Lesson 3: Decisions and Routing

Your classifier returns a priority and confidence. Now you need different things to happen for different results.

Urgent and confident? Notify the team. Needs action but not urgent? Create a task. Newsletter? Archive it. Uncertain about anything? Send it to a human.

The Decide Step

Add a decide step after classification:

machine classify_email

accepts
  subject as text, is required
  sender as text, is required
  body as text

responds with
  routed_to as text
  priority as text
  confidence as number
  reason as text

implements
  ask classify, using: "anthropic:claude-haiku-4"
    with task "Classify this email.\n\nFrom: ${input.sender}\nSubject: ${input.subject}\nBody: ${input.body}"
    returns
      priority as text, is required, choices: ["urgent", "today", "later", "ignore"]
      confidence as number, is required, range: [0.0, 1.0]
      action as text, is required
      reason as text

  decide route
    if classify.confidence < 0.7
      {routed_to: "human_review", priority: classify.priority, confidence: classify.confidence, reason: "Low confidence: " + classify.reason}
    else if classify.priority == "urgent"
      {routed_to: "notify_team", priority: "urgent", confidence: classify.confidence, reason: classify.reason}
    else if classify.action == "create_task"
      {routed_to: "create_task", priority: classify.priority, confidence: classify.confidence, reason: classify.reason}
    else if classify.priority == "ignore"
      {routed_to: "archive", priority: "ignore", confidence: classify.confidence, reason: classify.reason}
    else
      {routed_to: "inbox", priority: classify.priority, confidence: classify.confidence, reason: classify.reason}

decide is pure logic, like compute. No AI, no cost, instant. It reads the classification result and picks a path.

The confidence check comes first. If the AI is not sure, a human reviews it regardless of priority. Safety before speed.

What You See When It Runs

/run classify_email

With “URGENT: Server down” from [email protected]:

routed_to: "notify_team"
priority: "urgent"
confidence: 0.95
reason: "Production outage from operations team"

With “Quick question” from someone you do not recognize:

routed_to: "human_review"
priority: "today"
confidence: 0.58
reason: "Low confidence: Could be urgent or routine, unclear sender"

With “Weekly deals!” from [email protected]:

routed_to: "archive"
priority: "ignore"
confidence: 0.99
reason: "Marketing newsletter from retail store"

The Execution Trace

After running, check the trace:

/timeline [run_id]

You will see two steps:

[1] classify  ask    420ms  $0.0003  anthropic:claude-haiku-4
[2] route     decide   0ms  $0.0000

The AI step took 420ms and cost a fraction of a cent. The routing step was instant and free. The trace shows you exactly what happened, how long it took, and what it cost.

Three Kinds of Steps

You now know all three basic step types:

Step	What it does	Uses AI?	Costs money?
ask	AI reasoning	Yes	Yes
compute	Data transformation	No	No
decide	Conditional routing	No	No

The pattern: AI does the judgment. Logic does the routing and formatting. You choose where AI is involved and where it is not. That choice is visible in the code and in the cost trace.

What Comes Next

The machine routes to “notify_team” or “create_task” but does not actually do those things yet. Next lesson: connecting to real systems. Send a message in Microsoft Teams when urgent. Create a task in Planner when action is needed. That is where this becomes useful.

Next: Taking Action →