AI product moats

Foundation models are widely available. A thin chat wrapper around a public API is easy to copy. As of when this was written, models improve quickly and prices fall in fits. Strategy that bets only on "our prompt is clever" expires. Strategy that bets on workflow, data, distribution and eval discipline compounds.

When a customer asks, "Why wouldn't I just use ChatGPT?" your answer must be about permissions, audit logs, the system of record, offline eval quality on their documents, and the buttons next to the draft — not about your system prompt.

The big idea

Durable advantage usually sits in layers practitioners keep repeating:

• Workflow — you own the system of record UI where the job already happens.
• Data flywheel — consented traces, corrections and evals that improve your quality bar.
• Integrations — permissions, connectors, enterprise SSO, audit logs.
• Distribution — existing customers, marketplace, sales motion.
• Trust and brand — reliability, security review, clear failure modes.

See competition and moats. Distribution is underrated: an AI feature inside an already-loved tool (IDE, CRM, design suite) inherits trust and habit. A standalone AI tab fights every other chat box on the internet.

Walking through the one-sentence test

A support-draft feature inside the existing ticket UI: the agent already has the ticket, the policy corpus, SSO, an audit log and a human approve button. The customer stays in the system of record. That is a workflow plus data plus trust answer.

A standalone chat that pastes a ticket in and pastes a reply out: the customer can do that in any chat box tomorrow. That is wrapper risk.

Do not hard-code today's model failures as permanent UX debt, and do not assume magic next quarter. Design so a better model is a config plus an eval re-run. Keep artifacts as the source of truth. Avoid promising capabilities you can only deliver with a hypothetical future model — ship the copilot now, gate the agent behind quality bars.

Common mistakes

• Competing on the system prompt. Anyone can copy it.
• Training a foundation model as the first move. Almost no product team should.
• Ignoring distribution. A better model in an unused tab still loses.

Check yourself

A customer asks why not paste tickets into ChatGPT. What do you answer?

Permissions, audit logs, policy evals on their docs, and the approve button in their ticket UI.

API plus workflow plus an eval loop, or train your own foundation model?

API + workflow + eval/data loop is the default. Training a foundation model is rarely the product bet.

Key takeaways

• Defend with workflow, data, integrations and distribution.
• Wrapper risk shows up as "why not ChatGPT?" — answer in one sentence.
• Ride model improvements through eval harnesses, not prompt lock-in.

Further learning

The New Business of AI (a16z)At a technical level, artificial intelligence seems to be the future of software. AI is showing remarkable progress on a range of difficult computer science problems, and the job of software developers – who now…Andreessen Horowitz

Back to the map in AI product strategy.