Trust

Cost governance & safety

Per-task model routing, a hard budget brake, six roles, confirmed writes.

What it does

On cost: a built-in price table for 17 models, task-level routing that sends light work like email classification to the cheapest model, and full cost instrumentation across 23 task types. The monthly budget is a hard ceiling — at the line, model calls are refused. Not a warning; a brake. Usage slices by model, task, user and run, with a most-expensive-conversations ranking.

On safety: six roles — five customer-side tiers from read-only to admin, topped by a vendor tech tier: API keys, model configuration, tenant customization and maintenance endpoints belong to the vendor alone, so the customer admin stays focused on the business itself (users, teams, publish confirmations). Every endpoint enforces a minimum role by path and method; sales reps see only their own assigned inquiries; every write operation suspends for admin confirmation, backed by an anti-hallucination protocol — the AI may not declare success or invent links while an action is pending.

Core capabilities

Task-level model routing

Light tasks step down, heavy tasks step up — 17 model prices built in.

Monthly budget brake

At the cap, model calls are refused outright. Bills cannot surprise you.

Customer runs the business, vendor runs the base

Six roles enforced per endpoint — infrastructure config and business operations fully separated, every write confirmed.

Feature checklist

  • Six roles: viewer, sales, team lead, operator, admin, plus a vendor tech tier — credit-burning actions start at operator, infrastructure config is vendor-only
  • The tech tier has an anti-escalation gate: tech accounts can only be created and managed by the tech tier itself, never self-granted by a business admin
  • 23 task types fully cost-instrumented; usage aggregates by model, task, user and run
  • Session tokens stored hashed; HttpOnly + strict SameSite cookies; admin can force logout
  • Tiered rate limits: login 5/min, chat 20/min, content generation 6/min
  • A data-source health panel shows live, degraded or disconnected per integration
  • Anti-hallucination protocol: no success claims on pending actions, no duplicate article creation

Want to see this module run?

Book a demo and we'll run this module on real data.