Ray agentIn development
An agent that resolves in the same conversation
Ray will answer in the site chat only from the knowledge each business publishes and from read-only account data. It will guide step by step, report an order’s status and close the conversation; only sensitive topics will go to a person.
Early access with hands-on help from the team. Ray is in development.
Illustrative example: a customer of a fictional store asks how to request a refund and Ray answers with the steps; then they ask about the refund’s status, Ray reports it with read-only order data and the customer confirms the question was resolved, without going to the team.
ResolutionIn development
Resolve in the conversation, don’t hand off
Ray will try to close every question in the same chat. A person will step in only in three cases: a sensitive topic on the list, an explicit request from the customer, or a question with no source after clarifying it.
A question before handing off
If the question is ambiguous, Ray will ask for the missing detail instead of passing it to the team.Step-by-step guidance
With the exact names of menus and buttons, as they appear in the published article.Account data, read-only
The status of an order, a shipment or a refund, read-only and with its update time.Always a next step
If something can’t be done, it will point to the published alternative: another deadline, another channel or another form.Closing the loop
It will ask whether the question was resolved and whether anything else is needed before marking the conversation as resolved.
Illustrative example: the customer writes that their order has not arrived; Ray asks which order it is, the customer says, and Ray reports that it is on its way, with the estimated delivery date and the path on the site to track it.
Illustrative example with sample data. Lumora is a fictional brand.
TraceabilityIn development
The customer will get a clear answer; the team will see where it came from
Amounts, deadlines and conditions will come from published, current articles or from account data. The customer will get the answer without internal references, and in the console the team will see the cited article, its version and the data that was looked up.
Illustrative example in two panels. The customer sees only the answer, with the AI notice and the question of whether it was resolved. The team sees in the console the cited article and its version, the order data looked up read-only and the checks the answer passed.
Illustrative example with sample data. Lumora is a fictional brand.
Numbered steps
A procedure will come as short, numbered steps, not as one long paragraph.
Exact names
Menus and buttons in bold, with the same path as on the site, such as My account › Orders.
No article cards
The customer will get the written answer, not a link to read on their own.
AI notice
Every reply from Ray will carry the notice that it is an automated reply.
Sensitive topicsIn development
A short, explicit list that each business defines
Ray will come with a base list of sensitive topics that will always stay on, and each business will add its own. Only on those topics will Ray send a fixed holding message, and the team will receive the conversation with a draft, the sources and the context. Outside the list, Ray will try to resolve.
Personal data and deletion
Requests to access, correct or delete data.
Legal claims
Mentions of legal action, a formal complaint or a regulator.
Disputed payments and refunds
Unrecognized charges, declines and chargebacks. The draft will not promise amounts or timelines.
Account closure
A person will handle closure requests.
Health or safety
Any risk to a person will be handed off in full, with an alert to supervisors.
An explicit request for a person
If the customer asks to talk to someone, the request is always honored.
Illustrative example of the sensitive topics list in the console: six default topics, two added by the business and the fixed holding message.
Illustrative example with sample data.
On those topics, the team will get the conversation with the reason, the context and a draft from Ray in the Inbox and copilot.
Read-onlyIn development
Ray reports; people move the money
Ray’s tools will be read-only. It will be able to report the status of an order or a refund, but refunds, payments and account changes will still be made by people, in the business’s own systems.
Without an identity signed by the business’s server, the customer will be treated as anonymous and Ray will not look up their account data.
- The status of an order, a shipment or a refund
- The estimated delivery date and when each piece of data was last updated
- The published steps to complete a task on the site
- Move money or issue refunds
- Change an account’s data or close it
- Promise amounts or timeframes that are not published
SetupIn development
Each business will decide how Ray will work
In the console, the business will choose the mode, the share of traffic, the widget’s language, its own sensitive topics and the tone instructions, and it will test every answer before publishing it.
Illustrative example of the Ray agent screen in the console: autonomous mode, 25% of traffic, the widget in Latin American Spanish, sensitive topics and a test answer with its source, visible only to the team.
Illustrative example with sample data.
Widget language
Latin American Spanish, English or Brazilian Portuguese: the business will set it in the console and Ray will reply in that language. If a customer writes in another one, they will get a fixed line in the configured language.
Mode and traffic
Paused, copilot only or autonomous, with a share of customers that will grow in stages. An emergency switch will stop the automated replies.
Tone and instructions
Short tone instructions, such as addressing the customer formally. Replies will still come only from the published knowledge.
Tested before publishing
Every change to knowledge, instructions or the model will run through the test cases; if something gets worse beyond the tolerance, it will be put on hold.
See Ray with the business’s own knowledge
A short conversation with the team that builds Ray to review the company’s case: frequent questions, existing knowledge, sensitive topics and a migration plan.