AI SaaS: The Brutal Math

The easiest time to launch, the hardest time to profit

Today, people who were non-technical months ago can ship AI apps and even raise funding. That creates a misleading signal: shipping is cheap, distribution is doable, and demos look magical. But when you strip away pitch decks and VC oxygen, the underlying question becomes simple: does the business work when you have to pay your own bills? For most GenAI apps, the answer is uncomfortable.

Classic SaaS margins vs AI SaaS margins

Traditional B2B SaaS typically operates at 70–90% gross margins, with best-in-class companies around 80%+. The reason is structural: after the platform is built, the marginal cost to serve the next customer is close to zero.

AI-native SaaS breaks that assumption. In many GenAI-native products (especially “LLM wrappers”), gross margins are often closer to 30–60%, with the upper end representing exceptional execution. Even mature AI products can sit in the mid-50s range, which is “best-in-class” by AI standards, but still far below classic SaaS benchmarks.

This is not a temporary dip. It is what happens when your product’s core value is delivered through ongoing compute.

The hidden crutch: “dead subscriptions”

Some GenAI tools show healthier margins because of a phenomenon that does not last: dead subscriptions.

A dead subscription is when users pay but barely use the product. Bundled “AI tool packs” and hype-driven purchases create short-term revenue that inflates the appearance of product strength. But dead subs are not durable demand. Once budgets tighten and novelty fades, usage and renewals expose the real economics.

If your margin profile relies on customers not using what they bought, it is not product strength. It is temporary inefficiency.

Why GenAI unit economics break

Classic SaaS has high fixed cost upfront (R&D, product, platform), then low variable cost per additional customer.

GenAI flips this. Costs recur and scale with usage:

• API calls and inference compute per action

• Licensing costs (sometimes per output)

• Moderation and safety overhead

• Higher “support burden” because outputs must be validated

• Usage spikes from heavy users that can dominate cost structure

This is why many AI products cap usage. Not because founders enjoy limiting customers, but because unmetered usage can make costs explode.

The “all-you-can-eat” pricing trap

Users love flat-rate, unlimited plans. AI companies fear them.

With fixed-rate, unmetered pricing, a small minority of power users can generate disproportionate inference costs. That forces companies to introduce:

• Rate limits

• Usage caps

• Feature gating

• Tiered quotas

But once customers hit limits while paying, they often feel cheated and churn. So AI apps face a tough tradeoff: protect margins and risk churn, or protect experience and risk losing money per user.

A reality check using platform-scale AI

Large-scale AI products expose the same dynamic more dramatically: huge adoption does not guarantee healthy conversion or profitability.

A common pattern appears:

• Enormous free user base

• Low conversion to paid

• Significant ongoing compute costs even for non-paying users

• The need to meter usage even at premium tiers

This is why traditional SaaS metrics do not translate cleanly to GenAI-native products. The economics are different, and the path to profitability is narrower.

AI price wars make it worse

As models commoditize and competition increases, GenAI software faces price pressure. Tools layered on top of foundation models are especially vulnerable because their differentiation is thin and their core costs are tied to model pricing and compute.

In response, companies try:

• Expanding product lines to upsell

• Bundling to improve conversion

• Metering to control costs

• Impact-based pricing in B2B

Everyone is experimenting because no dominant template has emerged for most categories of AI apps.

What a sustainable AI business tends to look like

A practical litmus test emerges:

Prefer “SaaS with AI features” over “AI-first wrappers.”

A durable product solves a real problem even without AI. AI then makes the workflow faster, cheaper, or more accurate, but it is not the sole reason the product exists.

Pure GenAI-native businesses can still work, but the successful ones often share traits:

• Focus on high-volume text and document workflows (legal, HR, accounting, sales ops)

• Clear, repeatable jobs-to-be-done

• Integration into existing systems (CRM, document stores, ticketing)

• Pricing aligned to business impact, not generic subscriptions

• Strong controls to keep inference costs predictable

These may not always be unicorn-scale. But they can be profitable, stable businesses.

The real goldmine: boring, hard enterprise problems

The biggest opportunities are rarely “sexy.” They live in legacy enterprise workflows where the pain is real, budgets exist, and switching costs are high.

In industries like accounting, legal, and pharma, the hardest work is not sending documents back and forth. It is analytical, consultative, negotiation-heavy work where correctness, context, and judgment matter.

If you can automate genuinely hard professional work (not just summarize it), you build something that outlives hype cycles. That is where durable value tends to concentrate.

Bottom line

It is easy to get users to try an AI app with bold promises and fear-of-missing-out marketing. It can even be easy to get early payments. What is difficult is delivering value that keeps people returning.

The two metrics that outlive hype are the same as always:

• Retention

• Conversion

Before building or buying into an AI product, ask one question: Would this solve a real problem without AI? If the answer is no, the business is likely riding a wave, not building a foundation.