AI Decisions

An AI agent writes text and decides which tools to call. Often a process step needs less than that: one bounded answer that a gateway can branch on. Which team gets this ticket? Is this request urgent? How bad is the impact?

A decision model answers exactly that kind of question. Cloudflare’s Clef models do not generate text. They score the options you define in one pass and return a typed answer with a probability for each option. A gateway can then route on the answer, and route on how sure the model is.

The template

@bpmnkit/connectors bundles Cloudflare Clef Decision (io.bpmnkit.connectors.CloudflareClef.v1). It is an element template for a service task that runs on Camunda’s REST connector (io.camunda:http-json:1), so any cluster with the connector runtime can run it. You do not deploy a job worker.

To set it up on a cluster and test it step by step, follow the how-to Route a Camunda 8 process on a Cloudflare Clef decision.

Before you deploy, create the secrets CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID in the cluster’s connector secrets.

Questions and answers

Each question has a type and instructions. Answers come back under the same ids. With the default result expression, the answer to question team is in clef.team.

={
  "urgent": {"type": "noul", "instructions": "Is this support request urgent?"},
  "team": {
    "type": "choice",
    "instructions": "Which team should handle this request?",
    "criteria": {
      "billing": "Payments, invoices and refunds",
      "technical": "Outages, errors and configuration",
      "sales": "Plans and upgrades"
    }
  },
  "severity": {
    "type": "score",
    "instructions": "How severe is the customer impact?",
    "criteria": ["No impact", "Minor", "Major", "Critical"]
  }
}

The answers for these questions look like this:

{
  "urgent": { "type": "noul", "noul": 0.94 },
  "team": {
    "type": "choice",
    "choice": "technical",
    "probabilities": { "billing": 0.06, "technical": 0.91, "sales": 0.03 },
    "confidence": 0.88
  },
  "severity": {
    "type": "score",
    "score": 2.7,
    "legend": { "0": "No impact", "1": "Minor", "2": "Major", "3": "Critical" },
    "probabilities": { "0": 0.01, "1": 0.04, "2": 0.2, "3": 0.75 },
    "confidence": 0.71
  }
}

Example: triage a support ticket

This process asks the three questions above about a ticket. It pages on-call for an urgent technical ticket. It sends a ticket to a person when the model is not sure which team should have it. It queues the remaining tickets by team. The full source is apps/examples/src/07-ai-ticket-triage-clef.ts.

import { applyConnectorTemplate } from "@bpmnkit/connectors"
import { Bpmn } from "@bpmnkit/core"

const triage = applyConnectorTemplate("io.bpmnkit.connectors.CloudflareClef.v1", {
  model: "clef-flash",
  "body.state": "={subject: ticket.subject, body: ticket.body, plan: customer.plan}",
  "body.questions": QUESTIONS, // the FEEL context above
})
if (triage.problems.length > 0 || !triage.serviceTask) throw new Error(triage.problems[0]?.message)

const definitions = Bpmn.createProcess("TicketTriage")
  .withAutoLayout()
  .startEvent("start", { name: "Ticket Received" })
  .serviceTask("triage", { ...triage.serviceTask, name: "Triage Ticket" })
  .exclusiveGateway("route", { name: "Route Ticket" })
  .branch("incident", (b) =>
    b
      .condition('=clef.team.choice = "technical" and clef.urgent.noul >= 0.8')
      .serviceTask("pageOncall", { name: "Page On-Call", taskType: "pagerduty-alert" })
      .endEvent("endPaged"),
  )
  .branch("unsure", (b) =>
    b
      .condition("=clef.team.confidence < 0.6")
      .userTask("manualTriage", { name: "Triage by Hand", formId: "manual-triage-form" })
      .endEvent("endManual"),
  )
  .branch("billing", (b) =>
    b
      .condition('=clef.team.choice = "billing"')
      .serviceTask("billingQueue", { name: "Queue for Billing", taskType: "billing-queue" })
      .endEvent("endBilling"),
  )
  .branch("other", (b) =>
    b
      .defaultFlow()
      .serviceTask("supportQueue", { name: "Queue for Support", taskType: "support-queue" })
      .endEvent("endSupport"),
  )
  .build()

const xml = Bpmn.export(definitions)

The exclusive gateway takes the first flow whose condition is true, in the order above. An urgent incident is paged even when the team answer is not certain.

Set the confidence threshold from your own data. Start with a value that sends a good share of tickets to a person. Look at what the people decide, then lower the threshold.

Test the routing

In a test, mock the REST connector’s job type with the variables that the result expression writes. You can then test each branch without a call to Cloudflare:

import { createProcessTest } from "@bpmnkit/engine/testing"

const t = await createProcessTest({ bpmn: xml })
t.mockJob("io.camunda:http-json:1", {
  result: {
    clef: {
      urgent: { type: "noul", noul: 0.2 },
      team: { type: "choice", choice: "billing", probabilities: {}, confidence: 0.41 },
      severity: { type: "score", score: 1, legend: {}, probabilities: {}, confidence: 0.8 },
    },
  },
})
const run = await t.start("TicketTriage", { ticket: { subject: "Refund", body: "…" } })
// run is waiting at "manualTriage"

apps/examples/tests/ticket-triage-clef.process.test.ts tests all four branches.

Use the template in the editor

In the BPMN Kit editor, select a service task and pick Cloudflare Clef Decision in the Connector list. The panel then shows the template’s fields.

Use the template in Camunda Modeler

To use the template in Desktop Modeler or Web Modeler, write it to a JSON file:

import { writeFileSync } from "node:fs"
import { getTemplate } from "@bpmnkit/connectors"

const clef = getTemplate("io.bpmnkit.connectors.CloudflareClef.v1")
writeFileSync(".camunda/element-templates/cloudflare-clef.json", JSON.stringify(clef, null, 2))

The template is in the AI decisions category of the template chooser.

Limits

These limits are from Cloudflare’s model page:

• 1 to 64 questions for each call. A choice has 2 to 255 options. A score has 2 to 10 levels.
• The context window is 64k tokens. Long text state is truncated to fit.
• Images: 4 at most, 4 MiB each, 8 MiB in total. Data URLs only. Remote URLs are not accepted.
• A failed call (for example HTTP 401 or 429) fails the job. The template’s retries and retry backoff apply. To raise a BPMN error instead, set the error expression.