Size the Opportunity with Benefit Math Shown Step-by-Step

Workflow area: Data & Evidence · Tier: Foundation · Hands-on time: ~25 minutes

Opportunity sizing dies two deaths in most organizations: the analyst who will not put a number on anything, and the deck that claims "$2.3M annual savings" from arithmetic no one can inspect. Both lose in front of finance. The credible middle is a model where the assumptions are explicit and separately attackable, the math from assumption to conclusion is visible step by step, and uncertainty is carried honestly as a low/base/high range. AI is a strong drafting partner for exactly this structure — it is tireless about showing work when told to — but language models are unreliable calculators, so every number gets recomputed in a spreadsheet before it travels.

When to use — and when not to

• Use it to rank problems before solutioninga rough sizing across five pain points tells you which one deserves the project, and pb-06's confirmed waste register is ideal input.
• Use it for the benefits section of a business casethis playbook produces the benefit half; cost stacks and AI-specific cases are pb-38's territory.
• Use it when a stakeholder claim needs testing"we waste half our week on this" converted to modeled hours either substantiates the pain or right-sizes it; both outcomes are useful.
• Skip it when a real measurement is cheapif the warehouse can answer the question directly (pb-07), measure instead of modeling; models are for what you cannot yet measure.
• Skip point-precision theaterif the honest range spans an order of magnitude, present the order of magnitude and say what would narrow it; do not average your way to false confidence.

The recipe

1. 1.

   Gather the raw inputs and grade them. Volumes (transactions/month), time per unit, error/rework rates, fully loaded labor rates, and any hard costs. Grade each: MEASURED (from data, cite it), REPORTED (a stakeholder said it, cite who), or ASSUMED (your estimate). The grade travels with the number all the way to the final slide.

2. 2.

   Build the model with the assumptions ledger.

   Copy

   You are a business analyst building an opportunity-sizing model.
   Show all work. Do not round until the final step.
   
   Pain point: [ONE LINE — e.g. "duplicate vendor records force
   manual reconciliation at month-end close"]
   Inputs (with evidence grade):
   - [e.g. "340 duplicate vendors found last quarter — MEASURED,
     ERP dedup report Apr 2026"]
   - [e.g. "45 min average to reconcile one duplicate — REPORTED,
     AP team lead interview 12 May"]
   - [e.g. "fully loaded AP analyst cost $52/hr — ASSUMED, from
     salary band midpoint + 1.4 overhead factor"]
   
   Produce:
   1. ASSUMPTIONS LEDGER — every input, its grade, its source, and
      a LOW / BASE / HIGH value with one line on why that range.
   2. CALCULATION — step by step, one operation per line, showing
      units at every step, for LOW, BASE, and HIGH scenarios.
   3. RESULT — annualized range, stated as a range.
   4. SENSITIVITY — which single assumption moves the result most,
      and what evidence would narrow it.
   5. NOT INCLUDED — benefit types you deliberately left out
      (error-cost avoidance, opportunity cost) and why.
   
   Rules: never introduce a number not in my inputs without marking
   it ASSUMED with a stated basis. Keep units visible — hours vs
   days errors are how sizings die. No point estimates.
   

3. 3.

   Read a sample of good output.

   Copy

   CALCULATION (BASE)
   340 duplicates/quarter × 4 quarters = 1,360 duplicates/yr
   1,360 × 45 min = 61,200 min/yr = 1,020 hrs/yr
   1,020 hrs × $52/hr = $53,040/yr
   
   RESULT: $38k – $71k/yr (LOW–HIGH), base $53k
   SENSITIVITY: minutes-per-duplicate dominates (REPORTED, single
   source). A one-week timing sample of 20 reconciliations would
   narrow the range by ~60%.
   

4. 4.

   Recompute every line in a spreadsheet. This is non-negotiable: rebuild the calculation in Excel or Sheets with the ledger as an inputs block and formulas for the rest. You are checking the model's arithmetic (LLMs drop factors and mangle unit conversions) and creating the artifact finance will actually want — a sheet where they can change an assumption and watch the range move. If the spreadsheet and the AI output disagree, the spreadsheet wins and the discrepancy gets understood before proceeding.

5. 5.

   Attack it before someone else does. Fresh conversation: "Here is an opportunity sizing. As a skeptical CFO, find the weakest assumption, any double-counting, any benefit that would not actually convert to cash or capacity, and the question you would ask first." The classic catch: time savings that free up seventeen minutes per person per day convert to real money only if the capacity is redeployed — say explicitly whether you claim cost reduction or capacity release.

6. 6.

   Package with the sensitivity, not just the answer. The one-slide version: the range, the three biggest assumptions with grades, the sensitivity line ("the answer is mostly the minutes-per-duplicate estimate"), and the cheap next measurement that would firm it up. Offering the measurement is what separates analysis from advocacy — and stakeholders fund small measurements far more readily than large claims.

Make it yours

• Claude/ChatGPT for the drafting, always a spreadsheet for the record: the division of labor is fixed — AI structures the model and drafts ranges; the spreadsheet computes and becomes the source of truth.
• Ask for the spreadsheet directly: have the model output the assumptions ledger and formulas as a table you can paste into Sheets, with the formula text in a column — faster than transcribing, and it keeps the show-your-work property.
• Portfolio scale: use the identical ledger/calculation/sensitivity structure for every sized opportunity, and a roll-up sheet comparing base cases with evidence grades becomes leadership's prioritization view.
• Pair with pb-07: every ASSUMED or REPORTED input that could be MEASURED from the warehouse is a text-to-SQL question — upgrading two grades often narrows the range more than any modeling sophistication.

Quality check

• Every number traces to the ledgernothing appears in the calculation that is not an input with a grade and source; models smuggle in "typical" figures when allowed.
• The arithmetic was independently recomputedspreadsheet and AI output agree, or the discrepancy is resolved; unaudited LLM math never ships.
• Units survived every stephours/days, monthly/annual, per-unit/total conversions each visible and checked; unit slips are the most common fatal error.
• The range is honestLOW is genuinely conservative, and no point estimate appears anywhere a range should be.
• Reported inputs went back to their sourcesthe AP lead has seen that their "45 minutes" now anchors a $53k claim, and stands by it.
• Capacity-versus-cash is explicitthe model says which one it claims, and NOT INCLUDED names what was left out on purpose.