Fund workflows, not pilots: portfolio discipline beats enthusiasm
AI strategy, ROI, and use-case portfolio management
The question: Which AI projects should a company fund, stop, scale, or redesign?
Executives need a disciplined portfolio method to escape pilot theatre. MIT's GenAI Divide finding — ~95% of enterprise GenAI pilots showing no measurable P&L return — is not an AI failure; it is a portfolio-management failure.
What the lesson covers
ROI arithmetic, honestly: benefits = labour time × utilisation of freed time + quality uplift (error/rework reduction) + speed (cycle-time value) + revenue conversion + risk reduction + strategic option value. Costs = licences + compute/tokens + integration + data preparation + security/governance + evaluation + training + change management + MAINTENANCE (the chronically forgotten line). Most "AI ROI" decks count licence costs against gross time saved and call it a business case.
The adoption-lag reality: benefits follow the J-curve (Lesson 1) because complements come first. Time saved is not money saved until the freed time is redeployed — measure utilisation of freed capacity, or your ROI is fiction. High performers pair every use case with a workflow-level metric and a manager who owns the number.
Portfolio method: score use cases on value potential × implementation readiness (data, workflow, sponsor) × risk × learning value. Balance three buckets: quick wins (fund fast, harvest, publicise), capability bets (build the platform/data muscles), and strategic differentiators (where you must be better than competitors). Kill criteria are written BEFORE funding: the metric, the threshold, the date.
Why pilots die: no owner, no baseline measured, no workflow redesign (tool dropped on old process), no adoption plan, and no kill/scale decision date — zombie pilots consume attention indefinitely. The stop decision is a portfolio skill: stopping five zombie pilots funds one scaled workflow.
The generic-Copilot lesson: broad tool rollouts produce usage statistics, not P&L impact, unless specific workflows are redesigned around them and managers change what they expect from the freed time. Usage is an input metric. The scoreboard is cycle time, quality, cost-to-serve, revenue per FTE.
Key points
- Count ALL costs (integration, data, evaluation, maintenance, change) against NET redeployed benefit — not licences against gross time saved.
- Time saved ≠ money saved until freed capacity is redeployed; measure utilisation.
- Portfolio buckets: quick wins / capability bets / differentiators — with written kill criteria and dates.
- Zombie pilots are a management failure: no owner, no baseline, no decision date.
- Usage is an input metric; the scoreboard is cycle time, quality, cost-to-serve, revenue.
Framework — AI Portfolio Matrix
value potential × implementation readiness × risk × learning value → three buckets: quick wins / capability bets / differentiators. Every funded case gets: baseline, workflow metric, owner, adoption plan, kill criteria with a date.
The lab
Build a defensible ROI model and prioritise a portfolio — then try to kill your own case.
Open this lesson, its lab and its quiz
Sources and further reading
- The GenAI Divide — State of AI in Business — MIT Project NANDA
- The State of AI — McKinsey
- The Productivity J-Curve — Brynjolfsson, Rock & Syverson