GUIDE: AI ROADMAPEnterprise AI roadmap: how to build an AI strategy that reaches production

What is an AI roadmap, and how is it different from an AI strategy?

The terms overlap, so separate them.

• AI strategy answers why and where: which outcomes AI should move, where you will not use it, and how much risk you accept.
• AI roadmap answers what, when and who: the sequence of use cases and enabling work, with owners, durations and stage gates.
• AI readiness assessment answers can we: an honest look at data, people, processes and governance before you commit money.
• AI transformation is the outcome: AI built into how work is done, not bolted onto old processes.

A strategy without a roadmap stays a slide. A roadmap without a strategy becomes a list of pilots.

Why do most AI programmes stall before they pay off?

Adoption is no longer the problem. Value is.

• McKinsey’s 2026 survey of 1,719 respondents found nearly nine in ten organisations use AI regularly, but only 37 per cent attribute any EBIT impact to it, unchanged from 2025. About 6 per cent are high performers. [1]
• In Gartner’s survey of 2,002 CIOs and technology executives, only 13 per cent say their enterprise has realised significant or transformational value from AI, and “the likelihood of use case success remains one-in-five.” [3]
• In a Gartner survey of 782 infrastructure and operations leaders, only 28 per cent of AI use cases fully met ROI expectations. Many who reported a failure “expected too much, too fast”, and 38 per cent blamed poor data quality or availability. [4]
• Gartner also predicts that through 2026, organisations will abandon 60 per cent of AI projects unsupported by AI-ready data. [5]

The same sources show what separates the winners. Nearly three-quarters of McKinsey’s high performers “fundamentally redesign workflows” around AI, against one-quarter of others, and they are twice as likely to have committed senior leaders and defined processes to measure impact. [1] BCG’s 2026 survey of 1,330 senior leaders found the leaders “run AI as an enterprise-wide program”, invest in governance and the workforce alongside the technology, and measure impact in the P&L. [2] Gartner found CIOs with written rules for when a pilot should not go to production are 2.3 times more likely to be top performers, yet only 31 per cent have them. [3]

So the causes of stalling are predictable: no baseline, no business owner, data that is not ready, pilots with no exit criteria, and governance that arrives after the build. A roadmap exists to fix those five things.

What does an AI readiness assessment check?

An AI readiness assessment scores your organisation across four dimensions. Score each question 0 (not in place), 1 (partly), 2 (mostly) or 3 (fully and evidenced). It tells you where the roadmap needs most time; it is not a pass or fail.

Interpreting the totals (out of 12 per dimension):

• 0 to 4: foundation work first. Budget extra time for phase 4.
• 5 to 8: typical. Run a contained pilot while fixing gaps in parallel.
• 9 to 12: ready to scale. Focus on portfolio management and measuring value.

The lowest dimension sets your pace. The National AI Centre publishes free templates that fit this check: a business process mapping template, a data quality checklist, questions to ask AI suppliers, an AI policy template and an AI register template. [9]

How do you build an enterprise AI roadmap, step by step?

The template runs six phases over roughly nine months for a first wave, then repeats quarterly. Phases overlap; gates do not. Durations assume a mid-sized enterprise with two or three priority use cases. Halve them for one business unit; extend them where data or approvals are slow.

Phase 1. Mobilise (weeks 1 to 2)

Write down the two or three outcomes AI should move this year in P&L terms: hours returned, cost per transaction, cycle time, error rate. Appoint an executive sponsor with budget and a senior AI governance owner; the National AI Centre lists that and an AI policy as the first actions for any organisation. [7] An interim policy can be one page: approved tools, prohibited data, and who to ask.

