What AI SaaS actually means (and the two senses people mix up)

AI SaaS, defined

AI SaaS: AI SaaS is cloud-delivered software in which a machine-learning model performs the product's core function, which makes the vendor's cost of goods rise with usage instead of staying near zero.

The delivery half of the term is old and settled. NIST SP 800-145, September 2011: the consumer uses a provider's applications running on the provider's cloud infrastructure, with no control over the operating system, storage, or hardware underneath. That definition holds unchanged for AI SaaS. Nothing about the hosting model is new.

What changed is what sits inside the request. The compound "AI SaaS" only entered commercial use once inference got cheap enough to call on every request — OpenAI shipped the ChatGPT and Whisper APIs on 1 March 2023 and priced gpt-3.5-turbo at $0.002 per 1,000 tokens, ten times cheaper than the text-davinci-003 model it replaced. At that price you can put a model call in a product's hot path and still have a business. Before that you could not.

The two senses, labelled

People use "AI SaaS" for two different businesses. The pricing page tells you which one you are looking at.

	Sense A — AI-enhanced SaaS	Sense B — service-as-software
What you buy	Access to an application	Completed units of work
Unit of pricing	Seat per month	Resolution, action, run, credit
Who does the work	A human, faster	The vendor's agent, end to end
Vendor gross margin	Roughly flat with usage	Falls as usage rises, unless inference cost falls
Remove the model	Product still works, less pleasantly	Product produces nothing

Sense A is most of what gets marketed as AI SaaS: a seat-priced application with summarisation, drafting, or search bolted on. Sense B is the one that changes the P&L. Zendesk sells per-agent seats and also sells outcome-based AI resolution; Intercom prices Fin per resolution; Salesforce meters Agentforce in consumption credits per action rather than per seat. Three vendors, two pricing philosophies, one label stretched across both.

If someone says "AI SaaS" without saying which, ask what the invoice counts.

What it is not: AIaaS, MLaaS, and infrastructure

The term people most often confuse this with is AI as a Service (AIaaS), also written MLaaS. That is model and training capacity sold as a utility: Amazon SageMaker priced per instance-hour, the OpenAI and Anthropic APIs priced per million input and output tokens. Those are inputs. An AI SaaS product buys them the way a 2014 SaaS company bought EC2.

The test is what the customer is accountable for. Buy AIaaS and you own the prompt, the evaluation, the retries, the failure. Buy AI SaaS and the vendor owns all four, which is why the vendor can charge per outcome and why they now carry a variable cost you can read off a public pricing page.

It is also not the same thing as agentic workflow inside a marketing or ops team. That is a deployment pattern, and it breaks in specific, documented ways — the failure modes are catalogued in what agentic marketing actually is, and where it breaks.

Worked example: support, priced per resolution

Intercom sells Fin at $0.99 per resolution. Take 10,000 resolutions in a month: $9,900 of revenue, and no seats sold.

The cost side is what makes this a different business from the seat model. The vendor pays a model provider per token on every conversation the agent attempts, including the ones it fails to resolve and hands to a human. Attempts always exceed resolutions. So the unresolved conversations are pure cost against revenue that never arrives, and each resolution that does land carries a retrieval step, one or more model calls, and a re-check before the answer ships.

Compare the same deflection sold as seats. Zendesk's per-agent-per-month tiers do not move when a customer's ticket volume doubles overnight; the contract shape protects the margin. Under per-resolution pricing, a customer's bad product launch lands directly on the vendor's inference bill. Revenue and cost of goods are wired to the same meter.

Which is why cost caps stop being an engineering nicety and become a pricing prerequisite. Unbounded loops are the most common production failure in agent systems, and the fix is a step and spend cap enforced outside the model — not a sentence in the prompt asking it to be efficient. Prompt-level pleading fails the first time a tool returns something the model did not expect. The mechanics of enforcing those caps are set out in your first agent, with hard caps.

How to tell whether something actually qualifies

Four checks, applied in this order:

Read the invoice line. If every line item is a seat, the model is a feature. Metered actions, resolutions, or credits mean the vendor has accepted variable cost.
Ask what the model owns. If the output is a suggestion a human must retype, a human is still doing the work. If the output is the deliverable that ships to a customer, the model is doing it.
Ask for the eval and the rollback. Products whose core function is probabilistic need a scored test set, a version they can pin, and a path back to the previous behaviour. Vendors who cannot name theirs are selling Sense A in Sense B language. In March 2024 the SEC charged two investment advisers over false and misleading statements about their use of AI. AI washing is an enforcement category now, not a rhetorical complaint.
Ask what happens on a hostile input. Prompt injection is LLM01 in the OWASP Top 10 for LLM Applications. A product that takes untrusted text and acts on it without a boundary is not shipping AI SaaS; it is shipping an incident with a subscription attached.

I took Metadata.io from $0 to $15M ARR on the seat-and-platform model, raised $50M against it, and hold 6 patents in AI-driven marketing. That business would fail check 1 and check 2. I am building the opposite now — a company where the agents do the work and I pay their bills — and the honest version of that is in two years of building with AI.

The label matters less than the meter. If nobody's cost moves when the customer uses more, the AI is a feature and the SaaS is the product.