses and who owns each), which Foundations lists among the first actions: “Create and maintain an AI register.”¹

A quick win is not a project. It changes one person’s task and needs no redesign. The moment a quick win is shared across a team, scheduled, or allowed to act without a person pressing send, it has become a project and goes through the candidate process below.

Choosing candidate projects: score them, then order them

The four tests in AI business process automation: where to start (volume, rules, cost of a mistake, where the data lives) decide whether a process is a candidate at all. The Centre’s version scores five criteria from 1 to 5: volume, repetitiveness, data availability, error tolerance and current cost, with a total from 5 to 25 and 13 or more counting as high fit.¹² The fifth criterion, what the process costs you today, is the one to add to the four tests, because it is the one the roadmap’s measures will be judged against.

A roadmap then needs three more things from each candidate that a single-project test does not ask.

1. What it depends on. A process whose data sits in a system with no export, or that needs a connector (the link that lets an AI tool read or write another system) nobody has approved, waits until the dependency is cleared. Write the dependency next to the candidate, with its owner.
2. Whether it is shared. The first project should be one that several people touch, because that is where a written-down process pays back. A task one person does is a quick win, not a project.
3. Who owns it. If nobody can be named as the owner of the process, it is not ready. That is the sign-off question, below.

Before mapping, the Centre’s mapping page asks you to record for each step “time – how long the step usually takes to complete”, alongside who does it and what goes in and out.¹³ That table is the baseline. Its worked example, an accounting practice of under 20 staff, finds client onboarding taking “5 to 7 hours over 2 weeks” and sets a target of 2 to 3 hours after redesign.¹³¹⁴

The decisions that sit at the top

Most of a roadmap is a list of decisions with dates on them. The ones below sit at the top of the organisation, and “the owner” in this guide means whoever that is. In a business of a few dozen people it is the owner. In a larger organisation it is the chief executive, with the board approving the roadmap and any change that significantly affects staff, as clauses 2(c) and 2(f) of the board AI governance policy reserve to it. It never means the process owner. A process owner answers for one process. These decisions are about the organisation, and they cannot be delegated to whoever is keenest on AI.

The third decision below, what happens to freed time and to roles, is where the difference matters most. If a project would move a general manager’s twenty staff to another department, that is not the project’s call and not the general manager’s. It is a decision about roles, the third layer in How to write an AI policy, made at the top, approved by the board where there is one, and reached only after the consultation steps the modern award (the legal minimum pay and conditions set for an industry or occupation) requires, which are in Should your next hire be a person or AI?. Whether those twenty people keep their jobs, and whether the work they move to has the meaning the old work had, is answered in that decision and said to them by the people who made it, not discovered when the process goes live.

The decisions:

• What AI is for, and what is out of scope. The Centre’s readiness check asks “Have we defined what success will look like?”¹⁵ Out of scope is as useful. A first roadmap can leave out customer-facing AI and anything that decides about a person, and say so.
• The tool and the plan. One approved general-purpose assistant, Claude, ChatGPT, Copilot or Gemini, on a business plan, chosen on what it does with your data.
• What happens to freed time, and to roles. Say it before training starts, not after the first role gets lighter. The Centre’s return on investment (ROI) page puts the economics plainly: “Time saved only delivers value if it’s redirected to useful work, such as serving customers, improving quality or growing the business.”⁴ Whether a role goes to a person or to AI, and how to share the benefit, is in Should your next hire be a person or AI?.
• Build, buy or use what you have, for each project. The intake questions (what a proposed project must answer before it is approved) and the vendor questions are in Choosing AI software and connecting it to your systems.
• The process owner for each project, named before the project starts.
• Go, or stop, at each review. Each stage is funded on the evidence of the one before. Stopping a project that did not move its numbers is a result, and the register should record it.

Who signs off when work moves out of a role and into a system

When a task moves from a person’s job into an AI system, accountability does not move with it. It goes to the named owner of the process, never to the system. That owner signs off in one of two ways. Where volume is low or the stakes are high, they sign off decision by decision. Where volume is high, they sign off the rules, the thresholds and the limits the system works within, check a regular sample of what it did, and answer for the outcomes either way.

