Generative AI is dismantling one of software's most durable economic assumptions: that once a product is built, serving an additional customer costs almost nothing.

Traditional software often demanded heavy upfront investment, but developers could distribute the finished product at very low marginal cost. Generative AI works differently. Every prompt, every generated answer, and every agentic task must be produced—and paid for—anew.

Earlier this year, Stripe co-founder Patrick Collison argued that in the AI era, "software should be like pizza," made to order at the moment of use. Made-to-order software, however, comes with made-to-order costs.

The internet economy was largely designed around human intent: people searched, clicked, subscribed, and checked out, while businesses optimized around that behavior—acquiring customers, converting them into subscribers, and serving them cheaply at scale. AI agents now research, write code, call APIs, execute tasks, and even make payments on behalf of people or businesses. That promises speed and scale, but automated activity can push consumption—and costs—far beyond what businesses expect or are accustomed to managing.

Everyone is watching AI inflate compute bills; far fewer are paying attention to who is freeloading. Free-trial and multi-account abuse are the quiet killers of AI unit economics, and companies often discover the damage only after their margins are gone. The challenge is no longer simply making compute cheaper, but ensuring that usage translates into revenue rather than allowing unexpected or abusive consumption to erode margins.

Asia illustrates how quickly the shift is moving from experimentation to action. A 2025 Stripe survey found that 82% of business leaders were already implementing, or planning to implement, agentic AI, and half of respondents expect a larger share of sales to come through AI-driven channels by 2030. According to Bain, 85% of Southeast Asian shoppers are already using, or considering, AI tools to guide their shopping decisions.

Tokens Become the New Unit of Commerce

In the AI era, two different users can generate vastly different workloads, and every interaction carries a real cost. Tokens, the small chunks of text into which AI models break down language, have become the key unit for measuring and metering consumption, pushing AI companies toward usage-based or hybrid pricing—models that better reflect the cost of consumption and allow companies to track usage and collect payments in real time.

Lovable, an AI software creation platform that uses Stripe, illustrates the transition. The company began with subscriptions but has since shifted to usage-based billing: once customers exceed the free allowance in their plan, they are charged according to their AI token consumption.

That model introduces a new vulnerability: token theft. Abusers game the system by creating multiple accounts, exploiting free trials, and consuming AI tokens with no intention of ever paying. Every token a freeloader consumes creates an immediate operating cost the company will never recoup.

Stripe research found that more than one in six sign-ups at AI companies are linked to multi-account abuse. At that ratio, abuse scales with growth: faster an AI company adds users, the more freeloading accounts arrive alongside legitimate ones.

Timing matters. AI companies cannot afford to act only when a payment request fails; by then, the user has already moved on. Instead, abuse must be detected at sign-up, before the first token is ever consumed. Stripe, through solutions like Stripe Radar—a fraud prevention AI tool trained on data from millions of businesses worldwide—prevented $1.3 billion in fraud in Singapore last year.

Protecting Margins

The challenge is not simply detecting token theft or fraud, but deciding how to respond. Businesses need greater visibility into where their margins are exposed, along with the flexibility to set their own risk tolerance and tailor their response to different behaviors. Someone repeatedly creating new accounts to steal free AI tokens, for example, should be treated differently from a paying casual user whose usage suddenly spikes.

The indicators to watch are concrete: the share of new sign-ups flagged for multi-account abuse, the pace at which free allowances are drained, and whether abuse is caught at sign-up or only when a payment fails. As AI changes the cost of serving customers, protecting margins will depend not only on cheaper compute, but on making sure every token consumed counts.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune. This story was originally featured on Fortune.com.