Skip to main content
Open Quality & Testing → Simulation to test AI Assistant against a Zendesk ticket before allowing the same workflow to affect production systems. Simulation is designed to preserve realistic reasoning and retrieval while removing the most dangerous part of a test: accidentally performing the real mutation.

​
What Simulation runs

A Simulation uses the Assistant pipeline against the selected Zendesk ticket and current organization configuration. That means it can exercise:
  • ticket context;
  • procedures;
  • Response guidelines;
  • semantic knowledge retrieval;
  • configured AI models;
  • SOURCE tools;
  • planned Zendesk changes;
  • confidence; and
  • final response generation.

​
SOURCE tools

SOURCE tools run normally in Simulation. This gives the pipeline authentic read-only data such as current order/account status instead of a fake response that might hide integration problems. Be mindful that read-only does not mean free. A Simulation can still call an external API and consume provider/tool quota.

​
ACTION tools

ACTION tools are intercepted by the dry-run path. RAUM records what the action would have attempted without applying the real side effect. Examples of effects you normally do not want a test to perform include:
  • refunding money;
  • cancelling an order;
  • changing a subscription;
  • creating a return; or
  • mutating another external system.

​
When to use Simulation

Run representative simulations after changes to:
  • RUNBOOK procedures;
  • AI models;
  • high-impact AI Guidelines;
  • API Connections;
  • MCP tool configuration;
  • ACTION approval behavior;
  • Agent Mode policy; or
  • Zendesk automation scope.

​
What to review

After the run, inspect:
  1. the selected procedure;
  2. retrieved knowledge;
  3. SOURCE results;
  4. proposed ACTION calls;
  5. any planned Zendesk field/status updates;
  6. the final customer-facing answer; and
  7. confidence.
If the final answer is wrong, identify whether the failure came from knowledge, targeting, a tool response, procedure logic, or model behavior before editing random prompts.

​
Simulation is not a production permission

A successful dry run does not automatically enable an ACTION, Agent Mode, or automatic approval. Simulation proves that the configured pipeline can produce the expected behavior under the tested conditions. Production authorization remains controlled by the actual connection and automation settings. For an important workflow, simulate more than the happy path:
  • clearly eligible case;
  • clearly ineligible case;
  • missing-information case;
  • conflicting-information case;
  • tool failure;
  • low-confidence case; and
  • a case that should escalate to a human.

Agent Mode

Widen autonomous handling only after representative tests pass.

Connections

Configure SOURCE/ACTION behavior and approvals.