AI Implementation Roadmap Template: A 90-Day Plan That Actually Ships

A roadmap is not a five-year strategy deck. A useful AI implementation roadmap fits on two pages, covers the next 90 days in detail and the next 12 months in outline, and is updated every quarter against what actually shipped. This template is the one we use at Waymouth Tech for Australian SMB and mid-market clients.

The shape of a good roadmap

Before the template itself, the principles:

• Pick fewer things. Two or three workflows in the first year, not ten.
• Sequence by leverage and learning, not excitement. Boring document processing usually beats glamorous customer-facing AI for project one.
• Build the operating muscle first. The first project's biggest deliverable is not the system — it is the team's ability to run AI in production.
• Plan around evaluation. Anything you cannot measure, you cannot improve.

The two-page template

Here is the structure. Copy it into a document and fill it in. If you cannot fill in any section in plain English, you are not ready to build yet.

Section 1: Business context (half a page)

• What does the business do, and what are the top three operating constraints right now?
• Which functions have the most repetitive, document-heavy work?
• What is the budget envelope for the next 12 months of AI work?
• Who is the executive sponsor, and who is the delivery owner?

Section 2: Prioritised use cases (half a page)

List 5–10 candidate workflows. For each, capture:

• Workflow name and one-sentence description.
• Volume per week and time per case today.
• Estimated value of automation (dollars per month or hours per week).
• Complexity (low / medium / high) — based on data access, integrations and risk.
• Priority rank.

Then pick the top three. The rest stay on the list for later quarters.

Section 3: 90-day plan (half a page)

Break the 90 days into three 30-day blocks:

• Days 1–30 — Discovery and design. Workflow mapping, data assessment, architecture decision, evaluation plan, vendor and tooling selection. Output: a one-page design doc and a fixed-scope pilot brief.
• Days 31–60 — Pilot build. Build the smallest version of the workflow that runs end-to-end. Internal testing with the actual team. Output: working software with at least one week of internal use.
• Days 61–90 — Pilot operation and decision. Real users, real cases, measured against the success number. Daily standups in week one, weekly thereafter. Output: a go / no-go decision with evidence.

Section 4: 12-month outline (quarter page)

A rough sequencing of the next three to four use cases, the team or partner you will use, and the expected investment per quarter. Treat this as a hypothesis, not a commitment.

Section 5: Risks, controls and measurement (quarter page)

• Data residency and retention approach (default to AU-region cloud).
• Alignment with the Voluntary AI Safety Standard.
• Human-in-the-loop policy: where does a person review or approve?
• Monitoring: what dashboards, what alerts, what cadence?
• Success metrics: the one or two numbers that prove value.

How to actually run the 90 days

The template is the easy part. Running it is where most plans collapse.

Days 1–30: discovery without paralysis

Most discovery phases drag because no one is willing to make decisions with incomplete information. Set a hard 30-day cap. By day 21, the design should be 80% locked. Days 22–30 are for pressure-testing it with the people who will use the system.

You are answering five questions in this phase:

1. What is the workflow, step by step?
2. What data does it need, and where does that data live?
3. What is the simplest architecture that solves it?
4. How will we evaluate model outputs?
5. Who is on the build team, and what is the fixed scope of the pilot?

Days 31–60: build the smallest useful thing

Resist scope creep. The point of the pilot is to learn, not to launch. We typically restrict pilots to:

• One workflow.
• One user group (usually 3–8 people).
• One success metric.
• The minimum integrations needed to be real (often just an inbox, a file store and the AI provider).

If a stakeholder asks for "while we're at it, can we also..." — write it down on the backlog. Do not add it to the pilot.

Days 61–90: prove it, then decide

Move it into the hands of real users with real cases. Watch what happens. The first week is rough. The second week is encouraging. By the fourth week, you usually have enough evidence to decide.

The decision is one of three things:

• Go to production. Move into a 60–120 day production hardening cycle. See from pilot to production AI deployment.
• Iterate. The idea is right, the execution needs another cycle. Re-scope and re-time-box.
• Kill it. The workflow is not AI-amenable, or the value is not there. Document what you learned and pick the next use case from the backlog.

For more on realistic time expectations across the whole journey, see AI implementation timeline: realistic expectations.

Why this matters in Australia

Australian SMBs we work with consistently underestimate two things in their roadmaps: data plumbing and change management.

Data plumbing. "We have the data in our system" usually means "the data exists in three different SaaS tools, two spreadsheets and one staff member's head." Budget meaningful time for data access, schema mapping and clean-up. It is often 30–50% of pilot effort.

Change management. Australian workplaces tend to be quietly sceptical of imposed change. The roadmap should include user training, internal champions, and feedback loops from week one. Build with the people who will use it, not around them.

Also factor in the local regulatory backdrop. The Privacy Act reforms continuing through 2026, sector-specific frameworks like APRA's CPS 230, and the Voluntary AI Safety Standard all influence what controls need to be in the roadmap from day one. It is much cheaper to bake these in than to retrofit them after a security review.

A note on tooling choices

Resist the temptation to make the roadmap a tooling document. The right model, framework or platform should fall out of the workflow design, not drive it. Most SMB roadmaps eventually settle on:

• A frontier model API (Anthropic Claude, OpenAI GPT, or Google Gemini) via an AU-region endpoint where possible.
• A workflow or agent platform for the orchestration.
• An evaluation tool to keep the outputs honest.
• Your existing identity, storage and observability stack.

That is enough for the first three or four projects. Add complexity only when a specific workflow demands it.

What to do next

Open a blank document. Fill in the two-page template above. If you cannot fill in any section in plain English, that is your next piece of homework. If you can, you have a useful AI implementation roadmap — and you are well ahead of most businesses we meet.

For broader context, start at AI implementation consulting Melbourne.