Practical guide

How to create a market-sizing model with AI

A useful market-sizing model is a decision model, not a large number with a citation. It defines the customer unit, geography, period, eligibility rules, reachable segment, price or economic value, and uncertainty before doing arithmetic. AI can help classify source tables, draft formulas, and challenge scenarios, but the analyst remains responsible for source selection, exclusions,.

By Two Prune

A useful market-sizing model is a decision model, not a large number with a citation. It defines the customer unit, geography, period, eligibility rules, reachable segment, price or economic value, and uncertainty before doing arithmetic. AI can help classify source tables, draft formulas, and challenge scenarios, but the analyst remains responsible for source selection, exclusions, calculations, and the final interpretation.

Build the model with BOUND

Use five linked steps: Boundary, Observe, Unify, Normalize, Document. Keep a source-and-assumption register beside the calculations so a reviewer can reproduce each transition.

### B — Boundary the decision

Start with the decision the model will support. “How large is the market?” is too vague. A useful question is closer to:

> What is the annual contract-value opportunity among U.S. private-sector establishments in three named industries, with 50–249 employees, that meet our workflow and procurement criteria during the next planning year?

Record:

Do not mix establishments and companies. One company can operate several establishments, while a dataset may count jobs rather than people or employers. A model that changes unit halfway through can be numerically neat and conceptually wrong.

### O — Observe the source universe

Prefer a source whose population and definitions fit the boundary. The U.S. Census Bureau’s County Business Patterns methodology explains that CBP reports establishment data by geography, industry, and employment size, while excluding or treating some activities differently. It also documents disclosure protection, noise infusion, suppressed cells, and known coverage limitations. These features affect what a model may infer from a table.

The U.S. Bureau of Labor Statistics’ Quarterly Census of Employment and Wages overview describes QCEW establishment, employment, and wage measures, their unemployment-insurance administrative sources, geographic and industry detail, and exclusions. QCEW covers more than 95% of U.S. jobs, but that coverage statement does not mean it counts every potential customer or business in a proposed market.

For every source, retain the publisher, dataset, table or endpoint, release and reference dates, unit, population, classification, filters, revision status, suppression or imputation flags, and extraction date. A webpage link alone is not enough lineage.

### U — Unify classifications and denominators

Create a crosswalk before combining sources. It should show each required concept, the field used, the transformation, and the unresolved mismatch.

| Required concept | Source field | Transformation | Limitation | |---|---|---|---| | target establishment | CBP establishment count | retain selected NAICS and size bands | excludes nonemployers and most government activity | | workforce scale | QCEW employment | aggregate the same industry and geography | jobs are not unique workers or buying entities | | eligible workflow | first-party discovery sample | apply coded inclusion rule | sample may not represent the population | | annual value | approved price architecture | convert monthly to annual contract value | price and uptake are assumptions, not observed demand |

Do not fill a suppressed cell with an AI estimate and then present it as published data. Either use an authorized broader aggregation, obtain another appropriate source, or carry the gap as an interval. When classification years differ, document the crosswalk and run a sensitivity case.

### N — Normalize the calculation

Separate sourced counts from assumptions. A simple structure is:

text TAM units = in-scope population units SAM units = TAM units × eligibility rate reachable units = SAM units × reachable-share assumption annual opportunity = reachable units × annual value per unit

This is not a universal formula. Seat-based, usage-based, transaction-based, and outcome-linked offerings need different denominators. Avoid stacking percentages from unrelated studies. A survey response about interest is not automatically an eligibility, procurement, or adoption rate.

Use AI for bounded assistance: map requested concepts to candidate fields, draft auditable formulas, generate edge cases, and compare scenario tables. Require a person to approve the data source, classification, formula, and assumption. Recalculate key cells independently and test that changing one assumption affects only the intended outputs.

GAO’s Assessing Data Reliability frames reliability work around the intended use of the data. Apply that principle proportionately: a directional exploration may tolerate a wider interval than an investment or capacity decision, but the model should say which use it can support.

### D — Document the decision range

Publish a range and its drivers, not false precision. The model record should include:

Use clear labels: observed, derived, assumed, and not evidenced. A market model estimates an addressable opportunity under stated rules. It does not demonstrate demand, conversion, willingness to pay, competitive win rate, or future revenue.

  • customer unit: establishment, enterprise, team, seat, transaction, or another unit;
  • included industries and the classification version;
  • geography and reference period;
  • size or operating characteristics;
  • buyer, user, and use case;
  • product scope and price basis;
  • exclusions and why they matter;
  • decision horizon and currency.
  • base, conservative, and expansion scenarios;
  • source and assumption owners;
  • formula version and independent check;
  • missing and suppressed data treatment;
  • sensitivity ranking;
  • double-counting tests;
  • what evidence would replace each assumption;
  • review date and change triggers;
  • decisions the model must not support.

Sources: [1] [2] [3]

Worked example: a synthetic operations-software segment

An analyst models a fictional workflow product for a defined U.S. establishment segment. The extracted public table contains 18,400 in-scope establishments. A documented first-party review suggests that 22% meet the workflow eligibility rule. The planning team assumes 65% could be reached through its current channels and uses an illustrative annual contract value of $7,200.

text TAM units = 18,400 establishments SAM units = 18,400 × 22% = 4,048 Reachable units = 4,048 × 65% = 2,631 (rounded) Illustrative annual opportunity = 2,631 × $7,200 = about $18.9 million

The output is not reported as “the market is $18.9 million.” The review sheet says 18,400 is sourced, while eligibility, reachable share, and price are assumptions. It also shows:

| Scenario | Eligibility | Reachable share | Annual value | Result | |---|---:|---:|---:|---:| | conservative | 15% | 50% | $6,000 | $8.28m | | base | 22% | 65% | $7,200 | $18.94m | | expansion | 28% | 75% | $8,400 | $32.46m |

The largest decision sensitivity is eligibility, not arithmetic. The next evidence task is therefore a blinded review of representative workflows using a written eligibility rubric. The model is fit for prioritizing research; it is not yet fit for a revenue commitment.

Acceptance checklist

Accept the model only when a second reviewer can identify the unit, reproduce sourced counts, trace every filter and crosswalk, distinguish sourced values from assumptions, rerun the scenarios, and explain the largest uncertainties. Hold it when units are mixed, suppressed values are silently invented, percentages come from incompatible populations, formulas cannot be reproduced, or the conclusion exceeds the evidence.

Two Prune can provide assessment evidence about how participants research, calculate, verify, and communicate within a realistic market-sizing task. It does not validate a customer’s market estimate, provide investment advice, predict commercial outcomes, or make employment decisions.

Sources

  1. 1.County Business Patterns methodology · U.S. Census Bureau
  2. 2.Quarterly Census of Employment and Wages overview · U.S. Bureau of Labor Statistics
  3. 3.Assessing Data Reliability · U.S. Government Accountability Office

Explore this topic