A good AI strategy looks plausible in a presentation. A great AI strategy still makes sense in month four, when the data isn't where you expected it, the integration is more complex than scoped, and the internal champion who drove the initiative is on parental leave.

Most AI roadmaps fail not because the ideas were wrong but because they were built on assumptions that didn't survive contact with the organisation. Here are the structural issues we see most often.

The dependency problem: building phase 3 before phase 1 works

AI roadmaps tend to be written with optimistic dependencies. Phase 1 delivers clean, structured data. Phase 2 uses that data to train a model. Phase 3 deploys the model at scale. The presentation shows three neat arrows flowing right.

The reality: phase 1 takes longer than planned because the data is messier than expected. Phase 2 can't start on schedule. Phase 3 gets pushed. By month six, the entire roadmap has shifted and stakeholder confidence has eroded.

Build your strategy so that each phase delivers standalone value that doesn't depend on the next phase completing on time. A data cleaning initiative that makes your reporting more reliable is valuable whether or not the AI model in phase 2 ever gets built. Design for resilience, not for optimism.

The ownership problem: nobody owns the outcome

AI projects have a tendency to be owned by IT or by a central innovation team, with the operational teams expected to use the output. This almost never works. The people who will use the system daily need to be invested in its success — and that means having real input into how it works, not just being trained on it after the fact.

Assign an operational owner for every AI initiative. Not a technical owner — someone in the business unit who cares whether the process works and who will flag problems early rather than routing around the system. This person needs protected time to be involved, not just a role title in the project charter.

The measurement problem: measuring activity instead of outcomes

AI strategy reports frequently track the wrong things. "We completed 12 AI projects this year" tells you about activity. "Our invoice processing time dropped from 120 hours to 11 hours per month" tells you about outcomes.

For every initiative in your roadmap, define the outcome metric before work begins. Not the output (a model is deployed, a system is live) but the outcome (a process that was taking X hours now takes Y hours, with a quality improvement from Z to W). Review against those metrics quarterly, not annually.

The scope problem: starting too big

The first AI project in an organisation is not a transformation. It is a proof that AI can work in your environment — with your data, your integration constraints, your team. Scope it accordingly.

A well-scoped first project takes 8–14 weeks, touches one bounded process, and produces a measurable result that you can show to the rest of the organisation. That result is your internal case study and your credibility for the next project. An overscoped first project that runs for 9 months and delivers ambiguous results kills the programme.

The technology problem: locking in too early

The AI landscape is changing fast enough that a technology choice made 18 months ago may not be the right choice today. Build your architecture so you can switch model providers without rebuilding the integration layer. Use abstraction layers between your application logic and the AI model. This is more work upfront and pays dividends every time a better option becomes available.

The best AI strategy is the one that makes something measurably better before the next board meeting.

What a resilient AI strategy looks like in practice

The strategies that work are built around a short list of specific, high-value processes rather than a broad transformation ambition. They have named owners in the business, not just in IT. They track outcomes and review them quarterly. They start small, prove results, and use those results to justify the next phase. They are boring to read and interesting to execute.