Presales demos that close: build a narrow working slice, not a click-through

Presales teams selling AI have a specific problem: the prospect has already seen a chat window. A scripted click-through now reads as evasion, and a live open-ended demo fails in front of the room when someone types the question you did not anticipate. The way out is narrowness.

What a demo build contains

• —One workflow that reaches an outcome — a document posted, a claim triaged, a quote drafted — not a general assistant.
• —Real inference on seeded data that looks like the prospect's: their document types, their field names, their edge cases.
• —Guardrails: refusal below a confidence threshold, enforced citations, a hard scope boundary that fails gracefully.
• —A demo narrative: three questions to ask, one deliberate failure to show, one number to quote.
• —A hosted environment with one-click reset between meetings, so the tenth demo behaves like the first.
• —An honest boundary slide: what this slice does not do yet, and what a pilot would add.

Showing a failure on purpose

The most effective five seconds in an AI demo is a refusal. When the system declines a question it cannot ground and says why, technical buyers relax — they have been in the room where a confident wrong answer went to a customer. Demonstrating the guardrail sells the engineering better than another correct answer does.

Why it should not be throwaway

Built this way, the demo budget is a down payment on the pilot rather than a marketing expense. If the deals land, you already have the skeleton of the system; if they do not, you have learned which objection killed them and it is written down.

Never demo on a prospect's live production data because it is faster. Seed synthetic data that mirrors their structure, and you avoid a security review in the middle of a sales cycle.