The gap between a demo and a product

We see the same pattern over and over. A company buys an LLM subscription, builds a demo «on our data» — it all works. Six months later the project is dead. Not a single business metric moved.

It's not the model. Models are good enough today. It's the zone nobody wants to own: data, integrations, and SLA.

Data. For an agent to work, it needs fresh data in a usable format. That means ETL, quality control, refresh, a historical layer, access control. Not a week — months. In a demo, this data is hand-prepped. In production, nobody preps it.

Integrations. An agent that can't write into the CRM, send an email, drop a file, — is an LLM chat, not a product. Integrations are 30—50% of the codebase. And that's what breaks in production.

SLA. Agents will make mistakes. The question is who notices and how fast it gets fixed. In a demo, «model mistakes» are a funny screenshot. In production, they're a blown deal, a lost customer, or a regulatory fine. Monitoring, alerts, rollback, human-in-the-loop — they're part of the product, not «we'll add it later».

When we kick off a project, the first conversation isn't about the model. It's about what data you have, which systems hold it, how to extract it, who owns its quality, what happens at 2 a.m. if the agent stops responding. If you hear questions like these — you're working with product engineers. If only prompts — you're working with a demo.