Capability, Product and Business Model

Every AI pitch contains three questions, and many pitches answer only the first.

Can it be done? A model drafts the contract, finds the bug, reads the scan, answers the support ticket. That's a capability. Researchers, demos and benchmarks live here.

Will someone use it in their actual work? Every week, after the novelty wears off, inside the tools and approvals their job already has, at an error rate they can live with. That's a product.

Can that use be delivered again and again, at a cost below what it's worth to the customer, against competitors? That's a business.

In ordinary software these questions arrived in order, and answering the first one bought time for the other two. If you built something hard, it took competitors months or years to copy it, and you used that time to find users and a price. AI broke that sequence. The capability question is now answered for everyone at once, every time a lab ships a better model. So a technical edge in AI is perishable, and the company has to be carried by the product and the business, which are much harder to copy.

Scroll sideways to see the whole figure

How perishable? Epoch AI estimates, across five benchmarks, that reaching a fixed score has become 47 percent cheaper every quarter since 2023. That compounds to roughly thirteen times cheaper a year. A capability that set you apart in January can be a default setting by December.

In AI the first question gets cheaper every quarter, so the second and third have to carry the company.

Capability still matters. The useful question is what it buys you before it expires, and companies usually fail to convert it at one of three gaps.

Gap one: from "it can" to "they do"

Averages hide the failures users feel

A capability is reported as an average. A product is judged by its failures, and what counts is the failure rate the user can tolerate for that task. A marketer drafting ten subject lines tolerates plenty of misses, because picking the good ones is quick. A finance team reconciling payments tolerates almost none, because each miss costs an investigation.

Agents make this much harder, and the arithmetic is worth seeing once. Suppose a task takes ten steps and each step succeeds 95 percent of the time. The whole task succeeds only about 60 percent of the time, because the failures multiply: 0.95 times itself ten times is 0.6. If the user needs the result right 98 times out of 100, each of those ten steps must be right about 99.8 percent of the time.

So "95 percent on the benchmark" and "a product people trust" can be two different engineering problems. And consistency matters as much as the average: a tool that usually works but sometimes doesn't feels unreliable, and people stop handing it work that matters.

A benchmark measures a configuration

Researchers turning results into companies hit this gap before they have a single user: the model that achieves a result is often not the model that can be served.

In December 2024 a preview of OpenAI's o3 scored 76 percent on the ARC-AGI-1 reasoning benchmark, and 88 percent with more computing power. When the ARC Prize team tested the o3 actually released in April 2025, it scored 41 percent at low reasoning effort and 53 percent at medium. The released model was a different one, tuned for chat and product use. The compute budget behind the preview wasn't available in production. And the preview had been trained on much of the benchmark's public training data.

A benchmark score describes a configuration. A product ships the one it can afford.

For a founder the rule is simple: price the result you can serve at the speed and cost your customer will accept, and treat the best number in the paper as a ceiling.

The things no leaderboard shows

Four more conditions decide whether a capability becomes daily use.

Speed. An answer that takes two minutes changes how people work. They stop iterating and start batching, or they stop. When Cursor launched its own coding model in 2025, it sold it on speed as much as quality.

Integration. Work happens inside the editor, the ticketing system, the document store, the ERP. A tool that needs text copied into a separate window gets used for a few weeks. A tool that reads and writes where the work lives gets used without anyone deciding to use it again.

Trust. Before a team lets a tool act, someone has to believe it won't embarrass them. That belief comes from a track record on their own cases, a cheap way to check the output, and a way to undo mistakes.

Onboarding. People have to learn what to ask, what to check and when not to use it. If that's harder than the old way, adoption stops at the enthusiasts.

The public evidence for this gap is old but telling. In 2023 Sequoia Capital found that generative AI apps had a median of 14 percent of monthly users coming back daily, against 60 to 65 percent for the best consumer companies. Many people tried; few returned. My judgment is that these four conditions explain more of that gap than the model does.

Gap two: from use to a business

Every use costs money

Classic software has a convenient property: once it's built, serving one more user costs very little. So its gross margin, the share of revenue left after the direct cost of delivering the product, is high and stays high as customers use it more. Salesforce reported 77 percent for its fiscal year ending January 2025.

An AI product pays for every use. Each request reads and writes tokens, chunks of text a bit shorter than a word, and someone pays the model provider for each one. The more a customer uses the product, the higher the cost.

Take an illustrative product priced at $40 per user per month. Each task a user hands it (summarize these documents, draft this reply, fix this bug) takes eight model calls, each reading about 30,000 tokens and writing 1,500. On Anthropic's workhorse model, Claude Sonnet 5.5, at list prices, one task costs about 31 cents. Add $4 per seat for everything else and assume 21 working days a month.

About five and a half. Each daily task costs about $6.55 a month, and $36 of revenue remains after the other costs. To match Salesforce's 77 percent margin, a user could hand it less than one task a day.

Scroll sideways to see the whole figure

Two things follow, and both matter more than the specific numbers.

Your most enthusiastic customers are your least profitable. Suppose 40 percent of seats run half a task a day, 30 percent run two, 20 percent run five and 10 percent run fifteen. The blended margin is 36 percent. The heaviest tenth accounts for 45 percent of all model cost, and each of those users costs $98 a month against $40 of revenue. In classic software heavy use is the best sign of product fit. Here it's still a sign of fit, and it's also where the losses come from (Why AI Products Have Thinner Margins).

Choosing a model is a pricing decision. On the top model, Claude Fable 5.1, the same product breaks even at about one task a day. On the small model, Haiku 5.5, it barely notices usage. A team that sends every step to the strongest model because it scores best in their tests has made a margin decision without noticing.

