AI Market Sizing: Why Every Model Gives a Different TAM

AI Market Sizing: The Spread Is the Finding

AI market sizing is now the first draft of almost every strategy deck, due diligence memo, and investment case, and it has a problem that anyone who has tried it will recognise. Ask several models to size the same market and the answers can differ by multiples, each delivered with a confident CAGR and a named research firm as the source. The timing makes this matter. Firms are repositioning as AI-first, recent McKinsey research suggests the payoff from AI is still concentrated in a small group of leaders, and Bain has argued the industry needs something like six trillion dollars in annual revenue to justify the data centre build-out. Clients are asking sharper questions about market size, and the first number on the slide is increasingly written by a model.

Four Ways to Size a Market

Each method has a failure mode, and the useful question is which failures you can see and which stay hidden.

Why AI Market Sizing Numbers Diverge

When two models disagree about a market by a factor of three, the instinct is to assume one of them is wrong. Usually both are answering slightly different questions, and neither tells you which question it picked. The spread comes from a small number of sources, and once you know them you can read the disagreement rather than average it.

The Definition Problem in AI Market Sizing

Market definition is the largest driver by far. Does "AI in food and beverage" include processing equipment with embedded vision systems, or only software? Does a cybersecurity market include services, or just products? Is the figure end-user spend or vendor revenue? A model filling in those gaps silently will produce a number that is internally consistent and incomparable with the next model's number. The same is true of published reports, where research firms commonly size what sounds like the same market at figures that sit far apart.

Beyond definition, a few other factors widen the range:

• Base year and currency. A figure from an older report, compounded forward at an assumed growth rate, drifts further from reality with every year.
• Training data age. A model may anchor on a report that has since been revised, or on a pre-boom estimate for a market that has moved quickly.
• Compounding. A small difference in assumed growth rate becomes a large difference in a long-horizon forecast.
• Geography. Global, developed markets only, and a single region are frequently mixed in one answer.

The Fabricated Citation Risk

The most dangerous output in AI market sizing is not a wrong number. It is a wrong number with a real-sounding source. Models commonly attribute figures to well-known research firms, and the result falls into one of three categories: the report exists and says that, the report exists and says something different, or the report does not exist. From the slide, all three look identical.

Precision is a warning sign rather than a comfort. A figure quoted to two decimal places with a named publisher and a specific forecast year feels authoritative, and that is exactly the form a fabricated citation takes. We covered how these errors survive partner review in AI in consulting. In market sizing the rule is simple: if nobody on the team has opened the source, the number is an assumption and should be labelled as one.

A Triangulation Method That Holds Up

This approach uses AI where it is strong, structure and breadth, and keeps humans responsible for the parts that need evidence.

1. Write the market definition first. What is in and out, which geography, which year, which currency, and whether the figure is spend or revenue.
2. Ask models for structure, not the answer. Request the segments, buyer types, and drivers that would make up the market. Models are good at this.
3. Build bottom-up with your own inputs. Number of potential buyers, realistic adoption, and price, each sourced or explicitly assumed.
4. Add top-down anchors from verified sources only. Use published figures your team has actually read, with their definitions noted.
5. Run the same structured prompt across several models. Tabulate each model's assumptions, not just its final number.
6. Reconcile the gaps. Explain why bottom-up and top-down differ. The explanation is often more valuable to the client than either figure.
7. Present a range with sensitivities. Show what happens when the two biggest assumptions move, and say which one you would test first.

What a Defensible Estimate Looks Like

A market size survives client scrutiny when it shows its working. In practice that means a stated definition, every figure either sourced or labelled as an assumption, a range rather than a single point, a reconciliation between methods, and a date stamp so readers know how current it is. None of this is new to consulting. What has changed is that AI makes it easy to skip straight to a confident single number, and clients have noticed. A buyer who can generate their own TAM in thirty seconds is not paying for the number. They are paying for the reasoning behind it.

Pros and Cons of AI in Market Sizing

Used well, AI makes market sizing faster and more thorough. Used carelessly, it makes it faster and less honest.

• Pro: rapid structure. A segment tree and driver list in minutes gives the team a starting point to argue with.
• Pro: wider coverage. Models surface adjacent segments and buyer types a small team might overlook.
• Pro: visible disagreement. Comparing models shows which assumptions are contested before a client points it out.
• Con: invented sources. Plausible citations to real firms are the most common and most damaging failure.
• Con: silent definitions. Models rarely state what they included, which makes numbers incomparable.
• Con: anchoring. Once a big number appears on a draft slide, teams tend to defend it rather than test it.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how AI market sizing plays out in practice rather than presented as verified case studies.

A diligence team receives a target's management presentation claiming a very large addressable market, sourced to a research firm. An analyst runs the definition through several models and finds the figure includes hardware and services the target does not sell. The serviceable market is a fraction of the headline, and the valuation conversation changes accordingly.

A strategy team asks one model to size a niche software category and gets a precise figure attributed to a named publisher. A junior consultant tries to buy the report for the appendix and cannot find it. The number is quietly replaced by a bottom-up build before the steering committee, which is the right outcome reached by luck rather than process.

A team runs the same prompt across six models and gets estimates spread across a wide range. Tabulating assumptions shows four models include public sector buyers and two do not. That single definitional choice explains most of the spread, and it becomes the first question for the client.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

TAM, SAM, SOM, and Where AI Overreaches

Models are reasonably good at total addressable market framing and noticeably weaker as the funnel narrows. Serviceable available market depends on your client's actual product, channels, and geography. Serviceable obtainable market depends on competitive position, sales capacity, and pricing power, none of which a general model knows. A common pattern is a model presenting a TAM as though it were achievable revenue, or blending the three into one figure.

Founders face the same trap in fundraising, where an inflated TAM costs credibility with exactly the investors who know the market. The approach in our pitch deck playbook applies here too: let models challenge the number rather than produce it.

Why Talkory Wins

Triangulation is the heart of credible market sizing, and Talkory makes the AI part of it fast. It runs one structured prompt across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in a single pass, so a team can compare six sets of assumptions side by side rather than trusting one opaque number. Where the models agree on structure and drivers, that is a reasonable foundation to build on. Where they diverge on definition, growth rate, or source, the disagreement shows the consultant exactly which assumptions to verify and which questions to put to the client. It does not replace sourcing. It tells you where sourcing matters most.

Final Verdict

AI market sizing is useful when it is treated as a way to expose assumptions and unreliable when it is treated as a way to produce a number. The spread between models is not a flaw to average away. It is a map of the definitional and growth choices a consultant will have to defend in the room. Define the market first, ask models for structure, build bottom-up with your own inputs, verify every cited figure, reconcile the methods, and present a range with sensitivities. Clients can get a number from a chatbot. What they still need is someone who can explain it.