Phase 2. Assess readiness (weeks 2 to 5)

Run the readiness check with the people who do the work, not just their managers. Register every AI system already in use, including AI inside SaaS tools. [7] Record current cost, time and quality for each candidate process. Without a baseline you cannot prove value later. [3]

Phase 3. Prioritise use cases (weeks 4 to 6)

Score candidates with the matrix in the next section and pick two or three: one quick, low-risk win and one that changes a core workflow. Each gets a one-page business case with baseline, target, running cost, risk tier and a stop rule, for example: “If accuracy on the evaluation set is below 90 per cent after six weeks, or handling time falls by less than 20 per cent, we stop.” Organisations with rules for abandoning low-value AI in production are 2.6 times more likely to be top performers, yet only 17 per cent of CIOs have them. [3]

Phase 4. Build foundations (weeks 5 to 16, in parallel)

Fix only the data the chosen use cases need. Gartner’s steps for AI-ready data start with aligning data to specific use cases, and warn that “if the data has issues, then the data is not ready for AI.” [5] Decide where each workload runs by data class; our private AI guide compares hosted, private cloud and on-premises. Build an evaluation set of real examples and edge cases before the solution. Start training now; people take longer than technology.

Phase 5. Pilot (weeks 8 to 20)

Pilot inside the real workflow with real users and data, not a sandbox demo, and redesign the process around the AI where it helps. [1] Measure against the baseline weekly and apply the stop rule honestly. A pilot stopped after eight weeks beats a zombie pilot that runs for a year.

Phase 6. Scale and operate (months 5 to 9, then ongoing)

Production needs what a pilot can skip: monitoring, logging, a human override point, a fallback process and an accountable owner. [7] Track running costs from day one; about one in five organisations already limit AI use because of operating costs. [1] Our cost guide shows how to cut them. Each quarter, keep, improve or retire each system and pull the next wave from the scored list.

How do you prioritise AI use cases?

Use a weighted scoring matrix, agreed with finance and risk before anyone scores. Gartner recommends assessing each use case for “feasibility, risk, cost, and expected business impact” with “a shared scoring model” so use cases can be compared and ranked. [4]

Multiply each score by its weight and add them up for a total out of 5. A worked example:

Email routing wins the first wave despite lower headline value: it proves value fast at low risk. Credit decisions wait for foundations work and stronger human oversight.

How do you measure the ROI of an AI roadmap?

Measure at three levels, and set the baseline in phase 2.

Two rules keep this honest. Report cost per completed outcome, not cost per token; our AI cost calculator helps estimate it. And count time saved only if it is redeployed to something measurable. Gartner found 44 per cent of CEOs struggle to demonstrate AI-driven financial returns, and 63 per cent see radically redesigning the operating model as a top move to accelerate them. [6]

What governance does an AI roadmap need?

Governance belongs in the roadmap from phase 1, not as a review before launch. In Australia, start with the National AI Centre’s Guidance for AI Adoption, which sets out six essential practices. It comes in a foundations version for organisations starting out or using AI in low-risk ways, and an implementation version for higher-risk or more complex use. [7][8] There is no AI-specific act; the government “will not proceed at this time” with mandatory guardrails, and existing privacy, consumer, workplace and sector laws apply. [15] The six practices map to the roadmap like this:

Essential practices from the National AI Centre’s Guidance for AI Adoption. [7]

ISO/IEC 42001 (December 2023) is the international standard for an AI management system, built on Plan-Do-Check-Act. [10] It suits organisations that already run ISO 27001 or ISO 9001, or whose customers ask for certification. The NIST AI Risk Management Framework is a voluntary US alternative, now being revised. [11]

Two Australian obligations to plan around:

• Financial services. APRA’s April 2026 letter says existing prudential standards apply to AI and expects board literacy, an AI strategy aligned with risk appetite, “an inventory of AI tooling and AI use cases” and “human involvement for high-risk decisions”. [12] Our private AI guide maps CPS 230 and CPS 234 to an LLM deployment.
• Automated decisions. From 10 December 2026, organisations covered by the Privacy Act must explain in their privacy policy when a computer program makes, or substantially and directly helps make, decisions that could significantly affect individuals. [13] Add this to the phase 4 checklist for any such use case.

General information, not legal advice. Governance and regulatory content is current as at 8 November 2026, with links to the primary source. Confirm your obligations with your compliance and legal advisers.

