Live demo 2 · Customer support
Support Copilot
A support agent pastes a ticket. The copilot does the first five minutes of work: what the ticket is about, how the customer feels, how urgent it is, which help article fits and a draft reply. It never sends on its own when the ticket is risky: account security, legal language and an upset customer about to leave all go to a person.
Incoming ticket
Try a sample
The copilot’s work
Run the copilot or pick a sample to see the classification, mood, priority, the matched article and the draft reply.
Evaluation
Run the test suite against the live engine
The same checks run before every release. Guardrail cases are part of the suite, so a safety rule that breaks fails the build.
Part two · Customer success
Account health: who might not renew
The same copilot, one step earlier. A CSM enters what they know about an account and gets a risk tier, the reasons behind it and one next step from a playbook. What the customer said in words outranks the numbers, missing data is listed instead of guessed, and nothing goes to the customer.
Account signals
Try a sample
Risk read
Score the account or pick a sample to see the tier, the reasons, the quoted line and the next step.
Evaluation
Run the account health checks
Eleven invented accounts with the tier a CSM would expect, including missing data, a note that tries to change the score and a request for a discount.
Real vs simulated
- Real
- The workflow: triage, priority rules, a knowledge base lookup, escalation rules that sit outside the model and a draft a person approves. This is how I set up AI in support teams.
- Simulated
- The language model step and the help centre. A rules engine and a small invented knowledge base stand in, so results are instant and repeatable. The health score uses hand-set weights, not a model trained on real renewals.
- Privacy
- Nothing you type is stored. Card numbers and emails are masked before the ticket text is shown.