LLM API for workflow automation

Workflow automation in a production application

An accounts-payable workflow reads an invoice, matches it to a purchase order, asks the model to classify discrepancies, and either queues approval or opens an exception ticket. Workflow state lives in the application database, not in chat history.

The model decides only the ambiguous classification. Parsing, arithmetic, duplicate detection, writes, retries, and deadlines stay deterministic. Every external write carries an idempotency key.

How the workflow automation workflow operates

• Persist state and immutable input references.
• Run deterministic parsing and matching.
• Call the model only for the ambiguous decision.
• Validate the decision against policy and confidence rules.
• Write the next state idempotently or route to review.

Call gpt-5.6-terra through LLMFly AI

Send the request from your server. Replace the example content and placeholder tool schema with data and tools from your application.

curl https://app.llmfly.ai/v1/chat/completions \
  -H "Authorization: Bearer $LLMFLY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-terra",
    "messages": [
      {"role": "system", "content": "Classify the invoice discrepancy using the supplied policy. Return MATCH, REVIEW, or REJECT with the exact policy rule. Do not perform any external action."},
      {"role": "user", "content": "Classify this invoice and purchase-order discrepancy."}
    ],
    "response_format": {"type": "json_object"}
  }'

Why gpt-5.6-terra is the primary model

Terra handles the bounded judgment step while preserving enough reasoning for policy and exception classification.

Use gpt-5.4-mini for narrow classifications that pass validation. Escalate novel exceptions to Sol or a person rather than widening the model's permissions.

Acceptance checks for workflow automation

Failures to handle before deployment

• Using chat history as the state store
• Retrying non-idempotent writes
• Granting unnecessary permissions
• Having no terminal failure path

Returned output and run records

A structured classification, policy rule ID, explanation, and review reason when applicable.

Record the model ID, request ID, token usage, retries, validation result, and final disposition for every production run.