Where do we even start with AI strategy?

The wrong first question

Most AI strategy exercises open with "what could we build with AI?" It feels productive — you get a whiteboard full of ideas within the hour. It rarely is productive, because there's no shared answer to a more important question first: what are we actually trying to achieve as a business this year?

Skip that question and you get a backlog nobody can prioritise. Every idea sounds plausible. None of them can be ranked against each other, because they weren't generated against the same yardstick. Six months later you're still debating whether the chatbot or the forecasting tool should go first, and the honest answer is that neither was ever tied to a business outcome, so nobody can tell.

The right first question

AI strategy should answer three things before it goes anywhere near use cases:

1. What outcomes matter this year? Revenue growth, cost reduction, risk reduction, customer experience, employee productivity, regulatory position. Pick the two or three that are non-negotiable. If everything is a priority, nothing is.
2. How ready are we? Data, skills, governance, culture, and platform. Honest scoring here saves quarters of wasted delivery. A team with messy data and no clear owner for AI decisions cannot ship a customer-facing model safely in Q1, no matter how good the idea is.
3. What does success look like? Concrete measures, a timeline, and a named person accountable — not "the AI team" or "IT".

A worked example

Take a mid-sized insurer wanting to "use AI in claims." Asked cold for use cases, the team produces a list: auto-triage claims, chatbot for policyholders, fraud detection, document summarisation for adjusters. All sound reasonable. All compete for the same two data engineers.

Now run it through outcomes first. Leadership agrees the year's priorities are cutting claims cycle time and reducing fraud leakage — cost reduction and risk reduction, not customer experience. That single decision reorders the list: auto-triage and fraud detection move to the top because they map directly onto the chosen outcomes; the chatbot drops down, not because it's a bad idea, but because it serves an outcome nobody picked this year.

Then the readiness check bites. Fraud detection needs historical claims data tagged with confirmed fraud outcomes — and it turns out that tagging has been inconsistent for two years. That's not a reason to abandon the idea; it's a reason to sequence a three-month data remediation project ahead of any model build. Without the readiness step, that gap surfaces mid-project, after money and credibility are already spent.

Pitfalls to watch for

• Outcome-washing. Teams sometimes retrofit an outcome to a use case they already wanted to build, rather than deriving the use case from the outcome. If the outcome was written after the idea, it's decoration, not direction.
• Readiness scores that are too polite. Self-assessment tends to average out to "medium" on everything. Push for specifics: name the dataset, name the gap, name who owns fixing it.
• Accountability without a name. "The AI steering group will monitor this" is not accountability. One person's name against one measure, reviewed on a fixed date, is.
• Treating the canvas as a formality. The business model canvas step exists to show where value actually sits in the organisation — which customer segments, which cost lines, which channels. Skipping it means use cases get generated against a vague mental model of the business rather than the real one.

A simple sequencing you can steal

• Week 1: leadership alignment on outcomes and guardrails — what's in scope, what's off-limits (customer data, regulated decisions, whatever applies to you).
• Week 2: readiness assessment, scored honestly against data, skills, governance, culture, and platform.
• Week 3: business model canvas — where value actually lives, and where AI could touch it.
• Week 4: only now, use-case generation, scored against the outcomes chosen in week 1.

If you start at week 4, you'll spend the next six months rewriting your strategy every time the board changes its mind, because there was never a stable reference point to rewrite it against.

What to do next

Don't start by listing use cases. Start by getting three people in a room — whoever holds budget, whoever holds the data, and whoever holds the mandate to say no — and agree the two or three outcomes that matter this year. Write them down before anyone mentions a model, a chatbot, or a vendor. Everything else in this sequence depends on that list existing first.

Where Florence helps

Florence enforces this order structurally rather than relying on discipline. The Organisation profile and Business Model Canvas steps come before the AI Readiness Assessment, which comes before Insights and Strategy — you can't skip to use-case generation without passing through the earlier steps first. The assessment scores readiness across data, skills, governance, culture, and platform, and the Insights step turns those scores into what to fix before you build anything. By the time you reach Strategy, the outcomes, the readiness gaps, and the value map are already on record — so the use cases that come out the other end are ranked against something real, not against how loud each idea's champion is in the room.