Case Study: Artificial Intelligence Market Sizing: A Patent-Anchored Case Study

This artificial intelligence market sizing case study follows a Series-B enterprise-AI company preparing to raise capital for a new generative-AI governance product, and shows how we built a defensible TAM/SAM/SOM by reading public patent signals alongside published investment data – rather than reverse-engineering an analyst’s headline number. It is a representative scenario: every hard figure below is a cited public statistic, and every model output is framed as what the method produces, not a disclosed client result.

The Challenge

The client had strong engineering and early revenue but a term sheet stalled on one question the founders could not answer with conviction: how big is the market you are actually protecting, and is it growing fast enough to justify the round? Their deck cited a $1.3-trillion “AI market” figure lifted from a syndicated report. Investors saw through it immediately – a top-down number with no bottom-up logic and no link to the company’s defensible position.

The problem was not ambition; it was evidence. A generic trillion-dollar number tells a partner nothing about which slice of AI the company can win, how quickly that slice is compounding, or whether anyone else has already filed around it. Growth-stage diligence does not reward the biggest number on the slide – it rewards the number the founders can defend line by line when a partner starts pulling on the assumptions.

The board needed three things before the next partner meeting: a TAM/SAM/SOM they could defend line by line, evidence that the generative-AI governance segment was accelerating rather than peaking, and a view of whether the company’s patent position mapped onto the fastest-growing part of the opportunity or a crowded corner of it. The stakes were concrete – a mispriced market meant either a down-round narrative or an over-reach that would collapse in diligence.

Our Approach

We ran the engagement as a bottom-up market sizing tied to the IP position, sequenced so each step fed the next:

• Segment definition from patent taxonomy. We used the WIPO generative-AI patent taxonomy – GANs, large language models, diffusion models and downstream application areas – to carve the market into segments that map to real, filed technology rather than marketing categories. Defining the market this way means every segment boundary can be traced to a classification an examiner recognises.
• Growth signal from filing curves. For each segment we pulled the public family-count trajectory. LLM-related patent families rose from 53 in 2020 to 881 in 2023, and diffusion-model families from 18 to 441 over the same window – a filing acceleration that tells you where R&D capital is actually going 18-24 months ahead of revenue.
• Assignee benchmarking. We benchmarked the top applicants – Tencent, Ping An, Baidu, the Chinese Academy of Sciences and IBM – to see whether the governance and compliance sub-segment was already contested or still open, and to understand who the client would be filing against if it pushed into core architecture.
• Bottom-up TAM/SAM/SOM. We built the market model from named buyer populations and defensible per-seat and per-deployment assumptions, logging every assumption to a source, then cross-checked the top of the funnel against published AI investment flows so the ceiling was anchored to money that has actually moved.
• Opportunity ranking. Finally we ranked five adjacent segments by growth rate, contestedness and fit to the client’s granted and pending claims, so the board could see not just how big the market is but which part of it the company is best positioned to defend.

The discipline that made the model credible was the assumptions log. Every input – the buyer count, the attach rate, the average contract value, the growth multiplier – was written down with the source it came from and the reason it was chosen. When a partner later asked “where does the SAM growth rate come from?”, the answer was a specific patent-filing curve and a specific investment figure, not a gut feel.

What the Research Found

The patent record reframed the story. Generative AI is not a single market peaking on hype – it is a set of segments filing at very different rates. WIPO recorded more than 54,000 generative-AI patent families published between 2014 and 2023, with the count rising from 733 in 2014, and generative-AI patenting increasing by more than 800% since the transformer architecture appeared in 2017. That is an innovation curve still bending upward, not flattening.

The concentration signal mattered just as much. China accounts for more than 38,000 of those families and the United States roughly 6,300, but that concentration sits in core model architecture. The governance, evaluation and compliance layer the client was building was comparatively thin in the patent record – a white space, not a crowded field. The top applicants (Tencent, Ping An, Baidu, the Chinese Academy of Sciences and IBM) had piled into the model layer; almost none had staked the compliance layer.

