Reliable AI workflows need more than a good prompt

Build grounding, validation, retries, human approval, and observability into automated workflows.

Alex ยท

  • AI automation
  • Reliability
  • Human review

A prompt can ask an assistant to write a faithful summary. A workflow still needs a plan for missing evidence, invalid output, and delivery failures. For an editorial pipeline, those plans determine whether a bad draft reaches someone who can catch it before it goes out.

Give the model a boundary for its evidence

Pass the source material with stable identifiers and require those identifiers alongside factual claims. When sources disagree or a claim lacks support, make that an explicit outcome the workflow can route for review. Keep the source references with the draft so a human can inspect them at approval time.

Validate at each handoff

Check required fields and accepted values in code before using model output downstream. Verify that cited identifiers belong to the current input set. Schema checks help catch broken handoffs, while claim checks and human review help with factual quality. A passing schema check tells you the draft has the expected shape.

Retry only the failures that can recover

A temporary rate limit can justify a bounded retry with backoff. An unsupported claim needs better evidence or a reviewer. Give dispatch operations a stable identifier and record their outcome. If a timeout leaves delivery uncertain, check the destination or deduplication record before attempting another send.

Give approval a clear scope

Show the reviewer the draft, sources, warnings, and destination in one place. Associate approval with the reviewed version. If the text changes after approval, route that new version through review again. Give the operator a way to see whether a draft is waiting, rejected, or ready to send.

Make failures diagnosable

Track a run identifier, the stage that failed, and the disposition of the draft. Log enough context to diagnose routing and validation issues while limiting retained sensitive inputs. Walk through three desktop exercises before starting a pilot: a source disappears, the model returns invalid output, and the destination times out after accepting a send. For each, decide who notices and what they can safely do next.

See AI News Agency for an editorial workflow that puts human review between drafting and dispatch.

Back to writing