Impact estimation
Five value levers, conservative math for each, and the baseline discipline that makes benefits provable later instead of arguable forever.
Costs are knowable; impact is where estimates go to inflate. The discipline: name the lever, apply the conservative math for that lever, and freeze the baseline before anything launches.
- Cost-out — time saved × volume × loaded rate × capture rate. The capture rate (typically 40–70%) is the honesty coefficient: saved minutes only become money when they compound into headcount avoided, overtime cut, or capacity redeployed to named work. 'We freed 900 hours' is not a benefit; 'CS absorbs 18% volume growth without the two planned hires' is.
- Revenue — win-rate, retention, or capacity effects. Apply the attribution haircut: AI is usually one contributor among several, so claim the slice a skeptic would grant (Alder's faster quotes: claim 1–2 points of win-rate, not five). Revenue claims draw CFO fire first — armor them heaviest.
- Risk — expected-loss reduction (frequency × severity × mitigation share) for quantifiable risks; for the rest, an honest qualitative line. Never let 'incalculable risk value' carry a business case alone — it's real, and it's also the last refuge of a case that lost its numbers.
- Speed — cycle-time compression: quote in 2 hours vs 2 days, exception answered in 10 minutes vs 4. Speed monetizes through another lever (win rate, retention, detention fees avoided) — trace it there rather than double-counting it beside them.
- Quality/consistency — error-rate deltas on high-volume work: fewer invoice disputes, fewer re-deliveries. Cleanly measurable when you have the before-rate, which is the segue that matters:
The baseline discipline (do this or forfeit the argument later)
For every funded initiative, measure the before-state before launch: today's minutes per exception, today's quote turnaround, today's dispute rate — written down, dated, agreed by the owner. Six months from now, 'is this working?' becomes a comparison instead of a debate. Skipping baselines is the single most common measurement failure in enterprise AI programs, it is unrecoverable after the fact, and it hands next year's budget discussion to whoever tells the best story. Baselines are boring; they are also how strategy owners keep their jobs.
Stress-test this business case as a hostile CFO: [paste the one-pager - lever, math, assumptions, baseline plan]. (1) Attack every assumption - which are load-bearing? (2) Where am I double-counting across levers? (3) What's the weakest number a board member would circle? (4) Rewrite my headline claim at the confidence level you'd actually sign. Do not be polite.
Run every case that goes in the portfolio through this. The rewritten headline in (4) is usually the one you should present — cases that survive their own red-team get funded and stay funded.
Pilots over-perform: motivated users, curated scope, novelty attention, the builder hovering nearby. Discount pilot results 30–50% when projecting to scale — and design pilots to predict scale (typical team, typical mess, four weeks minimum) rather than to impress. A pilot that can't fail teaches nothing and licenses over-investment; the kill criteria you'll write in the portfolio workshop exist precisely so pilots are allowed to fail cheaply.