Your AI Roadmap Lists Projects. It Never Lists Capacity.

Every Australian enterprise AI roadmap I see has the same shape. A dozen coloured boxes. A timeline that stretches to next financial year. A quiet assumption that "the platform team will make it work." The board nods. Funding lands. Then delivery stalls because nobody counted the scarce things: seats, eval hours, data contracts, change load, and people who can own an incident at 2am.

Projects are easy to list. Capacity is what decides whether any of them ship. If your roadmap does not name capacity, it is a wish list dressed as strategy.

This piece is a practical rewrite. Keep the ambition. Add the constraints that actually govern throughput. Use it before the next pilot slate gets rubber stamped.

What a project-only roadmap hides

A project slide answers "what do we want to try?" It rarely answers:

• How many concurrent GenAI workloads can the approved stack carry without burning quotas?
• Who has spare hours to design evals, red-team prompts, and keep regression suites honest?
• Which source systems have owners willing to hold a data contract for retrieval or features?
• How much process change can frontline teams absorb this quarter without another "AI fatigue" revolt?
• Who is on the rota when an agent tool calls the wrong system of record?

Those questions feel operational. They are strategic. Throughput, risk, and credibility sit there. A roadmap that skips them trains executives to expect miracles and then blame "execution" when the miracles do not arrive.

Capacity is not one number

Treat capacity as five coupled budgets. If any one is empty, the project list is fiction.

1. Platform capacity

Licences, model quotas, vector store size, gateway rate limits, VPC endpoints, and logging volume. Count concurrent workloads, not logos on a slide. If three pilots and one production agent share one tenant with no quotas, you do not have a platform. You have a queue with vibes.

2. Evaluation capacity

Someone has to write golden sets, score retrieval, catch prompt regressions, and decide go or no-go. That work does not appear in most business cases. When it is missing, demos look sharp and production guesses. Budget person-weeks for eval the same way you budget engineering.

3. Data contract capacity

RAG and features need owners who will freeze schemas, publish freshness SLAs, and take a pager when a join breaks. If every pilot invents its own extract, you are buying technical debt with marketing budget. Name the contract owners before you name the chatbot.

4. Change and adoption capacity

Frontline teams can only absorb so many new tools, prompts, and "just use Copilot" mandates. Stacking roll-outs without measuring training load and support tickets is how shadow AI returns through the side door. Pace beats pile-on.

5. Operating capacity

Production AI needs an incident path: who can disable a tool, roll back a prompt, or cut a model when costs spike. If that rota does not exist, your roadmap ends at pilot theatre. Put names next to on-call and kill switches.

Rewrite the roadmap in one page

For each proposed initiative, force a capacity row before a funding row:

• Outcome: the decision or workflow that improves, not the model brand
• Platform draw: estimated tokens, seats, storage, and shared services touched
• Eval plan: owner, golden set size, pass bar, cadence
• Data contracts: systems, owners, freshness and PII class
• Change load: teams affected, training hours, support channel
• Ops: kill switch, on-call, rollback path
• Kill criteria: what evidence stops the project without drama

If a row is blank, the initiative is not ready for the roadmap. It belongs in discovery, not in the funded slate.

A ninety-day capacity discipline

You do not need a new PMO. You need a habit.

1. Weeks 1 to 2: Inventory live and queued AI work. Map each item to the five budgets. Publish the heat map to the steering group without soft language.
2. Weeks 3 to 4: Cap concurrent pilots. One in, one out, until platform and eval budgets show spare headroom.
3. Weeks 5 to 8: Stand up a shared eval lane and a short list of data contracts with named owners. No new production path without both.
4. Weeks 9 to 12: Rebuild the roadmap as a capacity-constrained portfolio. Fund fewer initiatives with clearer kill criteria. Report throughput (shipped and stopped) not slide count.

Boards respond to honesty. A shorter roadmap with named constraints beats a colourful wall of pilots that never leave staging.

What good looks like

A healthy AI strategy page reads like an operating plan:

• Fewer initiatives, each with a capacity claim someone will defend
• Shared platform and eval services treated as first-class products
• Explicit trade-offs ("we pause Copilot expansion to free eval hours for claims RAG")
• A standing review that can stop work without career damage

That is strategy. The project grid is just inventory.

Closing

If your AI roadmap only lists projects, you have described ambition. Add capacity and you have described a plan that can survive contact with Australian delivery reality: scarce specialists, regulated data, and teams already tired of tool sprawl.

Before the next steering pack goes out, ask one question per box on the slide: what capacity does this consume, and who owns that budget? If nobody answers, leave the box off the roadmap until someone does.