ToC Agent Subscriptions Are a Hard Business

Consumer AI agent subscriptions look attractive until you think about the cost curve.

Traditional SaaS has a beautiful assumption: once the software is built, the marginal cost of serving another user is low.

AI agents break that assumption.

The better the agent works, the more it may need to think, call tools, search, generate, retry, and run longer sessions. The best user is not always cheap. The power user may be the expensive one.

The value curve and cost curve move together

For many AI products, value comes from usage.

More conversations. More tasks. More context. More generated output. More tool calls.

That means the user’s perceived value and the provider’s compute cost can rise together.

This is very different from a notes app or a static project management tool. In a normal SaaS product, a heavy user may create some storage and database load. In an agent product, a heavy user can burn real model cost every day.

A flat consumer subscription has to survive that.

That is hard.

Model providers keep moving downward

The other problem is feature absorption.

If your consumer agent is mostly a wrapper around generic model capability, the model provider can eat the feature.

Today’s custom workflow becomes tomorrow’s default mode.

Today’s clever prompt becomes tomorrow’s system behavior.

Today’s agent product becomes a tab inside ChatGPT, Claude, Gemini, or a phone OS.

This does not mean independent products cannot win.

It means the product needs something the model provider does not automatically get.

Data moat is not a magic answer

People often say the answer is data.

Maybe.

But consumer data is messy. It is hard to collect, hard to permission, hard to clean, and often not unique enough to defend the product.

Another answer is vertical depth.

That is more believable. A product that deeply understands one workflow, one user type, one compliance boundary, or one output loop has a better chance than a generic agent.

But even that is not a noun you can own. It is an ongoing motion.

The moat is not “I have a workflow.”

The moat is “I keep learning the workflow faster than the platform absorbs it.”

My current conclusion

I do not think ToC agent subscriptions are impossible.

I think they are much harder than the demos suggest.

A durable product probably needs one of three things:

• a workflow where users bring valuable private context;
• a loop where output quality improves from repeated use;
• a distribution or community advantage the model provider does not care to copy.

Without that, the product is stuck between rising compute cost and shrinking differentiation.

That is why I am skeptical of generic consumer agents.

Not because agents are useless.

Because the business model has to survive the agent being used.