Network effects are claimed far more often than they exist.

The AI era will make this worse. Every company with users, data, or shared templates will be tempted to say the network gets stronger as more people join. Usually it does not. More users are nice. They are not automatically a network economy.

A network economy exists when the product becomes more useful to each user as other users join or participate. In AI, the more interesting version is often not a social network. It is a workflow network.

The product gets better because real work runs through it. The record of decisions gets richer. Exceptions get handled in one place. People learn how to coordinate inside the system instead of around it.

The weak version: more data

"More users create more data" is not enough.

The data has to improve the product in a way customers can actually feel. It also has to be hard for competitors to replicate, survive privacy and compliance constraints, and stay tied to the real job instead of being scraped from the edges.

A pile of chat logs is not a moat. A clean record of decisions, exceptions, outcomes, and human corrections inside a specific workflow might be.

The distinction matters because AI can make generic data cheaper. Models already know a lot. Public information, common support patterns, generic sales emails, basic coding examples, and standard operating procedures are increasingly available to everyone.

Unique workflow data is different. It captures how a company actually works.

Workflow networks

A workflow network forms when multiple parties use the system to get work done together.

Think suppliers and buyers, doctors and patients, brokers and clients, developers and reviewers, finance teams and budget owners, recruiters and hiring managers. The value is not just that people are present. The value is that the system becomes the place where requests, approvals, exceptions, and institutional knowledge move.

AI can strengthen that loop.

It can summarize history, route tasks, draft responses, detect anomalies, suggest next steps, and learn from corrections. If more real work moves through the network, the system gains better context. If that context genuinely improves the product, more work stays there.

That is a real loop, but only if you can see it in behavior.

Beware collaboration theater

Many products confuse collaboration features with network effects.

Comments, shared dashboards, team invites, and Slack notifications are not enough. They may improve the product, but they do not necessarily create power. If customers can export the data and coordinate somewhere else without losing much, there is no strong network economy.

The hard test is whether leaving breaks the workflow for other participants.

If one user churns, do others lose value? If one company leaves, does the ecosystem weaken? If the data stops flowing, does the AI get meaningfully worse? If a competitor starts from zero, how long does it take to reach acceptable quality?

Those are uncomfortable questions. Good. Strategy should make people sweat a little.

AI agents and network density

Agents may make network economies more important, not less.

When software starts acting on behalf of users, context and permission become valuable. An agent that understands who approves what, which supplier is reliable, how exceptions are handled, and what tone works with a given client is more useful than a generic agent pointed at a blank inbox.

But agents can also weaken networks by making switching easier. If an AI assistant can migrate data, rebuild workflows, and train users quickly, the old friction declines.

So the network has to be alive. Static data is vulnerable. Active coordination is harder to copy.

Operator test

To test for network economies in an AI product, ask:

  • What gets better for one user when another user participates?
  • Is the improvement visible in retention, usage depth, or willingness to pay?
  • Does the AI improve from workflow-specific data that competitors cannot easily access?
  • Are permissions and trust part of the network, or just raw data?
  • If a customer leaves, who else feels the loss?

The strongest AI networks may not look like classic networks at all. They may look more like operating systems for messy human work.

That is where the power may actually sit.

This is part 3 of 10 in Seven Powers in the AI Era.