An AI adoption roadmap that works fits into 90 days and five stages: audit how work really gets done and pick two or three use cases, pilot them against a measured baseline, write the policy and train people while the pilots run, scale only what beat the baseline, then measure and decide what comes next. The order matters more than the tools. Companies that skip the baseline end up with plenty of AI activity and no way of proving any of it paid off.

That describes a lot of companies. McKinsey's 2026 State of AI survey found nearly nine in ten respondents report regular use of AI in at least one business function, yet only 37 per cent attribute any EBIT impact to it, and just 6 per cent qualify as high performers attributing 5 per cent or more of EBIT to AI.[1] MIT's NANDA initiative was blunter in 2025: despite $30–40 billion of enterprise investment in generative AI, 95 per cent of organisations were getting zero return.[2] Its sample was small and its headline contested, but the direction matches most other research on the subject.

Why AI Pilots Stall

The failure pattern is consistent enough to be boring. A company buys licences, runs a lunch-and-learn and lets a few enthusiastic teams experiment. Six months later usage is patchy, nobody can say what changed, and the renewal conversation gets awkward.

Usually AI has been dropped into existing workflows rather than used as a reason to redesign them. Nearly three-quarters of McKinsey's high performers report fundamentally redesigning workflows because of AI.[1]

The organisation is also, almost always, already using AI unofficially. MIT found that while only 40 per cent of companies had bought an official LLM subscription, workers at more than 90 per cent reported regular use of personal AI tools for work.[2] Microsoft and LinkedIn's 2024 Work Trend Index found 78 per cent of AI users were bringing their own tools to work, and only 39 per cent had received AI training from their employer.[3] A roadmap drawn up as though adoption hasn't started ignores the adoption that has, along with its risks.

And the problem is organisational, while most rollouts treat it as individual. Microsoft's 2026 Work Trend Index attributes 67 per cent of AI's real impact to organisational factors such as culture, manager support and talent practices, against 32 per cent to individual mindset and behaviour.[4] You can't train your way out of a structure that gives people nothing better to do with the time AI saves them.

The 90-Day Roadmap at a Glance

Ninety days is long enough to get a real result from a pilot and short enough that the executive who sponsored it still remembers why. The stages overlap on purpose:

  • Weeks 1–2: audit how work happens and choose use cases
  • Weeks 3–6: pilot against a baseline
  • Weeks 3–8: write the policy and train people, alongside the pilots
  • Weeks 7–11: scale what worked and stop what didn't
  • Weeks 12–13: measure, report and set the next 90 days

Weeks 1–2: Audit How Work Actually Happens

Start with the work. A useful audit maps where time goes, where handoffs stall and where the same task gets done forty times a week by people who would rather be doing something else. It also maps where AI is already in use, sanctioned or otherwise, which becomes the first draft of what Australia's National AI Centre calls an AI register: a record of every AI system in use and how it is used. The same guidance recommends assigning a senior leader as the overall owner of AI governance.[5] Do that in week one. A roadmap with no named owner is a suggestion.

Then pick two or three use cases; twelve is a wish list. Good first candidates share a few traits:

  • The task is frequent enough that small time savings add up.
  • The output can be measured in hours, cost, error rate or turnaround.
  • A mistake is recoverable, so nobody is betting the company on week-four prompts.
  • The team that owns the work actually wants it fixed.

Tellingly, MIT's successful buyers sourced AI initiatives from frontline managers rather than central labs. MIT also found that while front-office gains are board-friendly, back-office deployments often delivered faster payback.[2] Resist choosing the use case that makes the best slide. Our AI advisory work starts from how a business actually runs for the same reason.

Weeks 3–6: Pilot Against a Baseline

Before anyone opens a new tool, measure the current state of each chosen task: how long it takes, what it costs, how often it has to be redone. Without that number every pilot feels like a success, because the people running it are usually the ones who asked for it.

The baseline also protects you from AI's uneven performance. In a field experiment with 758 Boston Consulting Group consultants, published as a Harvard Business School working paper, those using AI on tasks within its capabilities completed 12.2 per cent more tasks, 25.1 per cent more quickly, at more than 40 per cent higher quality than a control group. On a task selected to fall outside those capabilities, consultants using AI were 19 percentage points less likely to reach a correct solution.[6] Same tool, similar people, opposite results. Vendor case studies won't tell you which side of that line your use case sits on. A pilot will.

Keep the pilot small and honest: willing users, a comparison group or a clean before-and-after measure, and success and kill criteria written down before it starts. MIT's successful buyers benchmarked tools on operational outcomes rather than model benchmarks, and worked through early failures with their partners rather than walking away.[2]

Weeks 3–8: Write the Policy and Train the People

Policy and training run alongside the pilots because the pilots surface the questions a policy needs to answer. Which data can go into which tool. Who reviews output before it reaches a customer.

