4 ways to build an AI business

4/5/2026 ☼ AI ☼ meaningmaking ☼ not-knowing ☼ business models ☼ product management

tl;dr: Most people use “AI business” to mean one of two things — building the AI itself, or building products that call AI at runtime. There are actually four structurally distinct types. The other two — using AI to build products that then run without any AI, and building businesses that exist or matter primarily because AI proliferation has changed what markets are viable — are already tractable but remain mostly ignored.

Of late, when people say they’re building an “AI business,” they seem to mean one of two things. They are building the AI thing itself — foundation models, training infrastructure, chips. Or they’re building a product that calls AI at runtime — say, a product which calls an LLM when the user uses it. These two types of AI business absorb nearly all the attention, the venture capital, and the strategic analysis.

Two other structurally distinct types of AI business exist, are already tractable, and represent enormous underexploited opportunity. Most people haven’t noticed them because they’re still emerging and lack clear names. Lets call these two other types of AI business using AI to build Non-AI Things, and building Things Necessary/Important Because of AI. As usual, the high value work is being clear about category (which type of AI business you want to be in), and most people don’t seem to have done that thinking.

The four types

There are at least four possible relationships between a business and AI. The business can:

1. Build [AI Things] — the primary output is a model or model infrastructure.
2. Build [Things That Use AI] — AI is called at runtime; every user interaction involves a live model.
3. Use AI to build [Non-AI Things] — AI was used to build the product but the finished product is conventional software with no model called at runtime.
4. Build [Things Necessary/Important Because of AI] — the business exists or has grown substantially because AI exists at scale in the world.

Types 2 and 3 can look very similar from the outside, so it’s the most important conflation to fix. Type 2 calls the model every time you use the product. Type 3 used the model to build the product, but the product itself doesn’t run on AI. The runtime economics, the data exposure profile, the competitive moat, and the talent requirements are entirely different. Conflating them is the single most common error in how people talk about this space.

Type 1: Build [AI Things]

This is the infrastructure layer. Anthropic, which builds the Claude model family, and OpenAI, which builds the GPT series and ChatGPT, are canonical examples — their primary output is the underlying AI models, accessed via API, and everything else is distribution. The capital requirements are extreme and the relevant talent pool is small; most organisations that claim to be competing here are either building distribution for a type-2 product or haven’t thought clearly about which layer they actually occupy.

Type 2: Build [Things That Use AI]

GitHub Copilot, an AI coding assistant built into code editors, sends your code context to a model every time you pause typing. Grammarly, a writing assistant that checks grammar, style, and clarity, calls AI on each sentence as you complete it. These businesses depend on continuous model access, carry inference costs at scale, and their product quality is directly coupled to model capability. The design question that matters here isn’t “how do we use AI?” — that’s settled by the type. The real question is where the human-AI boundary belongs: which decisions must stay with the human, which can be offloaded, and what happens when the AI is wrong. The answer turns on meaningmaking, the work of making subjective decisions about what has value. Machines can generate outputs, but deciding which outputs are good enough — or should be rejected — is irreducibly human. For those building in this space, the design frame that works is tools that support meaningmaking rather than replace it — the product surfaces options, the human decides.

Type 3: Use AI to build [Non-AI Things]

This is the first category that most strategic analysis ignores, and the opportunity here is large.

The argument is economic, not technical. AI coding tools have made narrowly scoped, highly customised software cheap enough that tools previously impossible to justify commercially can now be built in an afternoon.

Rob Miller built a court submission preparation tool for a barristers’ chambers in a single afternoon. It runs on-premises for data protection compliance and makes PDF bundles with hyperlinked tables of contents that are formatted to court submission specifications. The task of preparing a digital court bundle previously required hours of manual work by an expensive human. Now it takes minutes, and no AI is called at runtime. Under old economics, writing a bespoke PDF bundler for a small chambers is commercially incoherent because the development cost can’t be recovered. With the new economics, it’s an afternoon project. These new economics open up a huge space for building and providing what I call Boring Tiny Tools (BTTs)—prosaic tools and software tailored to very specific needs with small addressable markets.

