Open to AI engineering rolesRun the live demo →
Provenance
01Client use caseMeridian Nine support

AI support that knows when not to answer.

Provenance answers routine customer questions from approved company documents, verifies every claim, and sends uncertain or unsafe requests to the right place.

Approve automaticallySend to a personBlock safely
Run the guided demo See what it solves
Live client workflow

Choose a support ticket to see the decision.

Each scenario runs through the same production-minded pipeline. Watch what gets automated, what needs a person, and what is stopped.

1Choose a ticket2Watch the checks3Inspect the outcome
Ready for a decision

Choose a guided scenario above to see the decision.

This is a presentation-only simulation. In production, reply thresholds, permissions, retention, and escalation routes are configured by the workspace operator. Every decision above is real — screening, retrieval, generation, and verification all run live against the actual pipeline. Sending a reply to the customer stays simulated — no email or real ticketing system is touched — but when a Slack connector is configured, the operator notification above is real, including the Approve/Reject buttons on tickets routed to human review.
What the client gets

Controlled automation, not blind automation.

The system handles the work it can prove, and keeps people in control of everything else.

01

Faster routine support

Approved answers can move from intake to a cited response without an agent researching the same policy again.

02

Lower answer risk

A response is withheld when even one important claim cannot be verified against approved documentation.

03

Clear human handoff

Liability, missing-policy, and uncertain cases arrive with evidence and a reason for escalation.

04

Auditable decisions

Each ticket keeps its source passages, verification result, route, and decision history.

How it works · inspectable by design

Every answer earns permission to be sent.

A customer message is screened, matched to approved knowledge, drafted, and checked claim by claim. The result is then approved, escalated, or blocked.

Explore the architecture
Measured on the demo corpus

Built to prove the decision, not just produce text.

The committed evaluation suite covers answerable, unsupported, adversarial, guided-scenario, and workspace-overlay cases. These are demonstration-corpus results, not universal production claims.

45
evaluated cases
0%
fabrication rate
0%
false-refusal rate