Checklist · 9 min read

AI workflow design & QA checklist

A pre-flight checklist for workflows that use models to research, classify, draft, recommend, or take action in GTM.

Key idea

Treat every AI workflow as a product: define its job, input contract, human control, failure behavior, and evidence of value.

1. Write the workflow brief

PromptA good answer
What is the job?Summarize verified account context for a rep before a first call
What does success look like?A rep can identify one relevant hypothesis in under two minutes
What is out of bounds?No invented facts, no sensitive data, no automatic customer contact
Who owns the outcome?Outbound operations lead, with Sales leadership as approver
What happens when uncertain?Cite missing evidence and route to a human review queue

2. Check the input contract

  • Every important input has a named source and freshness expectation.
  • The workflow can distinguish missing data from a negative signal.
  • Restricted, personal, or customer-confidential data is excluded by policy—not by prompt wording alone.
  • The model receives the minimum context needed for the job.
  • The output includes source links or evidence whenever it makes a factual claim.

3. Build quality assurance into the flow

  1. Create a test set

    Use 20–50 representative records, including messy inputs, missing data, edge cases, and known failures.

  2. Define a rubric

    Score factual grounding, relevance, policy compliance, usefulness, and correct escalation—not just whether the prose sounds good.

  3. Review before activation

    Start with a human approval queue. Only automate an action after the workflow proves reliable in that context.

  4. Log the evidence

    Store workflow version, input references, output, reviewer action, and final outcome so you can investigate and improve.

4. Measure the outcome, not the novelty

  • Adoption: Are intended users choosing to use it?
  • Quality: Is the output grounded, relevant, and accepted by reviewers?
  • Speed: Does it reduce time to a useful action?
  • Business result: Does it improve the downstream result it was designed to influence?
  • Risk: How often does it require correction, create an exception, or cross a policy boundary?