When a customer asks to change their shipping address, email, or phone number, an AI agent identifies the customer, confirms the new detail back to them, writes the change to the back-end system, and confirms the change is done. This is the first step on the ladder where a record actually changes.
A customer realises, an hour after ordering, that they put the wrong shipping address. They message support. Before AI, an agent looked up the order, asked the customer to confirm the new address, typed it into the order system, and replied to say it was changed. AI now does the same lookup, the same confirmation, the same write, and the same reply.
How this used to be a decision tree
The customer's path was a forced script. The bot said "I can help you change your address. What's your new street?" The customer typed it. "City?" "Postal code?" "Country?" Each field was its own branch. If the customer wrote everything at once ("change it to 23 Rue Lafayette, 75009 Paris"), the bot ignored the structure and asked for the street again. The tree assumed the customer would deliver information in the order the system wanted it.
Why AI doesn't make this a decision tree anymore
An AI agent reads the customer's full sentence, extracts the new address structure regardless of how the customer phrased it, reads it back for confirmation, and writes the change. The customer's input shape is no longer constrained by the system's field order. The shift, like the other rungs, is from instruction-following to goal-following.
What people in the field are saying
Debmalya Biswas's "Agentic AI for Customer Service Desk" calls this out: in the old chatbot world, the script controlled the order of inputs; in the agentic world, the customer controls it and the AI rearranges accordingly.
How does AI update an address today?
The customer's request reaches the AI. The AI identifies the customer, reads the new detail back ("you want to change your shipping address to X, is that right?"), waits for a confirmation, writes the change to the back-end system through an API, and tells the customer it is done. The shape of the work is the same as checking a status, with one important difference: a record is now being changed.
What does it take to make this work?
Write access to the back-end system via API. Verification that the customer is who they say they are: for a logged-in shopper this is the session itself; for higher-risk changes additional authentication is needed. A log or audit trail of the change. A rollback path in case the update is wrong.
Where does this go wrong?
The biggest risk is doing the write without proper verification, and the wrong customer's record gets changed. The second risk is silent failure: the write API returns an error, the AI does not catch it, and tells the customer the change was made when it was not. The third is invalid data: a typo'd email or a malformed address that the AI accepts without checking. The fourth is downstream timing: the change is technically applied, but the warehouse pulled the old address an hour ago and the parcel is already going to the wrong place.
Which tools update customer records?
Autonomous-agent tools that can take action are the right tier here.
- Sierra: autonomous resolution including write actions.
- Decagon: actions plus a log of what was done.
- Lorikeet: strict procedure following, well-suited to writes that need a specific order of checks.
- Gorgias: Shopify-native, can change orders pre-shipment.
- Yuma AI: e-commerce focused, similar capability.
- Fin (Intercom): actions through connected systems.
How would I start doing this?
Pick one low-risk write, such as a shipping address change for a logged-in shopper. Add an explicit confirmation step: "you want to change to X, is that right?" Log every change with who made it and when. Watch the first hundred for issues like wrong customer, invalid data, or silent API failures, before adding higher-risk fields like email or phone, and certainly before anything touching money.
Next on the ladder: the AI takes an action that involves money. Processing a refund.