Most businesses and entrepreneurs don’t seem to have cottoned on to this new paradigm of how software should be built and to the scale of the opportunity it represents. The opportunity is enormous because the organisations that need BTTs are almost too numerous to count. Who needs BTTs? Any organisation 1) that has non-discretionary repetitive human business processes that are too specific for enterprise software to serve and 2) which is too small for traditional custom software development to make sense. These organisations are everywhere. A 10-person law firm with a specific document formatting requirement is one. A 35-person local government office that processes a particular form type thousands of times a year is another. A 4-person supply chain team in a 200-person company running the same spreadsheet manipulation weekly is a third. None of these organisations appear in the addressable market of enterprise software vendors. But they all could use BTTs that are newly cheap to make and deploy.

The bottleneck here isn’t technical, because BTTs are never about cutting edge technology. The rate-limiter instead is knowing what to build and for whom. Identifying which processes are worth automating, scoping the tool correctly, and building trust with organisations that have never commissioned custom software — these are the skills that matter, and the approach this requires is amorphous, where scope is discovered through observation rather than defined in advance.

Type 4: Build [Things Necessary/Important Because of AI]

The fourth category is the least mapped. It has two distinct sub-types.

The first is businesses that are newly necessary — they didn’t exist before AI proliferation created the conditions for them. Scale AI, a data labelling and annotation company, exists because training AI systems creates enormous demand for human-annotated data — hundreds of millions of labelled examples for vision systems, RLHF pipelines, and other supervised applications. The market for this was tiny before the AI boom. Lakera, an AI security company, addresses prompt injection security — a threat surface with few close analogues in traditional security practice. Conventional security firms lack the training data, evaluation methods, and threat models relevant to this domain. Both businesses are responses to structural conditions that AI created.

The second sub-type is businesses that are newly important — they existed before, but AI proliferation has substantially enlarged their purpose or market. DeepLearning.AI, which offers online courses in AI and machine learning, was founded in 2017 as a niche platform for practitioners. When ChatGPT arrived, it created an entirely new audience of knowledge workers needing AI literacy. The platform now serves seven million learners. C2PA, the Content Credentials standard (an open protocol for embedding verifiable provenance in images and video) was a niche technical initiative before generative AI made photorealistic synthetic images routine. Origin verification became a genuine social problem that has driven investment in cryptographic provenance infrastructure. The standard now has over 5,000 member organisations.

The “newly important” sub-type is the most poorly defined as a category. There’s no clean label for “businesses that were fine before and are now structurally more important because of AI.” Absence of a label is a signal that the category is underexplored. Those who map it first have a real advantage.

Where attention and opportunity diverge

Type 1: Build AI Things	Type 2: Build Things That Use AI	Type 3: Use AI to Build Non-AI Things	Type 4: Build Things Necessary/Important Because of AI
The AI model (or hardware) is the product — everything else is distribution on top of it.	The finished product calls a live AI model at runtime; the AI is continuously embedded and running during use.	AI built the software but is absent from the finished product — the shipped tool runs with no AI at runtime.	The business exists or has substantially grown because AI proliferation created conditions — markets or needs — that didn’t exist before.
Anthropic, OpenAI	GitHub Copilot, Grammarly	Court bundle-making tools; Boring Tiny Tools broadly	Scale AI, Lakera, prompt academies

Types 1 and 2 are crowded, well-understood, and capital-intensive. Type 1 requires frontier research capability and compute budgets already reaching hundreds of millions of dollars per training run — and climbing. Type 2 requires finding product-market fit in an environment where model providers can replicate your feature set in a product update.

Type 3 requires minimal capital, is tractable today with current tools, and addresses a market — the long tail of organisations with specific repetitive processes — that conventional digital transformation consulting has structurally ignored. Type 4, particularly the “newly important” sub-type, is large and growing, and the absence of clear naming means most of it remains unsearched.

Focusing nearly all attention on the already oversaturated and overhyped types of AI business (1 and 2) and nearly none on the underexplored but clearly valuable and voluminous types (3 and 4) is a big conceptual oversight.

The map is the tool

Knowing which type of AI business you’re actually building determines what skills matter, what the unit economics look like, who your real competition is, and what risks will materialise at scale. Most organisations building “with AI” haven’t asked themselves which of these four things they are doing. That question — the categorisation question — is where rigorous strategy begins.