The demo works because nobody gave it system access.

The claim

Most impressive demos run in a sandbox: curated data, a friendly interface, and a person quietly carrying outputs into the real system. The demo proves the model can talk about the work. It says nothing about whether the system can do the work.

How it fails

Production value lives in the integration layer: reading from systems of record, writing back without a human copying and pasting, holding permissions correctly, surviving latency and rate limits. None of that is visible in a demo. Teams postpone integration because it is slow and political, and each postponed quarter makes the demo more polished and less true.

Warning signs

• The demo runs on a snapshot export, not a live system.
• Every end-to-end run has a person in the middle carrying data between systems.
• "Integration" has been next quarter's work for more than one quarter.
• Nobody has written down the list of systems the agent must read from and write to, with the permissions each requires.

What it costs

The demo earns continued funding on evidence that cannot survive production. The integration bill arrives at the worst moment - after the executive team has already promised the outcome.

What would change the answer

One thin slice running against real systems of record with real permissions, however narrow. A pilot that does one real task end to end beats a demo that pretends to do fifty.

Where this maps

This failure is a Reach gate failure. The PROOF Scorecard tests this gate in about four minutes; the full review proves it with evidence.