Guardrails that stop promises your policy cannot keep.

The worst thing an AI support agent can do isn't a slow answer. It is a confident promise nobody can keep: "Your refund has been processed" when it hasn't, "We'll send a replacement tomorrow" when there is no stock, or a friendly pretence of being a person. Each one becomes a second ticket, an angry customer and a teammate cleaning up.

Better prompts alone don't prevent that. What does is a set of checks between the AI and the customer, and an agent built to say "I don't know" and step aside. This guide covers the guardrails Aide runs on every reply, and the few settings worth choosing yourself.

The check before every reply

Aide writes each reply from your knowledge, the customer's orders and the scenarios that match the conversation. Then, before anything is sent, the reply is verified. In the chat, the shopper sees "Double-checking the answer" while that happens.

Verification rejects a reply that:

• promises a refund, credit or action the topic's AI level does not allow
• contains a placeholder, such as an unfilled blank
• claims to be a person

A rejected reply is revised or handed to your team. When a draft says something your knowledge or the order's records contradict, it is held back, and the handoff note tells your teammate so instead of presenting it as fact.

Good grounding means fewer replies get that far. Aide reads the product pages an article names for live stock, sizes and price. It checks where an order ships before suggesting a return process, and never promises a restock on a style tagged discontinued. It keeps to what a teammate already told the customer.

Refunds and order changes

A promise only does damage if the AI can act on it, or claims it did. Aide's order actions have their own gates:

• The shopper proves the order is theirs first, by logging in or with the order number and checkout email.
• Aide's refund limit starts at zero, so every refund it offers waits for a teammate until you raise it.
• On a topic set to Drafts, a change the shopper confirms goes to your team, and Aide tells the shopper a teammate will review it.
• Aide says exactly what it will do and waits for a yes. If the shopper moves on, the offer lapses.

Here is what that sounds like on a refund:

"Sorry the jacket didn't work out. I can refund the $64.00 you paid for it on order #1187 to your original payment method. Shall I send the request to our team?"

If Shopify declines a change, Aide tells the customer. A retried reply never refunds or cancels twice. Your team decides per topic how far Aide goes, as the ticket-type guide explains.

Saying "I don't know"

When the answer isn't in your knowledge, Aide says so and offers your team, instead of guessing or handing off without a word. Asked for a sizing chart your help center doesn't have, it says it doesn't have that handy and offers a teammate. For an order with no tracking number yet, it shares what it knows and says the tracking number isn't available.

If Aide can't write a reply it is confident in, the conversation goes to your team. Both behaviours follow your handoff settings.

Honest about being an AI

When a customer asks whether they are talking to a person, Aide answers in English with a fixed line that no setting or instruction can change. For an agent named Ava at a store called Juniper, it reads: "I'm Ava, an AI agent for Juniper. I can bring in a person from our team any time you like." In other languages, Aide says the same thing in its own words. A reply claiming to be human is rejected, whatever the agent's name or instructions say.

A way out that always works

A customer who wants a person should get one without arguing with a bot.

• "Talk to a person" in the widget hands off at once, with your handoff message and no AI reply first. A short message such as "agent please" does the same.
• "Customer asked for a person" and "Policy requires a person" always hand off, whatever else your settings say.
• On email, Aide stops replying once a teammate is assigned. A thread that gets three AI replies within an hour goes to your team with an internal note.

The same holds when the AI service itself fails. If the service behind Aide's replies is down or overloaded, Aide notices within a few failed requests and sends new conversations to your team within seconds, instead of making customers wait through retries. It tries one request about every 30 seconds and resumes on its own once one succeeds. Conversations handed off during the outage stay with your team.

Testing and measuring

You can see what Aide would say before any customer does:

1. Try the agent from its page. The preview states what it is answering from.
2. Open a scenario and preview Aide's draft on real conversations from the last 28 days.
3. On a connected helpdesk, compare Aide's drafts on solved tickets with the reply your team sent.
4. Run a helpdesk inbox in Watch only, where Aide decides and drafts but writes nothing, before widening it. The helpdesk guide covers the stages.

Aide's own team also grades replies to order questions against a written definition of an acceptable reply, per request type: where is my order, address change, cancel, refund, return and shipping policy. There is no partial credit. One wrong fact fails a reply, and so does an unnecessary handoff, so the bar can't be met by passing everything to a person. An AI judge grades first, a person checks every fail and a share of passes, and the person's verdict is final. Every confirmed failure becomes a test that future changes must pass, and Aide publishes no accuracy figure until the bar has been met on enough real tickets to mean something.