#Examples

One policy, three modes. Suppose your administrator authored a policy called customer-refund with declared inputs customer, product, cost, and issue. The same policy can serve three calls — different mode, same policies array, same inputs bag.

For the request body shape and response field reference, see Policy engine overview.

#Validate — should this refund go through?

{
  "request_id": "5b91d2c4-6a3f-4d1f-9c4a-2f0e7b1f8d4e",
  "results": [
    {
      "slug":              "customer-refund",
      "version":           4,
      "mode":              "validate",
      "status":            "ok",
      "outcome":           { "passed": true, "reasons": [] },
      "failed_predicates": null,
      "detail":            null,
      "citations":         [],
      "latency_ms":        2
    }
  ],
  "summary": null
}

#Generate — draft the response to Sarah

Same policy, mode flipped to generate. The same inputs are threaded into the LLM prompt — the engine produces policy-compliant text.

curl -X POST https://aiengine.velgent.com/api/v1/policies/evaluate \
  -H "Authorization: Bearer velgent_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "policies": [{ "slug": "customer-refund" }],
    "mode":     "generate",
    "inputs":   {
      "customer": "sarah@acme.com",
      "product":  "blender",
      "cost":     200,
      "issue":    "stopped working after 4 months"
    }
  }'

{
  "request_id": "7c81f1a4-2d9e-4f3b-b1a2-9e3d7f0a6c5b",
  "results": [
    {
      "slug":       "customer-refund",
      "version":    4,
      "mode":       "generate",
      "status":     "ok",
      "outcome":    {
        "generated_text": "Hi Sarah, I've reviewed your case for the blender. Since the failure happened within the manufacturer's warranty window, you qualify for a full refund of $200. I'll have it processed today...",
        "reason":         "Generated by claude-sonnet-4-6."
      },
      "failed_predicates": null,
      "detail":            null,
      "citations":         [],
      "latency_ms":        1840
    }
  ],
  "summary": null
}

#Decide — pick the right action

Same policy again, mode decide. The administrator has populated the policy's action catalogue with two options: approve_refund (with a Workday payload template) and escalate_to_manager. The LLM picks one based on the inputs.

curl -X POST https://aiengine.velgent.com/api/v1/policies/evaluate \
  -H "Authorization: Bearer velgent_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "policies": [{ "slug": "customer-refund" }],
    "mode":     "decide",
    "inputs":   {
      "customer": "sarah@acme.com",
      "product":  "blender",
      "cost":     200,
      "issue":    "stopped working after 4 months"
    }
  }'

{
  "request_id": "2a4f9e1c-7b8d-4a1f-9c2e-5e0a3b8f1d4e",
  "results": [
    {
      "slug":       "customer-refund",
      "version":    4,
      "mode":       "decide",
      "status":     "ok",
      "outcome":    {
        "action_id": "approve_refund",
        "payload":   {
          "system": "workday",
          "amount": "200",
          "user":   "sarah@acme.com"
        },
        "reason":    "Customer is within 30-day window and product is defective."
      },
      "failed_predicates": null,
      "detail":            null,
      "citations":         [],
      "latency_ms":        1620
    }
  ],
  "summary": null
}

The payload is the action's payload_template from the catalogue, with ${input.X} placeholders substituted against the request inputs. Your caller fires it downstream — the engine doesn't execute actions itself.

Next: Chains →