Portfolio thinking

With costs and impacts estimated, the amateur move is ranking by ROI and funding the top of the list. Ranked lists over-fund the safe and starve the strategic; they also ignore that early initiatives change the economics of later ones. Portfolio logic fixes both. Three buckets, deliberate proportions:

• Quick wins (~60–70% of initiatives, ~40% of spend) — high-confidence, 90-day payback shapes: proven capability, green data, tolerant error profile, eager owner. Alder: exception triage, tender extraction, contract summarizer. Quick wins aren't the strategy — they're the trust engine that funds the strategy: visible relief, referenceable owners, momentum the change plan (Module 5) will spend.
• Strategic bets (1–2 at a time, ~40–50% of spend) — initiatives that move a line the board tracks, on quarters-long timelines with real uncertainty. Alder: predictive ETA for top accounts (attacks churn, the #2 board worry). Bets get program management, executive sponsorship — and pre-agreed checkpoints, because bets are where sunk-cost thinking does its expensive work.
• Options (~10% of spend, deliberately cheap) — small probes into things too early to bet on: a two-week eval of a promising vendor category, a red-flagged inventory item's unblock experiment, one team's structured trial of an emerging capability. Options buy information, and their success metric is a decision ('pursue / park / drop'), not a benefit.

Sequence for compounding, not just returns

Initiatives share substrate: the document pipeline built for tender extraction serves the contract summarizer; the eval-and-baseline habits from the first quick win de-risk the ETA bet; the data cleanup for ETA unlocks the detention predictor next year. When two candidates score similarly, fund the one that leaves more behind. A portfolio sequenced this way gets cheaper per initiative every quarter — which is the actual, defensible meaning of 'AI flywheel', a phrase you should otherwise ban from the roadmap.

Kill criteria: written at funding, or never

• Every initiative gets a review date and a kill/continue condition written at funding time, when heads are cool: 'triage: ≥60% auto-routed correctly at week 6, else re-scope'; 'ETA: beats dispatcher guess on 30% of lanes by Q2 checkpoint, else stop'.
• Killing an initiative that met its kill criteria is a process success — say so publicly, redeploy the team gracefully, and watch the quality of future proposals rise (people propose honestly when failure is survivable).
• The portfolio review (quarterly, one hour, the artifact list from this module) is where zombie initiatives go to die. Programs without kill discipline accumulate walking-dead pilots that consume the change capacity Module 5 needs — the most expensive line item no model shows.