Tests and casesIn development
Every change is tested before reaching customers
Each business will have its own set of cases. After every change to knowledge, instructions or the model they will all run, and a change that makes the results worse will not reach customers without a recorded exception.
Early access with hands-on help from the team. Ray is in development.
Illustrative example of a candidate change on hold: 23 of 24 cases pass, the sensitive-topics floor is complete, one case got worse in Brazilian Portuguese, with buttons to view the case or record an exception.
Illustrative example with sample data. Illustrative data.
The business's casesIn development
Each case says what is expected of Ray
Each case will be a real customer question, with one of four expected outcomes. Asking for clarification and saying there is no source count as correct: what is avoided is guessing.
- Resolve with a sourceThe case's question: How long does domestic shipping take?
Reply with the published, valid article that applies, with the exact time frame.
- Ask for clarificationThe case's question: My order never arrived.
Ask which order it is about before replying, instead of handing off.
- Say there is no sourceThe case's question: Do you ship to a P.O. box?
Say there is no published information and suggest a person, without guessing.
- Hand offThe case's question: I want my data deleted.
Send the holding message and pass to a person, with a draft.
Illustrative example of the test-case table of a candidate version: five cases with the expected outcome; four pass and one got worse, so the change is on hold.
Illustrative example with sample data.
RegressionsIn development
If it gets worse, it will not reach customers
If results get worse beyond the tolerance, the change will be put on hold and customers will stay on the previous version. Publishing it anyway will take an exception with a reason.
What is recorded
Each run will store the model, the version of the instructions and that of the knowledge.By language, no averages
Results will be reported for each enabled language, never averaged.Cases that grow
A draft rejected with a correction in the Inbox will become a proposed case, without personal data.
Illustrative example of recording an exception: the case that got worse, the written reason, who records it and the buttons to record and publish or to cancel.
Illustrative example with sample data.
Safety floorIn development
Sensitive topics are always tested
The cases of the base list of sensitive topics will run on every change, for every business, and cannot be turned off. Each business will add cases for its own topics.
Illustrative example of the sensitive-topics floor: six cases that cannot be turned off, all passing.
Illustrative example with sample data.
HistoryIn development
Every run will be recorded
What changed, how many cases ran, the result and who published or recorded an exception. The history will be read-only.
Illustrative example of the run history: five runs with date, change, number of cases, result and owner; one on hold and one with a recorded exception.
Illustrative example with sample data. Illustrative data.
Going liveIn development
In stages, and with an emergency switch
After testing, Ray will go live gradually. An emergency switch will stop automatic replies, and new conversations will go to the Inbox.
- 01
Paused
Ray will not reply. Useful for preparing the knowledge.
- 02
Copilot only
Ray will write drafts and the team will decide what gets sent.
- 03
Autonomous by percentage
Ray will reply to a share of customers, and the share will grow in stages.
Build the first test cases with the Ray team
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.