Should you build, buy or partner?

Decide per use case, not once for the whole roadmap.

McKinsey found 32 per cent of organisations have decided against buying at least one software product because they could build it with agentic coding tools. [1] Building is cheaper to start, not cheaper to own.

What does an AI roadmap look like in different industries?

The template does not change by sector. The use cases, data constraints and regulators do.

Financial services and insurance

Start with document-heavy work where humans stay in control: submission intake, claims document extraction, servicing queries and knowledge search. Keep credit, claims and hardship decisions behind human approval. [12] Quantana case studies include an AI-powered commercial insurance submission platform with approval governance and read-only audit records, and an insurance portal giving a global broker’s clients self-service policy information.

Retail and consumer businesses

AI in retail works best where there is plenty of transaction data and a clear metric. Common candidates:

• Demand forecasting and inventory, from sales history and signals such as seasonality and weather, to cut stockouts and overstock.
• Personalisation of products and offers from browsing and purchase history.
• Customer service, with AI answering routine questions and drafting replies for staff to approve.
• Catalogue content, generating and cleaning product descriptions at scale.

McKinsey found respondents in consumer goods and retail most often use AI agents in marketing and sales. [1] Score data readiness first: personalisation and forecasting fail quietly on inconsistent product and customer data. Quantana’s Windo case study is an AI-native commerce app that lets micro-retailers run a store from a phone. In hospitality, the Greenpark Hotels AI email agent sorts email into bookings, escalations, FAQs and priority queries.

Healthcare and life sciences

AI in healthcare splits into two groups. Administrative use (scheduling, referrals, correspondence, documentation support, coding, case processing) is lower risk and usually the first wave. Clinical use (diagnostic support, imaging, deterioration prediction, treatment recommendations) carries regulatory obligations from the start.

In Australia, the Therapeutic Goods Administration regulates software, including AI and large language models, as a medical device when its intended purpose includes diagnosis, prevention, monitoring, prediction, prognosis or treatment of disease. [14] It warns about “scope creep”: a digital scribe that summarises consultations may not be a device, but adding diagnosis suggestions can make it one. [14] Check intended purpose in phase 3 for every healthcare use case, and treat health information as sensitive when choosing where it runs. Quantana’s CasX.ai case study is an autonomous pharmacovigilance platform with 100 per cent audit traceability, more than 50,000 case records processed, and a reported 50 to 80 per cent reduction in pharmacovigilance cost across deployments.

Mining, industrial and remote operations

Data sits on site, connectivity can be poor, and compliance dominates. Good first candidates are compliance tracking, maintenance knowledge search and document processing, deployed privately or offline where needed. Quantana’s PX4 Software case study is a mining tenement compliance platform supporting 60+ regulatory jurisdictions across six continents.

How is an AI roadmap different from a digital transformation strategy?

A digital transformation strategy moves work onto digital platforms: cloud systems, online channels, connected data, digital marketing. An AI roadmap assumes much of that exists and asks which decisions and tasks AI should now perform or support, and how to govern it.

The two depend on each other. Digital transformation produces the clean data and modern systems AI needs; Gartner’s warning about AI projects without AI-ready data is really a warning about unfinished digital transformation. [5] And AI changes digital priorities, because the payoff comes from redesigning workflows, not digitising old ones. [1][6]

In practice, a low data score means your AI roadmap starts with digital transformation work: system consolidation, data ownership, governed access. A low process score means redesign first: map the workflow, decide where people stay in control, then add AI. Growth work such as search, content and lead generation is simply more use cases, scored on the same matrix.

How long does an AI roadmap take, and what does it cost?

Building the roadmap in-house usually takes two to six weeks (phases 1 to 3); with AI Compass we deliver it in 14 days. The first wave typically reaches production three to six months later, and the roadmap is refreshed quarterly. Budget in three parts: the roadmap work (mostly people time), foundations (data and governance fixes, often the largest item) and running costs. For running costs, see our cost guide and the LLM price index.