Fundamentals
What does read-only AI mean, and why use it?
By the RoutMe team. Last updated .
Read-only AI is an assistant that can look up and show information but has no ability to change anything in your systems. For customer support, it means the assistant can check an order or a policy, yet cannot refund, edit or cancel, so a wrong or manipulated answer cannot become a wrong action.
Why it matters for a public chat
A chat widget on a storefront talks to anyone, including people who try to push the assistant off script. A prompt can be written to ask a model to ignore its instructions. If the assistant holds a refund tool, that request becomes a real risk. If it holds only lookup tools, the same request has nothing to act on.
Enforce it where it cannot be argued with
There are four layers where read-only can be enforced, and the safest designs combine them.
| Layer | How | Strength |
|---|---|---|
| Credentials | An API token with read scopes only | Strongest: the system refuses writes |
| Tools | Only expose tools that perform reads | Strong: no write tool exists to call |
| Platform | Only offer tools the server marks read-only | Useful filter, depends on the server being honest |
| Prompt | Instruct the model not to change anything | Weakest: can be argued around |
RoutMe applies the third row for the chat assistant: it is offered an MCP tool only if the server marks it with readOnlyHint: true, and a tool with no annotation is left out. Its base instructions also state that the assistant can only look things up, and per-widget instructions cannot turn on actions. The documentation notes one exception: flow Agent nodes can call every tool on the servers attached to the node, so attach only servers whose tools you accept.
The MCP specification says clients must treat tool annotations as untrusted unless they come from trusted servers. So read-only annotations help when you operate the server, and the credentials are the real guarantee.
What read-only still covers
- Order and shipment status
- Stock and price checks
- Policy answers from your documents
- Appointment availability
- Account facts such as plan or seats
These are the questions most stores see repeatedly, and none of them needs a write.
What stays with people
Refunds, cancellations, address changes, exceptions and anything where judgment or money is involved. You can still automate parts of this in a flow, where rules you wrote decide what happens. The agent proposes, and the flow decides. See how to automate returns safely.
What read-only does not protect against
- Wrong information. If the underlying record is wrong, the assistant repeats it.
- Oversharing. A read tool that returns too much data can leak it. Return the minimum fields.
- Misreading. The model can still phrase something badly. Grounding and logging reduce this.
That is why tools should verify ownership and return narrow results, as in how an agent looks up order status safely.
A practical way to audit read-only
Ask three questions of any assistant you are about to publish. First, list every tool it can call, and mark each one as a read or a write by reading the tool's code, not its description. Second, look at the credential behind each tool and check what that credential is allowed to do in the underlying system, since a tool named "get order" backed by an admin token can still write if someone changes the tool later. Third, try to break it: ask the assistant to cancel an order, to change an address, to refund a purchase, and to reveal another customer's details. A correctly built read-only assistant should decline each of them, and your logs should show no write attempts reaching the platform. Repeat the test after every change to your tool server.
Sources
Explore the product: RoutMe for Shopify, solutions and the documentation.