AI SaaS: Building a Product on Top of the Models

AI SaaS is subscription software whose core work is done by an AI model, usually a foundation model from a provider such as OpenAI, Anthropic or Google, called through an API and wrapped in a product that does one job for a defined customer. The model supplies general capability. The product supplies everything the model does not: the workflow, the context, the interface, the checks and the accountability.

It is the most ambitious of the four ways to productize what you know, and the one where the gap between a demo and a business is widest. A demo needs a prompt. A business needs sound economics, a defensible position and a clear answer to who is responsible when the model is wrong.

The economics are not classic software economics

Classic software is expensive to build and almost free to serve: once the product exists, one more customer costs very little. AI software pays for every use. Each request consumes tokens that someone has to buy, so a heavy customer can cost real money to serve.

That shows up in the margins. ICONIQ’s State of AI survey of around 305 executives at software companies building AI products, published in July 2026, put average gross margins on AI products at 45% in 2025, projected to reach 53% in 2026. Bessemer Venture Partners’ State of AI 2025 found its fastest-growing group of AI start-ups running at about 25% gross margins, and even its steadier, healthier group, at around 60%, sits slightly below traditional software peers.

Three practical consequences follow.

• Price has to track cost. Flat per-seat pricing over unlimited AI usage lets your heaviest users cost more than they pay. ICONIQ found consumption-based pricing among AI builders rising from 35% to 42% in six months and outcome-based pricing from 18% to 23%, with companies blending 1.7 pricing models on average.
• Overruns come from usage, not headcount. In the same survey, token spend was the most common source of budget overruns once products moved from single requests to multi-step agent workflows.
• Routing is a margin decision. ICONIQ’s respondents ranked reducing inference costs and routing strategies as their top two levers for improving margins: send routine tasks to smaller, cheaper models and keep the expensive ones for the hard cases.

Where the defensible part sits

Every AI SaaS company rents its intelligence from the same handful of providers. So the model cannot be the moat. Your competitors can rent the same one, and the provider can ship your feature to everyone.

What is defensible is what the model cannot get on its own.

• Domain context. The rules, examples, standards and edge cases of one field, written down and maintained. It is the same asset knowledge work increasingly depends on, packaged as a product.
• Proprietary data, handled properly. Data your customers trust you with, and the improvements that come from seeing many cases. Data sovereignty matters twice here: for your customers, and for you.
• Workflow integration. A product wired into the systems a customer already runs is far harder to replace than a separate tool with a chat box.
• Accountability. In regulated or high-stakes work, buyers pay for someone who stands behind the output. No model provider offers that.

Own the layer that matters

The architectural rule is the one vendor lock-in applies to companies buying AI, turned around for companies building it: make the part that is hard to move the part you own, and make the part that is easy to move the model underneath.

In practice that means calling models through an interface you control, keeping your prompts, rules and evaluation tests as versioned assets in your own systems, and testing new models against those tests before you switch. A product built that way gets better and cheaper every time the providers compete for your business. A product hard-wired to one provider’s API absorbs every price rise and every retired model. The same build versus buy reasoning holds inside your own product: buy the commodity components and build the layer that is genuinely yours.

The obligations you take on

Selling an AI product makes you more than a user of AI. Under the EU AI Act, a company that places an AI system on the market under its own name is its provider (Article 3), and providers carry the obligations. Two matter for most AI SaaS products.

• Transparency. Systems that interact directly with people must be designed so those people know they are dealing with an AI (Article 50).
• High-risk uses. If the product is used for purposes the Act lists as high-risk, such as decisions on employment, credit or access to essential services, the provider duties become substantial: risk management, technical documentation, human oversight and a conformity assessment. A company that puts its own name on someone else’s high-risk system can become its provider too (Article 25).

Data protection runs alongside. If your product processes your customers’ personal data, you are usually their processor under the GDPR, which means a data processing agreement with each customer, and your model provider becomes a sub-processor in that chain. AI governance covers the operating side of meeting these duties. None of it is exotic, but it has to be designed in rather than discovered after the first enterprise customer asks.

Small product or large company

Not every AI product needs to be a venture-scale company. Many of the most sensible ones are a micro-SaaS: one job, one niche, a small team, priced to cover its running costs. The venture path makes sense when the market is broad, the data advantage grows with every customer and you are prepared to fund thin margins while it does.

Either way, the question to settle first is not which model to use. It is what you know that the model does not, and whether enough customers will pay for that every month.

The part of an AI product nobody can rent is its context: the instructions, examples and standards that make a model produce your kind of work. If you are building an AI product, or have built one that now has to hold with real customers, tell Daniel what it does, what it runs on, and what the business needs from it. Talk to Daniel.

Frequently Asked Questions

What is AI SaaS?

AI SaaS is subscription software whose core function is performed by an AI model, usually a foundation model accessed through an API, wrapped in a product that does a specific job for a defined customer and adds the workflow, context and interface the model lacks.

How is AI SaaS different from traditional SaaS?

Mainly in its economics. Traditional software costs little to serve once built, while AI software pays for every use in model fees. ICONIQ’s 2026 survey put average gross margins on AI products at 45% in 2025, and AI builders are moving towards usage- and outcome-based pricing to match.

What makes an AI SaaS product defensible?

Not the model, which competitors can rent too. Defensibility comes from domain context, proprietary data handled with care, integration into customers’ existing workflows and accountability for the output, none of which a model provider can ship as a generic feature.

How should you price an AI SaaS product?

In line with what it costs to serve. Flat per-seat pricing over unlimited AI usage can make heavy users unprofitable. Many builders now combine a subscription with usage-based or outcome-based components, and route routine work to cheaper models to protect margins.

Does the EU AI Act apply to AI SaaS companies?

Yes, if they offer the product in the EU. A company that places an AI system on the market under its own name is its provider under the Act. Transparency duties apply to systems that interact with people, and products used for high-risk purposes such as hiring or credit decisions carry substantial additional obligations.

How do you avoid depending on one AI model provider?

Call models through an interface you control, keep prompts, rules and evaluation tests as your own versioned assets, and test alternative models against those tests regularly. That lets you switch providers when price or quality changes without rebuilding the product.