Australian companies have a sensible starting point. The federal government's Guidance for AI Adoption, released in October 2025, condensed the ten guardrails of the earlier Voluntary AI Safety Standard into six essential practices, and remains non-binding.[7] Its foundations version is written for teams new to AI governance, and the practices read like a checklist for this roadmap: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control.[5]

The policy should be short enough that people read it and specific enough that they can act on it: approved and prohibited tools, data handling, output review, IP and disclosure. Write it from the usage your audit found rather than the usage you wish you had, or teams will quietly route around it. That is the approach behind our AI policy development work.

Training is where most companies under-invest. BCG's 2026 AI at Work survey of close to 12,000 people found 72 per cent say the skills expected of them have shifted, but only 36 per cent feel adequately upskilled. Of frontline employees who use AI regularly, 42 per cent save eight hours a week, yet 66 per cent receive limited or no guidance on what to do with the time they save.[8] That last figure has nothing to do with prompting skills. Train people by role, on the tools they will actually use, and decide in advance where the reclaimed hours should go.

Weeks 7–11: Scale What Worked, and Only That

By week seven you should know which pilots beat their baseline. Scale those. Stop the rest without ceremony. A pilot left to linger consumes budget and teaches the organisation that nothing is ever really measured.

Scaling rarely means buying more seats. It means rebuilding the workflow around the new capability and removing steps that existed only because the old process was slow. BCG found the share of organisations using AI to reshape workflows end to end, or to invent new business models, nearly doubled from 22 per cent in 2025 to 42 per cent.[8]

Two decisions sit inside this stage. One is build versus buy: MIT found that pilots built through strategic partnerships were twice as likely to reach full deployment as those built internally.[2] The other is who carries the change. Internal champions drawn from the pilot group will do more for adoption than any all-staff email.

Weeks 12–13: Measure, Then Set the Next 90 Days

Close the loop against the baselines from week three. Report hours saved, cost per output, rework rates, active usage and any incidents, and be specific about which changes came from the AI and which from the process redesign that travelled with it. McKinsey's high performers are twice as likely as other respondents to have defined processes for measuring AI's impact,[1] the least glamorous of their habits and probably the easiest to copy.

Then update the AI register, check that monitoring matches the risk of whatever is now in production,[5] and choose the next use cases. The second cycle runs faster because the groundwork already exists. It can also aim higher: most of McKinsey's high performers use AI to pursue growth or innovation as well as efficiency.[1]

What Being Left Behind Actually Looks Like

Plenty of businesses are still at the start line. The National AI Centre's SME tracker found 44 per cent of Australian small and medium businesses reported some level of AI adoption in February 2026, while 19 per cent said they simply don't know how to use AI in their business.[9] For larger organisations, though, the more likely way to fall behind is having AI everywhere, a licence bill to match, and no way of saying what any of it achieved.

Ninety days, a handful of well-chosen use cases and an honest baseline won't produce a dramatic board slide. They will produce a result you can defend and a method you can run again. If you want help running the first cycle, our AI advisory team does exactly that, and the easiest place to start is a conversation.

Frequently asked questions

What is an AI adoption roadmap?

An AI adoption roadmap is a staged plan that sets out which AI use cases a company will pursue, in what order, who owns them and how success will be measured. A practical version runs over about 90 days: audit how work gets done and choose two or three use cases, pilot them against a measured baseline, write policy and train people in parallel, scale only what beat the baseline, then measure and plan the next cycle.

Why do most AI pilots fail to scale?

Most pilots stall because AI is added to existing workflows instead of being used to redesign them, because nobody defined success against a baseline before the pilot started, and because no senior person owns the result. Without a baseline, every pilot looks like a success and none of them can be defended when budgets are reviewed.

How long does it take to see measurable results from AI adoption?

A well-scoped pilot can show a measurable result against its baseline within four to six weeks, and a first full cycle of audit, pilot, policy, scaling and measurement fits into roughly 90 days. Organisation-wide impact takes longer and usually comes from repeating that cycle with progressively more ambitious use cases.

Do we need an AI policy before we start piloting AI tools?

You need basic guardrails before any pilot begins, covering which data can go into which tools and who reviews output before it reaches a customer. The full policy is best written alongside the pilots, based on how people are actually using AI. In Australia, the National AI Centre's non-binding Guidance for AI Adoption offers a sensible starting framework of six essential practices.

How do you measure the ROI of AI adoption?

Measure each use case against a baseline taken before the pilot: time per task, cost per output, error or rework rates and turnaround time. Track active usage and incidents alongside those business metrics, and be clear about which gains came from the AI and which came from the process redesign that accompanied it.

References