Frame and Size an Opportunity with Auditable Math

Workflow area: Strategy & Roadmap · Tier: Foundation · Hands-on time: ~30 minutes

Models are genuinely good at the framing half: turning a themes doc into opportunity trees and crisp jobs-to-be-done statements. They are genuinely dangerous at the sizing half: they will produce a confident "$4.2B TAM" assembled from misremembered analyst reports and silent multiplication errors. The fix is structural — force every number to carry a source and every calculation to be shown step by step.

When to use — and when not to

• Use it after synthesis, before strategywhen pb-02 gave you evidenced themes and you need to turn them into framed, sized opportunities for a roadmap discussion.
• Use it to stress-test a framing you already haveasking for three alternative problem framings is cheap and regularly reveals that you framed around your solution.
• Use it for order-of-magnitude sizingis this a ₹10 crore or ₹1,000 crore opportunity? AI-assisted bottom-up math answers that fast enough to compare ten opportunities.
• Skip it for board-grade market sizinga fundraise or a build-vs-buy decision needs sourced analyst data and a human analyst; AI-recalled market figures are not citable.
• Skip it when you have no evidence inopportunity trees generated from a product name alone are product-management theater: impressive structure, invented content.

The recipe

1. 1.

   Feed the evidence, ask for framings. Paste your synthesis doc (pb-02's output) into ChatGPT or Claude:

   Copy

   Below is a synthesized research findings doc for [PRODUCT, ONE LINE].
   
   1. Draft an opportunity tree: desired outcome at the root, then
      opportunities (user needs/pains, not features), each linked to
      the theme(s) that evidence it by theme number.
   2. For the top 3 opportunities, write a JTBD statement:
      "When [situation], I want [motivation], so I can [outcome]."
   3. For each of the 3, propose one alternative framing of the same
      underlying problem that we might be missing.
   
   Rules: no opportunity without a cited theme. Mark any branch you
   inferred rather than found in the evidence with [INFERRED].
   
   Findings doc:
   [PASTE SYNTHESIS DOC]
   

   Prune the tree yourself: kill [INFERRED] branches you can't defend, and check that no "opportunity" is secretly a feature request wearing a JTBD costume.

2. 2.

   Size bottom-up, with every step shown. For each surviving opportunity:

   Copy

   Help me size this opportunity bottom-up. Show every calculation as
   an explicit step — formula first, then numbers, then result. Never
   skip or combine steps.
   
   Opportunity: [ONE-LINE JTBD]
   My inputs (use ONLY these; where I give a range, compute low/base/
   high):
   - Target segment: [e.g. "Indian mid-market logistics firms,
     50-500 employees"]
   - Segment count: [NUMBER + SOURCE, e.g. "~9,000 — govt registry"]
   - % with this problem: [e.g. "30-50%, from our survey n=140"]
   - Reachable share in 3 yrs: [e.g. "5-10%"]
   - Price point: [e.g. "₹3L/yr ACV, current median"]
   
   Then: list which input the result is most sensitive to, and state
   the 2 weakest assumptions. Do not add market data from memory —
   if an input is missing, ask me for it.
   

3. 3.

   Audit the arithmetic by hand. Good output looks like:

   Copy

   Step 3: Serviceable = firms with problem × reachable share
   Base: 3,600 × 7.5% = 270 firms
   Step 4: Revenue = 270 × ₹3,00,000 = ₹8.1 crore/yr
   Most sensitive input: % with problem (survey n=140, single region)
   

   Recompute every multiplication yourself — calculator, sixty seconds. LLMs transpose digits and drop zeros precisely when the numbers look most fluent, and a 10x error in step 3 propagates silently to the headline.

4. 4.

   Verify every input's source. Each number should trace to a survey, a registry, your billing data, or an analyst report you actually opened. Any figure the model volunteered from memory gets deleted or replaced with a sourced one — this is the sizing equivalent of pb-02's citation audit.

5. 5.

   Write the one-pager yourself. Framing, base-case size with the low/high range, the two weakest assumptions, and what evidence would change your mind. The assumptions section is the part execs should argue with — never let the model bury it.

Make it yours

• Google Sheets as the audit trail: have the model output the sizing as a formula table (input cells, calculation cells, source column), paste into Sheets, and let the spreadsheet do the arithmetic — the model structures the model, Sheets computes it, and every review meeting can poke the cells.
• Claude for the adversarial pass: paste the finished one-pager into a fresh conversation and ask it to attack the sizing as a skeptical CFO — double counting, TAM/SAM confusion, price-point optimism. Cheap rehearsal for the real review.
• Team scale: standardize the input list (segment source, problem rate, reachable share, price) as a template so every PM's sizing exposes the same assumptions in the same order — comparable opportunity one-pagers are what make prioritization meetings short.

Quality check

• You recomputed the mathevery multiplication verified by hand or in Sheets; never forward AI arithmetic unchecked.
• Every input has a named sourceno number in the doc exists because the model remembered it.
• The range surviveslow/base/high made it into the one-pager; a single-point TAM is a tell that uncertainty got laundered out.
• Opportunities trace to evidenceeach tree branch cites a theme from real research; no hallucinated users, no invented pains.
• The weakest assumption is stated in boldif a reader can't find what would break the sizing, the doc is advocacy, not analysis.