Moats for AI companies

AI products often start with a thin moat: the underlying models are available to everyone via APIs, and a clever prompt wrapper is trivial to copy. Durable AI companies accumulate advantage along paths that map back to Helmer's 7 Powers. "Our moat is the model" is usually false unless you train frontier systems few others can afford.

Where advantage actually accumulates

Data flywheels. Product usage generates proprietary interaction data that improves the system in a way rivals cannot buy. Pure "we train on public web data" is not a cornered resource. Labeled outcomes from your workflow — which support tickets resolved, which code changes passed CI, which recommendations converted — can be.

Workflow integration. The model is a feature; the moat is being embedded in how work gets done: tickets, IDEs, CRMs, design files, compliance logs. Switching costs rise when your system holds history, permissions and automations.

Distribution. Owning a channel (a large existing user base, a marketplace of plugins, a sales motion into a vertical) beats a slightly better model with no path to customers. See go-to-market and distribution.

Proprietary evals and trust. In regulated or high-stakes domains, the company that can prove reliability — eval suites, audit trails, human review loops — earns brand and switching costs that a demo cannot match.

Cost structure. If you can serve the same quality at structurally lower inference cost (custom routing, caching, distillation, better batching), you get a scale-economy edge. Rivals on expensive frontier APIs alone will struggle on gross margin.

Competing with incumbents

Incumbents have distribution, brand and often switching costs. Well-funded rivals have time runway. You need a Power they cannot easily match — typically counter-positioning or a niche network they will not prioritize.

Patterns that work:

1. Segment monopoly: win a vertical so completely that the generalist looks wrong.
2. New enabling technology: something that was impossible becomes cheap (LLMs for document workflows) and you re-bundle the job.
3. Superior wedge UX: 10× better on the one job that starts the relationship, then expand via switching costs.
4. Ecosystem / platform: invite others to build on you so the network compounds.

What usually fails: raising a large round to "out-feature" the market leader, or competing only on price without a cost-structure advantage.

You build an AI writing assistant for enterprise security teams. Competing on "better prose" is hopeless. Competing on SIEM connectors, incident-labeled evals, and audit logs creates switching costs and a cornered resource. The horizontal tool stays better at generic writing; you become the default inside that workflow.

As an engineer, prefer work that deepens integration, data feedback loops, or multiplayer value over one-off features that demo well. When a PM asks for parity with a competitor, ask which Power you are building.

Common mistakes

• "Our moat is the model" when you call the same API as everyone else.
• Out-featuring the incumbent on their terrain. Flank instead.
• Competing only on price without a structurally lower cost.

Check yourself

Which data can be a cornered resource?

Labeled outcomes from your workflow that a rival cannot buy at similar cost — not public web text.

How do you beat a horizontal chatbot in a vertical?

Win the workflow: connectors, evals on that domain's history, compliance. Not "better prose."

Key takeaways

• Assume the model commoditizes. Design value into data, workflow and distribution.
• Evals and trust are brand and switching costs in high-stakes domains.
• Lower inference cost is a scale-economy edge.
• Beat incumbents by flanking and counter-positioning, not by cloning their roadmap.

Further learning

7 Powers (Hamilton Helmer)7 Powers – by Hamilton Helmer7powers.com

Back to competition and moats.