AI for founders: the 'is this a product?' filter

A founder we know pitched us their idea. A B2B AI assistant that summarized customer-support tickets. We liked them. We didn't think they had a company.

The idea was a feature. Zendesk would build it. Intercom would build it. The CRM vendors would build it. Standalone, the wedge was too narrow and the moat was nonexistent.

The founder's job is the filter: which AI ideas are products, and which are features that will get absorbed.

The filter

Three questions to apply to any AI idea:

1. Could a 30-person engineering team at a $500M company build this in a quarter?

If yes, it's a feature. Someone with distribution will build it once they notice. You're racing the clock.

2. Does the value compound on customer-specific data?

A product that gets better the more a customer uses it has a moat. A product that's only as good as the underlying model has none.

3. Is the customer's switching cost something other than "the model"?

If your customer would switch to a marginally better model the moment one appeared, you have no moat. If switching costs them workflow rework, training rework, integration rework, you have a moat.

A product passes if it's a yes-no-yes or a yes-yes-yes. The pure-no answers, however clever the idea, are features.

What's a feature today

Wrappers around frontier models with a nicer UI.
Domain-specific chatbots without proprietary data or workflow.
"AI for X department" where X is well-served by Microsoft/Google/Salesforce.
Generic content generation, even with brand-voice fine-tuning.
Document Q&A wrappers without specific compliance or domain expertise.

These all ship. They all also get out-competed in 18-36 months by an incumbent's feature release.

What's a product

AI tooling that integrates deeply with workflow software the incumbent doesn't make.
Compliance-laden domain AI (healthcare, legal, finance) where the regulatory work is itself the moat.
AI on customer-specific data with strong privacy/sovereignty positioning.
Tooling for AI engineers themselves (evals, observability, routing).
AI applied in industries the major incumbents don't serve well (manufacturing, construction, agriculture, government, etc.).

These have moats. They take longer to build. They're also less dependent on model improvements for their value.

The model-leverage question

A new question that defines this era: are you betting on the model getting better, or the workflow getting deeper?

Model-bet companies depend on frontier improvements. When a new model drops, their product gets better automatically. They also have less to defend. Most consumer AI products are model-bets.

Workflow-bet companies depend on integration and domain depth. New models help but don't transform their advantage. They have more to defend. Most B2B vertical AI is a workflow-bet.

Pick one deliberately. Mixed bets confuse customers and investors alike.

The 'AI-enabled' framing

Increasingly we recommend founders avoid "AI-first" positioning. "AI-enabled software" or "we put AI to work" is more honest about what the product is: software that solves a real problem, with AI as an ingredient.

The reason: "AI-first" companies have to justify their existence every time a new model drops. "AI-enabled" companies have to justify their existence every time they fail to solve the customer's problem. The second conversation is easier.

The investor pattern

VCs in late 2025 have shifted hard. The "AI" prefix that earned a 2x multiple in 2023 is no longer the differentiator. What earns now:

Specific industry expertise.
Defensible distribution.
Real revenue (not pilots).
A model-neutral architecture.
Clear path to gross margin > 70%.

Founders pitching "we're using GPT-4 to do X" without a moat are getting harder passes than they did 18 months ago. The market sorted.

The build-or-wrap question

Three patterns:

Pure wrapper. Your product is a prompt + a UI. Easy to ship; trivial to copy. Don't raise on this.
Wrapper plus integration. Your product is the integration into a workflow, with AI as one component. Harder to copy. Raisable.
Custom model + workflow. Your product is a domain-specific model on proprietary data. Hardest to copy. Most defensible.

The build-vs-wrap question depends on what you're building. The defensible answer is usually "wrap the model, build the workflow."

What founders get wrong

Building the AI before the customer. Easy to fall in love with model output. Customers care about outcomes.
Pricing on tokens. Customers don't think in tokens. Price on outcomes.
Burning runway on model improvements. Your competitive position rarely depends on the next model release.
Hiring AI engineers before product engineers. The product engineer who learns AI compounds; the AI engineer without product sense often doesn't.

Close

Founders are the filter for which AI ideas are companies. Not every idea is. The honest "this is a feature" conversation is the most valuable conversation a founder can have early, before the round, before the team, before the customers expect you to last.