I’ve read more AI strategy documents than I care to admit. Most of them look the same. There’s a section on the “AI opportunity,” a list of use cases pulled from a consulting framework, a capability maturity model, a proposed org structure, and a multi-year roadmap with neat phases.
They look great in PowerPoint. They almost never survive contact with reality.
The failure rate for enterprise AI strategies is staggeringly high. Gartner and others have put numbers on it, ranging from 60% to 85% depending on how you define failure. But I think the number is even higher if you’re honest about whether the strategy actually changed anything versus just creating busywork.
The Three Ways AI Strategies Die
Death by use case list. The most common failure mode. A team brainstorms 50 potential AI use cases, scores them on a 2x2 matrix of “feasibility vs. impact,” picks the top 10, and declares that the strategy. The problem: picking use cases isn’t a strategy. It’s a shopping list. There’s no coherent thesis about where AI creates competitive advantage. No honest assessment of what the organisation can actually execute. No prioritisation beyond gut-feel scoring.
I’ve seen organisations with 30 active AI proofs of concept and zero in production. They didn’t have a strategy. They had a portfolio of science experiments.
Death by infrastructure project. This is the opposite failure mode. The team decides that before any AI can happen, they need a “modern data platform,” a feature store, an ML ops pipeline, a model registry, and a data catalogue. Eighteen months and several million dollars later, the infrastructure is partially built and nobody has shipped an AI product to a customer.
Infrastructure is necessary. But building infrastructure without a clear first use case is like building a highway without knowing where it goes. You need enough infrastructure to support your first two or three production systems. Build what you need, when you need it.
Death by org chart. Some companies think the AI strategy problem is really an organisational design problem. Should AI be centralized? Federated? A center of excellence? Embedded in business units? They spend six months debating structure and hiring a Chief AI Officer before they’ve shipped anything.
Org structure matters, but it matters less than people think. I’ve seen centralized teams ship great AI products and I’ve seen federated teams do the same. The structure that works is the one that matches your company’s culture and decision-making style, not the one the consulting firm recommends.
What a Real AI Strategy Looks Like
A useful AI strategy answers four questions. That’s it. If your strategy document doesn’t answer these clearly, everything else in it is decoration.
Where does AI create defensible value for us specifically? Not “AI can improve efficiency” which is true for every company. What specific problems do you have, that AI can solve, in a way that builds competitive advantage? At Cochlear, the answer to this question is very different from what it was at Westpac, because the businesses are different, the data assets are different, and the competitive dynamics are different.
What can we actually execute in the next 12 months? Not what’s theoretically possible. What can your team, with its current skills, data, and infrastructure, realistically ship to production? Be brutally honest here. If you’ve never put an ML model in production, your 12-month plan shouldn’t include deploying six.
What do we need to build to execute? This is where infrastructure, team, and skills gaps get addressed, but in service of specific goals, not as abstract capability building. You need a feature store because your first three use cases require real-time feature serving. Not because “modern AI organisations have feature stores.”
How will we know it’s working? Every AI initiative should have a clear success metric that connects to business outcomes. Not model accuracy. Business outcomes. Revenue increased, cost reduced, customer satisfaction improved, time saved. If you can’t draw a line from the AI system to a business metric, the initiative doesn’t belong in your strategy.
The First 90 Days
If you’re building an AI strategy from scratch, here’s what I’d actually do in the first 90 days.
Spend the first month listening. Talk to 20 people across the business. Not just leaders. Talk to the people who do the work. Ask them what’s painful, what’s slow, what’s manual, what breaks. You’ll find better AI use cases in these conversations than in any brainstorming workshop.
Pick one or two problems from those conversations. Not the biggest or most exciting ones. Pick problems that are clearly painful, where you have good data, and where the solution is technically straightforward. Your first AI win needs to be a clear win, not a stretch goal.
Build and ship something in the second and third months. Even if it’s simple. Even if it’s not using the latest model architecture. Get something into production that real people use and that measurably improves their work. This accomplishes something no strategy document can: it proves you can deliver.
Everything after that builds on the credibility you’ve earned by actually delivering something. The strategy evolves based on what you learn from shipping, not from what looked good in a slide deck.
The Honest Conversation With Leadership
Most AI strategy failures can be traced back to misaligned expectations with senior leadership. The board saw a demo of ChatGPT and now they want an AI strategy. The CEO read an article about how AI is transforming healthcare and wants to know what you’re doing about it.
You need to have an honest conversation early. Here’s roughly what I say:
“AI will create real value for this business. It will also take longer and cost more than you expect. The first year is about proving we can ship reliable AI products and building the muscle to do it repeatedly. The big, transformative stuff comes in year two and three, once we’ve built the team, infrastructure, and organisational trust to execute at scale.”
This conversation is uncomfortable. But it’s far less uncomfortable than explaining in month 18 why none of the 15 use cases from the original strategy have shipped.
Stop Writing Strategy Documents
The best AI strategies I’ve seen aren’t documents. They’re living systems: a clear thesis about where AI creates value, a prioritised backlog of problems to solve, a team that ships regularly, and a feedback loop that adjusts priorities based on what’s working.
Write the strategy on one page. Spend the rest of your time building things. Adjust the strategy every quarter based on what you learn. That’s it.
The companies winning with AI aren’t the ones with the best strategy decks. They’re the ones that shipped something six months ago, learned from it, and shipped something better.