Order-taking for an identified user

We've seen that a custom AI agent can know your business (the RAG) and look up your live data (the tools). The most powerful capability remains: acting on the customer's behalf. Let's take the example of order-taking.

First, know who you're talking to

As soon as we're talking about an action that commits an account, one rule becomes essential: you must reliably know who you're talking to. That's the difference between an anonymous visitor and a properly identified user.

An anonymous visitor can ask questions and check availability. But to place an order, the agent must be certain of the person's identity, because this action touches personal information, a delivery address, and potentially a payment.

Identification happens before the agent gets the right to act. In practice, the user is authenticated by your usual system: signing in to their account, a secure session token, or any other mechanism you already use. The agent then inherits a secure context telling it, for example: "you're talking to customer number 4837, Marie Tremblay, authenticated." At no point does the agent rely on the customer's word to establish identity: it comes from a trusted source, not from the conversation.

The order, step by step

1. The identified customer says: "I'd like to order two Aria chairs."
2. The agent first checks availability — as in a previous post — to confirm the requested quantity is in stock.
3. The agent personalizes the exchange using the customer's profile: it knows their saved delivery address, their language, their history. It suggests: "I can have them delivered to your usual address at 120 Main Street. Shall I confirm?"
4. Before finalizing, the agent always recaps: products, quantities, prices, address, timeline. This explicit confirmation is crucial: the agent never commits a transaction without a clear "yes" from the customer.
5. On confirmation, the agent calls a create_order tool — this time to write to your system. The order is created, a tracking number is returned, and the agent shares it with the customer.

Guardrails under your control

• The agent can only act for the account it's tied to: it cannot order on someone else's behalf.
• Every action goes through a tool you've defined, with your own validations (quantity limits, eligible products, etc.).
• Every step is traceable: you know which order was created, for whom, and when.

Know, look up, act

It's this progression — know through RAG, look up through tools, act securely for an identified customer — that turns an impressive but generic technology into a true digital collaborator serving your business.

All articles