Buying an AI agent: what 'done for you' must actually include

The demo is the easy part

In 2026 a persuasive agent demo takes an afternoon: connect a model, paste your price list, ask it something friendly. This is why the market is full of them. What you are actually buying, if you are buying well, is the unglamorous ninety percent around the model — the grounding, the integrations, the guardrails, the monitoring — because that is the difference between a toy that impresses in a meeting and a system you can leave alone with customers.

Grounding: the agent should know your business, not the internet

An agent answering from general knowledge will eventually invent a price, a policy or a promise, and it only has to do it once to cost you a customer. Done-for-you means your documents, catalogue and policies are indexed so every answer traces to something you actually wrote — and when the answer is not there, the agent says so and hands over to a human. Ask any vendor one question: where does an answer come from, and what happens when it cannot be found? The quality of the reply tells you everything.

Integrations: conversation without action is theatre

A support agent that cannot read the order status is a FAQ with manners. The value appears when the agent acts — checks stock, books the appointment, issues the invoice, updates the CRM — under explicit permissions that define what it may do alone and what needs sign-off. Scoping these permissions is where real projects spend their time, and where quotes that skipped the thinking fall apart in the second week.

The paper trail and the safety rails

Every conversation logged, every action attributable, escalation rules that hand hard cases to people, and a tone that stays yours in every language you serve — these are not enterprise luxuries. They are the reason you can put an agent in front of customers without holding your breath. Insist on seeing the escalation flow and the logs in the demo, not the confetti.

What a serious engagement looks like

A real done-for-you build maps the workflow first, ships a narrow agent that does one job completely, measures it against the humans doing the same job, then widens scope with evidence. Weeks, not quarters; one process, not 'everything'; a monthly loop where transcripts are reviewed and the agent gets measurably better. Buy that, and you have hired an employee who never sleeps. Buy a demo, and you have bought a demo.