AI strategy as a portfolio decision
Most AI programs run into difficulty at problem selection rather than model selection. The work is choosing which expensive, repetitive workflows deserve funding, and declining the rest.
The requirement this answers
“Define and drive the enterprise AI strategy and roadmap”
Start from a costly workflow rather than from a capability you want to use.
Where programs stall
Programs that begin with a specific, measurable cost tend to survive an early setback. Programs that begin with a general intention to adopt AI often cannot demonstrate what they were worth.
About 88% of organizations now use AI somewhere, and roughly a quarter of initiatives meet their expected return. In my experience the gap is rarely the model.
How I structure it
I run AI strategy as six repeating steps: problem selection, platform fit, direction (customer-facing or internal), governance and risk, adoption, and measurement. Each has a funding gate. A use case that cannot name its baseline metric before build does not get funded.
Platform fit deserves more attention than it usually gets. Generative, agentic, and classical machine learning solve different problems. Forecasting and demand planning are structured-data problems. Multi-step transactional work needs tool-calling and write access. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, largely over cost, unclear value, and inadequate risk controls.
Sequencing outward and inward work
Internal-facing AI is usually the better first investment. It is a lower-stakes environment for building organizational capability, and it is reversible. Customer-facing AI carries higher visibility, which raises both the scrutiny and the cost of underperformance.
I manage the two as separate tracks with separate measurement, because they have different risk profiles and different payback horizons.
How I apply it
- Inventory workflows by cost and repetition before discussing technology
- Assign each candidate a platform class (generative, agentic, or classical ML) and justify it
- Run outward-facing and inward-facing work as two tracks with separate measurement
- Gate funding on a small, fixed KPI set agreed before the first sprint
What good looks like
- Every funded use case has a named business owner and a measured pre-build baseline
- At least one use case has been stopped at a gate
- Run cost is budgeted alongside build cost
- Leadership can state the purpose of the AI portfolio in one sentence