The checking can be delegated. The accountability cannot. On a support desk handling hundreds of tickets a day, the owner does not read every reply the system sends or approve every escalation. They name the reviewers who do, write down what a reviewer looks for and what they send back, decide how big the sample is and how often it is read, and answer for what the reviewers find and for what they miss. A reviewer who is given the rules and the sample is doing the owner’s checking. The owner is still the one who signs, and the one the board asks.

The Centre’s oversight patterns line up with this. “Check before acting” is for high impact, where “A person approves every AI output before using it”. “Watch while acting” is for medium impact at high volume. “Spot-check” is for low impact at very high volume, where “A person reviews a sample of AI outputs”.¹⁴ The federal guidance says the same, a named person accountable for every AI system and an accountability that leaders cannot delegate or outsource. The quotes and the law behind them are in Who is accountable when AI makes the decision.¹³

Above the process owner sit the directors. Their duty of care under section 180 of the Corporations Act (the duty to act with the care and diligence a reasonable person in their position would) extends to the adoption and use of AI. In ASIC v Bekier, the Star Entertainment case decided in March 2026, the Federal Court said technology may assist that duty but cannot displace it. The duty and the judgment are in Your duty of care over AI.¹⁶¹⁷¹⁸ How a business sets the rules a system may act within is in When the work leaves a role, the process owner signs off, a section of Choosing AI software and connecting it to your systems.

The Commonwealth’s own rule for automated decisions runs the same way: where legislation lets a program decide, “a natural person must be nominated (by position or title) as having ultimate responsibility for the decision”, and since October 2026 every agency must name “an accountable use case owner for each in-scope AI use case”.¹⁹²⁰ Those rules bind agencies, not businesses. The roadmap borrows the shape.

In a roadmap, this comes down to one rule. No project starts without the name of its process owner written next to it, and no project goes live until that person has signed off the test results. The Implementation practices ask for the same, a documented go-live authorisation from the accountable person based on test results (5.1.5), and the intake and sign-off rules are in Choosing AI software and connecting it to your systems.²

When the roadmap changes roles, consultation starts at the decision

Modern awards and enterprise agreements (the terms a business and its own staff agree in place of the award) carry a consultation term, a clause that obliges the employer to tell affected staff about a change and hear their views. It starts when the employer decides on a major change in technology likely to significantly affect staff, and Should your next hire be a person or AI? sets out what the clause requires and when it bites.²¹

A roadmap is a schedule of decisions, so it should carry the consultation step on the same line as the decision that triggers it. In the example, that is the choice of first project in months two to three, before the redesign in months four to six. The Centre’s readiness check asks the same question twice: “Have workers been meaningfully consulted?”¹⁵ The policy side is in How to write an AI policy.

Measuring it: the baseline is the measure

The Centre’s method for time savings is three steps: “Measure how long the task takes today. Measure how long it takes with AI support. Multiply the time saved by the cost of staff time to estimate value.” It adds that “You may need to track time saved over several weeks or months to get a clearer picture”, and lists the other places value shows up: quality and accuracy (compare “error rates or rework costs before and after”), capacity and scale, and revenue and customer outcomes.⁴ The Centre’s executive director put it in one line in September 2026: “If you want to save time, know where the time is going now.”²²

Three cautions from the evidence shape what you measure.

People’s estimates of their own speed-up are not a measure. In the 2025 randomised trial by Model Evaluation and Threat Research (METR), experienced software developers “expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had sped them up by 20%”. Measured, they took 19 per cent longer.²³ METR now says it believes those early-2025 results “no longer reflect the current impact of AI models”, and the finding was in one occupation.²³

The lesson that survives is about measurement rather than about AI. Ask the process, not the person. Count invoices matched first time, days from enquiry to quote, replies sent per hour, and keep the staff survey for confidence and trust.

