AI agents aren’t human customers, and that breaks everything we know about how markets behave. They don’t have purchasing habits, don’t experience friction and don’t develop brand loyalty. After spending years as a quant at the global hedge fund Citadel, I’m cursed to always think in finance terms, but it’s helped me spot the patterns hidden in plain sight.

What I’m seeing is that AI agents are now our customers, but we are still measuring them like they are people. We are applying human metrics — customer acquisition cost (CAC), lifetime value and retention rates — to entities that don’t persist long enough for those metrics to mean anything.

We need new ways to measure penetration and new criteria for what makes customer relationships last.

Why agent customers break the model

For decades, venture-backed companies paid huge sums to bootstrap network effects because customers stuck around. After Uber launched, rides were heavily discounted. They operated at a loss, betting that customers would get hooked and stay. Now it’s hard for any other ride-share to threaten Uber’s incumbency, though many have tried. That upfront spending built something that has endured.

AI agents break the durability assumption. Unlike people whose loyalty can be earned, they appear, transact and disappear.

You can pay a ton of money to attract AI agents to use your product, but there’s no moat being built around it. Those relationships are fragile because agents don’t accumulate loyalty, and each interaction resets the market. With every transaction costing as much as the first one, the new model is far less cost-efficient than the prior CAC regime.

Learn to count differently

Tools like Cursor and Claude Code now generate entire applications on demand. They automatically select which services to integrate into new apps based on patterns in their training data. For example, they might choose Google login over Facebook, Stripe for payments or Salesforce for CRM. Coding agents are just one class; others will emerge in sales, procurement and consumer transactions, each shaping markets differently.

With developers no longer making these purchasing decisions, the real customer has become the large language model. The shift is already widespread. Stack Overflow’s 2024 survey found that 62% of professional developers were using AI tools in their development process, with 82% of those using them to write code.

To know whether your product is winning with AI agents, stop counting “users” as if they are your old human customers. Instead:

• Assess what percentage of code bases that need authentication and implement your login solution.
• Measure how often your API gets selected for apps that need payments.

This works like Nielsen ratings for television, which samples households to see what people are watching at a given time. Measure how often your API gets selected for apps that need payments when the selection is made by an agent. You can think of this as an Agent Penetration Rate (APR): the percentage of agent-driven choices for a given component (e.g., Google login vs. Facebook login).

When transitioning to agent customers, compare how agents select your product versus how humans do. You could also consider stopping spending on acquisition and see what happens. With humans, some stick around. With agents, growth stops immediately. That will tell you all you need to know.

This is a market inefficiency issue as well as a product-strategy problem. Businesses are allocating capital toward customers that no longer behave like capital assets.

Investment implications

Venture-backed startups should be much more skeptical about paying upfront amounts to acquire users because, with AI agents, you are paying for transactions that evaporate. What matters now is identifying where the customer acquisition cost doesn’t go to zero.

Vertically integrated software is one such case. When there aren’t many customers between you and the end user (and that end user is an old bureaucratic business less likely to churn), those relationships prove more valuable. For private equity firms looking to acquire companies, prioritize vertically integrated software where customer relationships are durable.

Entertainment and the service industry are two other obvious exceptions where the primary consumer remains human. Some argue that even in agent-driven markets, network effects persist. If you dominate AI-generated code for long enough, your code shows up more in GitHub repositories, and future LLMs trained on that data will be more likely to select your frameworks.

The caveat in this scenario is that the network effect is far less cost-efficient than the previous version.

Investors evaluating deals need to identify who the actual end customer is. If it’s a blink-and-you-miss-it AI agent making decisions, the old assumptions don’t hold. If it’s a human or a durable institutional relationship, they might.

Humans are the ghost in the machine

We have spent decades building businesses around the concept of a “user” as a stable entity you could identify, track and retain as discrete Customers A, B and C. Each has its own lifetime value, acquisition cost and probability of sticking around.

Elsewhere, I have called the waning of this category (and eventual death) the “twilight of the user,” and the implications go beyond fixing a few metrics. Customers that pop up only to disappear negate the business logic of the very category we have used to organize our business. You cannot calculate lifetime value for something that doesn’t have a lifetime, or measure retention for something that has no memory of you.

The AI agents making purchasing decisions today in generated code are already here. Many businesses may not have noticed yet because the old metrics still produce numbers. Recognize this shift now or your next fundraise will assume customers that no longer exist.