Control

The agent answers from what you gave it, does only what you allowed, and asks before anything that costs money.

Rules enforced in code, not prompts

A prompt instruction is a suggestion. Rules live outside the model at three points: a gate before any tool runs, a gate before anything sends, and a verifier on every reply.

  • A rule the model cannot talk its way past
  • Every refusal is recorded in the audit log

Never quote before lane and volume are captured

Tool blocked

Hazardous cargo always goes to a human

Escalated

Discounts above 10% need a manager

Approval required
Enforced at three points outside the model: before a tool runs, before anything sends, and on every reply.

Approvals for anything that spends

Requested ad campaigns, budget moves, discounts past a limit. A person decides, with the number attached.

  • The agent can request, never apply
  • Thresholds are yours to set, per rule
Approvals
1 waiting

12% discount on quote Q-1148 · ₹4.2L

AI

Requested by the agent mid-conversation. Above the 10% threshold, so it stops here.

ApproveDecline
A person decides, with the number attached — never applied silently.

The rest of control

Five effects, not free text

Block a tool until fields are captured, require a human first, force escalation on a topic, require approval above a threshold, or append an instruction. A closed set is auditable; free text is not.

The agent will not deny being an AI

It does not volunteer it either. Denial is refused by the verifier, outside the model, and is not configurable.

Grounded knowledge with citations

Chunks link back to the document they came from, so an answer can point at its source instead of asserting something unbacked.

Workflows as rows

Trigger, condition, action — edited like any other configuration. A visual canvas is a product in itself and is deliberately deferred.

See it with data in it

The demo is the real product, seeded with two businesses that sell completely differently.