Real companies hit this wall in 2025. In June, Cursor switched its $20 plan from a fixed number of requests to a credit pool priced at the model providers' rates. Users got surprise charges; the company apologized, refunded them and explained that "the hardest requests cost an order of magnitude more than simple ones." In July, Anthropic announced weekly limits for subscribers running its Claude Code agent around the clock, affecting fewer than 5 percent of them. Both were products so good that people used them far more than the price assumed.

Thin margins and expensive customers multiply

Margin is only the first line. A business also pays to win each customer and needs them to stay long enough to repay it.

Suppose winning one seat costs $480 in sales and marketing, the customer acquisition cost. At a 77 percent margin, each $40 seat earns back that cost in about 16 months. At the 36 percent blended margin above, it takes about 33. A thinner margin more than doubles the time each customer has to stay before the company earns anything from them.

That's why distribution matters even more in AI than in classic software. An incumbent adding an AI feature to an account it already has pays far less to win that user than a startup does (Distribution Beats the Better Model). And retention depends on gap one: a product that never became part of someone's daily work doesn't renew, however impressive the trial.

Thin margins and expensive distribution multiply each other. A company can survive one of them, rarely both.

Gap three: when the edge resets

The third gap runs through all three layers at once. A model release re-answers the capability question for everyone, and a company whose product and business rest only on that answer finds both exposed in the same week.

Jasper is the case to remember. It launched in early 2021 on OpenAI's GPT-3, selling AI-written marketing copy, and reported $45 million of revenue that year. In October 2022 it raised $125 million at a $1.5 billion valuation, with more than 70,000 customers. Six weeks later OpenAI released ChatGPT, free. Two months after that came ChatGPT Plus at $20 a month; most of Jasper's customers reportedly paid $80. By September 2023 Jasper had cut staff, reportedly lowered its revenue forecast by at least 30 percent, and changed its CEO.

ChatGPT wasn't the only cause, but the three layers make the mechanism clear. Jasper's capability was OpenAI's capability. Its product was a good interface for marketers, but a free chat window covered most of what an individual marketer needed. And its business charged four times the price of the model owner's own product for overlapping value. One release exposed all three. Jasper's reported response was to move toward marketing teams at midsize and enterprise customers, where brand rules, collaboration and approvals matter: an attempt to move its value up from capability to product.

Chegg shows the same mechanism from the other side. It had a brand students knew and a working subscription business, a real product and a real business. But what it sold was a capability: answers to homework questions. In May 2023, after it acknowledged ChatGPT had become a serious competitor, its shares fell by more than a third in a day. By October 2025 it had cut about 45 percent of its remaining staff and turned toward selling skills training to businesses.

A product protects a capability only when it is anchored in work the model doesn't do alone.

Distribution and habit protect a company only if what travels through them is something the model can't deliver by itself. When the core of the offer becomes free, a channel to customers is a channel for something they no longer need to buy.

What a durable position looks like

The opposite case is a company built on other labs' models that kept growing while those labs shipped competing products. Cursor, the code editor, passed $1 billion in annualized revenue in November 2025, months after Anthropic launched Claude Code. Three things in its position fit the argument. It sits in the editor, where programmers spend their day, so it gets used without anyone opening it. Each better model from any lab improves it, because users choose among models instead of the product compensating for one model's weaknesses (A Product That Improves With the Model). And it took part of its cost and speed into its own hands with its own models. Its 2025 pricing episode shows the business layer was still the hardest one.

The general rule is the one in Where the Moat Moves: what stays defensible is made of time, rights and other people's decisions. Data you're allowed to use, permission to work inside a customer's systems, a place in their daily work, a relationship that already exists. A capability is made of effort, and AI is making effort cheap.

"But great technology always finds a business"

A good founder or researcher will push back: the companies that matter were built on real breakthroughs, so focus on the capability and the rest will follow.

The objection is right about the technology and often wrong about who collects. Google researchers published the transformer architecture in 2017, and by 2019 Google was using it to rank search results. The product that opened a new market with it, ChatGPT, came from OpenAI five years later. The technology found a business; the new market belonged to someone else. Jasper runs the other way: GPT-3 found a business fast, and when the model's owner sold the same capability directly, the business moved to the owner.

The objection holds in one situation: when a capability stays scarce long enough for its owner to build the other two layers on top. That's the position of the leading labs themselves, whose edge costs billions to reproduce and is renewed every few months. Even they are moving hard into products, distribution and enterprise sales, which suggests they don't expect capability alone to hold. For a startup renting its capability from a lab, the scarcity belongs to the landlord.

This could change if one lab opened a lead others couldn't close for years, or if the most capable models became as cheap as today's small ones. As of October 2026, neither is happening.

What to do

Read any AI company, including your own, through three sets of questions.

Capability: What would the next model release do to this? How much of the result survives at the cost, speed and safety settings the product can actually serve?

Product: Who uses it every week without being reminded? Inside which existing tool? At what error rate do they stop trusting it?

Business: What does the heaviest tenth of users cost against what they pay? How many months does a new customer take to repay what it cost to win them? What happens to both numbers if a competitor who already has the customer ships something good enough?

A company that answers the first question well and the other two vaguely has a result. A company that answers all three has a business whose technical edge can expire without taking the company with it.

Check yourself

Why does a technical edge in AI expire faster than in ordinary software?

The capability question is answered for everyone at once whenever a lab ships a better model. Reaching a fixed score has become far cheaper every quarter.

Why can a product's heaviest users be its least profitable ones?

Each use pays the model provider for tokens, so cost rises with usage while the price stays flat. In the example the heaviest tenth cost $98 a month against $40 of revenue.

Why was Jasper exposed when ChatGPT arrived?

Its capability was OpenAI's capability, its interface was mostly covered by a free chat window, and it charged several times the price of the model owner's product. One release exposed all three layers.