The setup costs come first, and they are easy to miss. The Centre’s ROI page lists the costs that are “often overlooked”: “staff time for training”, “testing and change management”, “data preparation”, “governance activities” and “ongoing oversight”.⁴ At national level the Productivity Commission makes the same point in its own terms: “some productivity benefits may go unmeasured in the economic statistics”, and “investments required for AI adoption may be challenging to measure, potentially delaying gains in measured productivity improvement”.²⁴ Plan for the first months to show the costs before the returns. That is why the funding decision sits at month six, not month two.

Hours saved is a weak headline. Freed time is redeployed, and a saving that disappears into the day does not show up anywhere. Measure at three levels instead:

Level	Measure	Where the number comes from
The task	Minutes per run, rework per run, before and after	Two weeks of timing in month 1, repeated after go-live
The process	Turnaround, first-time accuracy, volume handled per person, share of AI outputs a person changed	The process’s own records: the accounting system, the inbox, the job log
The business	Did volume or revenue grow without a matching hire; did the freed time go where the owner said it would	The month-6 and month-12 reviews

The Centre’s redesign page gives the monitoring loop for the middle row: “Capture changes – record which outputs people changed and why”, “Monitor results – track accuracy, speed and quality of outputs” and “Review regularly – schedule regular checks to decide what to keep, change or stop”.¹⁴

Where you stand nationally. The ABS found “Around 12 per cent of Australian businesses reported the use of AI in their workplace in 2024–25”, with about 35 per cent of large businesses, 22 per cent of medium businesses, and “around 11 per cent” of small and micro businesses.⁹ The Centre’s tracker found “43% of Australian SMEs reported some level of AI adoption” in the December 2025 to February 2026 quarter, and that among users “the most common safeguard is checking AI outputs before they affect customers”, in place at about half of them.²⁵

The two use different questions and different size bands. The tracker counts businesses of 200 to 500 employees as SMEs, where the ABS calls 200 and over large.⁶ Neither tells you whether your AI is working. The Productivity Commission said in January 2024 that “there is scant empirical evidence that AI uptake causes an increase in firm productivity”, because firms that use AI may simply be the more productive ones.²⁶ Only your baseline answers the question for your business.

Free help with a roadmap exists, if you fit a centre’s sectors

The federally funded AI Adopt Centres “provide free specialist services for eligible SMEs in National Reconstruction Fund (NRF) priority sectors”, and business.gov.au lists their services as “training courses, one-on-one consultations and roadmaps, technology demonstrations and AI safety guidance”, and the ARM Hub centre offers “a customised AI adoption roadmap for their business”, free.²⁷ The Centre’s own page lists “AI adoption roadmaps aligned to your goals” among the services the centres can offer.²⁸ The National AI Plan of December 2025 commits “$17 million for the AI Adopt Program to support SMEs” and says the government “will bring this program into the NAIC’s remit”.⁸

Time it into your plan. The program’s guidelines fund the centres “over four years until 2026/27”, require the centres’ projects to be completed “by 31 March 2027”, and say “Further funding for this program is not guaranteed”.²⁹ A twelve-month roadmap that starts now runs past that date, so use a centre in the first two stages.

The Centre’s August 2026 blog describes one centre as “For: Businesses in all industries”, which is wider than the business.gov.au sector test. Check with the centre before you count on it.³⁰

The government does not publish a roadmap template. The Centre’s tools page, as at 18 October 2026, offers a business process mapping template, a data quality checklist, questions to ask AI suppliers, an AI policy guide and template, an AI register (as a document and a spreadsheet), a team readiness canvas and three team scenarios, and no roadmap or strategy template.³¹ The table above is meant to be copied, and the AI Roadmap Template below is its short form.

The AI Roadmap Template

Copy the roadmap below into a document and fill it in. It holds the four stages with the decision each ends in, the candidate processes with the baseline you take in month one, and the decisions only the owner can make. The example above is the long form, and this is the short form that goes on the agenda of a meeting you already hold. Make it as long as the plan needs.

What changed

This is version 1.0, first published 18 October 2026 and last updated 5 November 2026. The version moves at a major change, such as a law or a piece of guidance the guide rests on changing; a smaller fix moves the updated date only. Earlier versions are kept from the first major change on.

No changes since first publication.

