By clicking “Accept”, you agree to the storing of cookies on your device to enhance site navigation, analyze site usage, and assist in our marketing efforts. View our Privacy Policy for more information.
All demos
AI adoption and enablement

Tickets that fulfil themselves

IT · Service desk · 1:30

Demo · fictional organization · real working software

The problem

IT queues fill with requests a service catalog could fulfil, so engineers spend their days granting and revoking access by hand. An agent could do most of it, but only if the business can bound what it may do.

How it's resolved

Quellbrook Freight, a fictional freight company of about 1,400 staff across depots. The operator sets a policy for its IT agent, one row per access tool: whether it is on, how many units a call may handle, which kinds a person must approve, and how many calls a minute. The Traide MCP server enforces it, outside the model, so the agent cannot change it. On Monday the AI reads six messages, and code checks each reading before any call. A route planner seat and a carrier portal API key are allowed, placed as $0 requests billed to the firm, and their keys issued in under a second. Group access to a shipments data share always reaches a person, so its row holds the call. Ten seats against a limit of five are refused by the row, and the AI drafts a reply that is not sent. The IT lead opens the held call, sees the exact arguments and their digest, and approves it with one click; the agent resumes it, and the API lets it run once, only if the digest matches. A leaver’s access is listed and revoked, item by item, and every call is on the record as the server wrote it.

What you'll see

  1. 01The agent’s policy: one row per access tool, set by the operator and enforced outside the model
  2. 02Monday’s inbox: the AI reads each message, and code checks its reading before any call
  3. 03A seat is allowed by its row, placed as a $0 request, and its key issued 948 ms later
  4. 04Group access always reaches a person: its row holds the call
  5. 0510 seats against a limit of 5: refused by the row, no order placed, a reply drafted and not sent
  6. 06The IT lead opens the held call, with its exact arguments and digest, and approves it: it runs once, in 1,284 ms
  7. 07A leaver’s three items are listed and revoked, none still active, and all 14 calls are on the record

Under the hood

Each business term in the video, and the Traide building block it runs on.

In the businessOn Traide
The agent’s policyPer-tenant MCP operator policy: one row per tool, enforced by the Traide MCP server
RequestOrder: $0, billed to the firm
A seat or key, issuedEntitlement from a license key pool
A held callMCP approval request: the stored arguments and their digest, claimed once
Access removedEntitlements by email, then an entitlement revoke for each
The recordThe MCP server’s audit line: tool, argument keys, outcome, milliseconds

Guardrail. The policy lives in the API and is enforced by the MCP server, outside the model: the agent cannot raise a limit. A held call runs only after a person who is not the agent approves it, exactly as stored and once. Group access always reaches a person. The AI reads and drafts; code checks every reading before any call. Every request is $0, billed to the firm, in test mode.

  • Tool use
  • Classification
  • Guardrails
Strategy call · 30 minutes

Have a problem like this one?

You'll talk to a senior engineer, not a sales rep, and leave with a clear first move — whether or not we work together.