Meanwhile the demand side was clearly funded. Stanford’s AI Index reports that generative AI attracted $33.9 billion in global private investment in 2024, up 18.7% year on year and now more than a fifth of all AI private investment, against total corporate AI investment of $252.3 billion. Regional depth was there too: US private AI investment alone reached $109.1 billion. And the filing appetite is spreading to Europe – the EPO logged a record 201,974 patent applications in 2025, with computer technology up 6.1% on the back of AI advances.

Put together, the bottom-up SAM the client could credibly claim was far smaller than the $1.3-trillion headline – but it was growing faster than the aggregate, sat in a defensible patent white space, and was backed by verifiable investment flows. That is a story a partner can underwrite.

The Outcome

The founders replaced the borrowed trillion-dollar slide with a three-layer model they could defend line by line: a TAM anchored to the governance-and-compliance spend implied by funded generative-AI deployment, a SAM scoped to the buyer segments their product actually serves, and a SOM tied to their sales capacity over the raise horizon. Each number carried a citation or a logged assumption.

Crucially, the sizing was wired to the IP position. Because the fastest-filing segments (core LLM and diffusion architecture) were also the most contested, and the client’s claims sat in the less-crowded governance layer, the market story and the defensibility story pointed the same way. Instead of claiming the whole trillion-dollar market, the company claimed a smaller, faster, protectable slice – and could show why competitors would find it expensive to follow.

The round advanced on a number the company could stand behind in diligence. More durably, the founders kept the assumptions log as a living model: when the next WIPO and AI Index data dropped, they refreshed the curves and the sizing moved with the evidence rather than with the pitch.

What This Means for Similar Matters

• A top-down “the AI market is $X trillion” slide is a liability in diligence. A smaller, bottom-up SAM with a source-cited assumptions log wins the room.
• Patent filing curves are a leading indicator. Segment-level family counts show where R&D capital is committed 18-24 months before it shows up as revenue.
• Size the segment you can defend. When the market model and the patent white space point at the same sub-segment, the fundraise story becomes coherent.
• Concentration cuts both ways – a field dominated by a few assignees can still contain an open adjacent layer worth building in.
• Keep the model alive. Because every input is sourced, refreshing it against new WIPO or AI Index data is a data update, not a rebuild.

The Three-Layer Model, Line by Line

The deliverable was a single model with three defensible layers. The TAM was the total governance-and-compliance spend implied by funded generative-AI deployment – derived from the investment flows above rather than a syndicated headline. The SAM narrowed that to the buyer segments the product actually serves: regulated enterprises deploying generative AI where evaluation and audit are non-negotiable. The SOM was the realistic three-year capture given the client’s sales capacity and pipeline. Each layer was built upward from named populations and per-unit assumptions, and each assumption cited a source. The point of the exercise is not the final number but its auditability: a partner can start at the SOM and walk all the way up to the TAM without hitting a single unsupported leap.

Why Patent Curves Lead Revenue

Companies file before they ship. A patent family is typically published 18 months after priority, and the invention it protects usually reaches the market later still – so a rising family count is a paid-for commitment to a technology direction that has not yet shown up in anyone’s revenue. That is why segment-level filing curves are such a useful demand signal for artificial intelligence market sizing. When LLM families climb from 53 to 881 in three years while an adjacent layer stays flat, the market is telling you where capital conviction sits and where it does not – well before the analysts catch up. We read those curves per segment, not in aggregate, because the aggregate hides exactly the divergence a founder needs to see.

Data Sources

The market and patent data referenced above comes from:

• WIPO – Patent Landscape Report: Generative AI (2024) — 54,000+ GenAI patent families 2014-2023; China 38,000+, US ~6,300; 800%+ growth since 2017; LLM families 53 to 881 and diffusion 18 to 441 (2020-2023).
• Stanford HAI – AI Index Report 2025 (Economy) — Generative AI private investment $33.9B in 2024 (+18.7%), >20% of AI private investment; total corporate AI investment $252.3B; US private AI investment $109.1B.
• EPO – Technology Dashboard / Patent Index 2025 — Record 201,974 European patent applications in 2025; computer technology up 6.1%, driven by AI advances.

Discuss an AI Market Sizing Study

Raising capital or planning a build? We size the opportunity bottom-up and tie it to the patent position that protects it.

Frequently Asked Questions