What we could not confirm

• How many Australian businesses keep any baseline for their processes, such as time or error counts. We found no source that asks. The ABS question counts use, the tracker’s questions ask about expectations and practices, and we did not find a survey that asks whether a business measured its AI.
• Whether the Bekier judgment is under appeal. We read the liability judgment from the Federal Court’s own published copy on 19 October 2026 and rely on it for one point. The penalty judgment of 17 June 2026 records at [87] that Mr Bekier intended to appeal the liability findings. We found no notice of appeal and did not search the Court’s list. The passage concerns the directors of a listed company and is the judge’s observation, not a rule for small businesses.
• The dates of the National AI Centre’s planning pages. The “Planning for AI” pages print no publication or update date. The downloads on its tools page are dated 21 April to 5 May 2026.
• Whether the Foundations PDF changed after October 2025. The file on ai.gov.au sits in a July 2026 folder, is labelled October 2025, and its tools page says it was published 5 May 2026. Its 30 actions match the web page word for word on 18 October 2026.
• How many AI Adopt Centres there are, and who is eligible. business.gov.au lists four centres and a sector test. The Centre’s blog of 3 August 2026 names three and describes one as open to all industries. The program’s 2023 guidelines set 31 March 2027 as the end of the centres’ funded projects, and no page says what happens after.
• The tracker’s question count. Its methodology page says the survey asks “4 questions” and then lists 5.
• A numbering slip in the Implementation practices. The PDF numbers the practices in section 4.2 as 4.1.2 to 4.1.7. The web version numbers them 4.2.1 to 4.2.6. The total of 73 is the same either way.

We re-check this piece every quarter, and when the ABS publishes its 2026-27 business survey or the National AI Centre issues a new version of its guidance. The date at the top of the page changes with it.

Notes

1. National AI Centre, Guidance for AI adoption: foundations, web page and the PDF it links (“Guidance-for-AI-adoption-foundations_2.pdf”, labelled October 2025, published on the page 5 May 2026), “How to use the guidance” and practices 1 to 6 (actions 1.1 to 6.4), read 18 October 2026. The count of 30 actions and the word search are ours. ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩

2. National AI Centre, Guidance for AI adoption: implementation guidance, October 2025, introduction (pp 2 to 4) and implementation practices 1 to 6, including 1.4.1, 2.1.5, 4.1.1, 5.1.5 and 5.2.2, read 18 October 2026. The count of 73 practices (10 wholly for developers, 3 with a developer-only part) is ours; the web version of the page, published 5 May 2026, carries the same practices. ↩ ↩ ↩ ↩ ↩

3. Department of Industry, Science and Resources, Voluntary AI Safety Standard: the 10 guardrails, published 5 September 2024, guardrail 1 and items 1.1, 1.2.2 and 1.2.7; and the standard’s landing page, updated 2 December 2025, which says the Guidance for AI Adoption “evolves” the standard; read 18 October 2026. ↩ ↩ ↩

4. National AI Centre, Measure return on investment, “Assess the full cost” and “Time and productivity gains”, undated, read 18 October 2026. ↩ ↩ ↩ ↩ ↩

5. Australian Bureau of Statistics, Characteristics of Australian Business, 2024-25, released 25 June 2026, the ICT section, read 18 October 2026. ↩ ↩

6. National AI Centre, How we collect data on AI adoption, “Respondents”, “What we ask SMEs” and survey question 3, undated, read 18 October 2026. ↩ ↩ ↩ ↩

7. National AI Centre, Prepare your business, undated, read 18 October 2026. ↩ ↩

8. Department of Industry, Science and Resources, National AI Plan, December 2025, “Implementing the plan and measuring success” (p 8), the funding list, and Action 4, “Scale AI adoption” (pp 18 to 19), read 18 October 2026. ↩ ↩

9. Australian Bureau of Statistics, media release, Business adoption of artificial intelligence accelerates in 2024-25, 25 June 2026, read 18 October 2026. ↩ ↩

