AI Product Market Sizing: How to Size a Market That Does Not Exist Yet

A traditional market sizing exercise for a project management tool looks roughly like this: count the number of knowledge workers in your target industry, multiply by the average per-seat SaaS price for that category, and you have your TAM. It is crude but directionally useful because a project management tool competes with other project management tools for an existing budget line item.

AI-native products do not work this way. They frequently replace activities rather than software tools. An AI that handles first-line legal contract review does not compete with contract management software (a defined SaaS category with known spend). It competes with the hours a paralegal billed to do the same task. The market is not the contract management software market. It is a portion of the legal services labor market.

This creates several specific problems for standard TAM/SAM/SOM:

No single approach works for every AI product. The right method depends on whether you are replacing software, replacing labor, or creating net-new economic activity. Often the most credible sizing exercise uses two approaches and triangulates between them.

Top-down sizing works when your product is a clear upgrade to an existing software category. An AI-powered CRM that replaces Salesforce's standard tier is easier to size top-down than a net-new AI product category, because the CRM market is well-documented and the buyer already has a budget line for CRM software.

The top-down chain for an AI product looks like this: global market for the predecessor category, then your addressable geography and segments (your SAM), then a realistic penetration rate given your competitive position (your SOM). The critical error in AI top-down sizing is using the "AI market" as the starting category (inflated and ill-defined) rather than the specific predecessor category your product displaces.

Bottom-up sizing is the most rigorous approach and, for AI products, the most honest. It requires you to know (or credibly estimate) three numbers: how many target users exist, how many units of value each user generates per period, and what you can charge per unit. Multiply and you have a market size built from real economics rather than from analyst reports and percentage-share assumptions.

For AI products, the "unit of value" is often not a seat. Define it as the specific output or transaction your product delivers: a completed research report, a processed invoice, a reviewed contract clause, a resolved support ticket. Count how many of these outputs occur in your target market per year. Multiply by the price you can charge per output at scale.

Value-based sizing starts from the economic value your product creates rather than from market categories or seat counts. It is the most defensible approach for AI products that replace high-cost professional labor, because it ties your market estimate directly to the value chain you are entering.

The framework has three inputs: the total labor cost of the activities your product automates or assists (this is your value pool), the productivity multiplier your product creates (how much time does the AI save as a fraction of total task time), and the capture rate (what fraction of the savings the market will pay to your product rather than retain internally).

The capture rate is the most debated input. AI products in competitive markets typically capture 5 to 15% of the economic value they create over the medium term. Higher capture rates (15 to 30%) are possible when there is strong product differentiation, high switching costs, or a workflow monopoly. Anything above 30% is unlikely to hold at scale because it invites competition and buyer resistance.

AI market sizing mistakes fall into two categories: methodological errors (using the wrong inputs or formula) and framing errors (choosing the wrong market to size). Both are common, and both produce estimates that are 5 to 10x too high and then fail to survive investor or executive scrutiny.