10. National AI Centre, Start small and build momentum, “Where to start”, undated, read 18 October 2026. ↩

11. Salim Ismail with contributors, The Organizational Singularity: How AI Breaks the Firm and Rewrites It to Solve Everything, OpenExO, v25, June 2026, chapters 8 to 10, read 6 November 2026. “An Edge Twin is a board-mandated, CEO-sponsored parallel business unit, typically 3-5 humans plus an advanced agent cluster, that rebuilds specific mothership workflows from scratch using the Intelligence Stack, proves they outperform the original on core benchmarks, and then replaces them”; “parallel-run before you deprecate anything”; “Migrate workflows easiest first”; “detect → spawn → migrate → deprecate → repeat.” ↩

12. National AI Centre, Identify opportunities, Table 1 and Figure 1 (the AI opportunity assessment matrix) and “Improve the process first”, undated, read 18 October 2026. ↩ ↩ ↩

13. National AI Centre, Map your processes, “Choose a process to map” and the worked example, undated, read 18 October 2026. ↩ ↩

14. National AI Centre, Redesign your workflow, Table 1 (oversight patterns), “Create feedback loops” and the worked example, undated, read 18 October 2026. ↩ ↩ ↩

15. National AI Centre, Get ready for AI, “Align AI to your business goals” and “Prepare and support your people”, undated, read 18 October 2026. ↩ ↩

16. ASIC, Report 798: Beware the gap: Governance arrangements in the face of AI innovation, 29 October 2024, page 34, read 18 October 2026. The report reviewed financial services and credit licensees; the sentence on directors’ duties is general. ↩

17. Federal Register of Legislation, Corporations Act 2001, section 180, compilation C2026C00432 of 19 September 2026; the register served our reader only the compilation’s front page on 18 October 2026, so the section is described from ASIC’s report. Section 180(1) read in full from compilation 149 (19 September 2026) on 19 October 2026: a director or officer “must exercise their powers and discharge their duties with the degree of care and diligence that a reasonable person would exercise” in the corporation’s circumstances and in that office. ↩

18. Federal Court of Australia, Australian Securities and Investments Commission v Bekier (Liability Judgment) [2026] FCA 196, Lee J, 5 March 2026, at [1956], read 18 October 2026 through a text proxy because the court’s site refused our reader directly. ↩

19. Office of Parliamentary Counsel, Instructor Guidance Note: Automated decision-making, last updated 5 March 2026, paragraph 12. Read 5 November 2026. ↩

20. Digital Transformation Agency, Policy for the responsible use of AI in government: Standard for accountability, version 2.1 of 21 October 2026, “Accountable use case owners”. Read 5 November 2026. ↩

21. Fair Work Ombudsman, Consultation and cooperation in the workplace: best practice guide, content last updated 6 May 2026, “Consultation about major workplace change”, read 18 October 2026. ↩

22. Lee Hickin, Executive Director, National AI Centre, You’ve got an hour back. Now what?, ai.gov.au blog, first published 18 September 2026, updated 23 September 2026, read 18 October 2026. ↩

23. METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, 10 July 2025, with the note added in 2026 that results “current as of early 2026” have been released, read 18 October 2026. ↩ ↩

24. Productivity Commission, Harnessing data and digital technology, inquiry report no. 111, December 2025, box 1.1, “How large is the productivity dividend from AI?”, read 18 October 2026. ↩

25. National AI Centre, AI adoption insights: December 2025 to February 2026, first published 7 May 2026, updated 3 June 2026, read 18 October 2026. ↩

26. Productivity Commission, Making the most of the AI opportunity, research paper 1: AI uptake, productivity, and the role of government, January 2024, section 1, p 3, read 18 October 2026. ↩

27. business.gov.au, AI Adopt Centres, undated, read 18 October 2026. ↩

28. National AI Centre, AI adopt centres, “Available support”, undated, read 18 October 2026. ↩

29. business.gov.au, AI Adopt Program: grant opportunity guidelines, November 2023, section 3 (“Grant amount and grant period”, “Project period”), read 18 October 2026; the program page says the grant is closed to applications. ↩

30. National AI Centre, Get hands-on support from an AI Adopt Centre, ai.gov.au blog, 3 August 2026, read 18 October 2026. ↩

31. National AI Centre, Planning tools and templates, downloads dated 21 April to 5 May 2026, read 18 October 2026. ↩