What an AI workflow audit is
An AI workflow audit maps one value stream from trigger to outcome. It looks at the work, the data, the handoffs, the decisions, the reporting, and the repeated administrative load around that workflow so AI opportunities are evaluated in context.
When to use this checklist
Use it before buying a tool, rebuilding a process, adding automation, or asking a team to adopt AI. Pick one workflow that matters — sales, onboarding, delivery, support, renewals, hiring — and walk it end to end with the operators who actually run it.
Intake checklist
- Is intake structured enough that two operators would capture the same facts?
- Are requirements, source documents, stakeholders, and success criteria captured before work begins?
- Does the request get scored for urgency, value, complexity, and owner?
Handoff checklist
- Does every handoff include context, owner, next action, and due date?
- Where does work wait because the receiving team lacks the full story?
- Which handoffs depend on Slack threads, memory, or one senior operator?
Customer journey checklist
- Where does the customer repeat information the business already has?
- Which moments feel slow, unclear, or inconsistent from the customer's point of view?
- Where could AI summarize context before a customer-facing interaction?
Employee workflow checklist
- Which tasks exist only to reconcile tools or restate context?
- Where do employees copy, paste, reformat, or chase information every week?
- Which steps drain senior judgment without improving the outcome?
Data flow checklist
- Can the workflow be drawn on a single page without omitting a step?
- Is every step owned by exactly one role?
- Are handoffs documented, or do they live in memory?
- Where does the same record exist in more than one system?
- Which steps exist only because a tool can't talk to another tool?
- Which steps exist only because a previous decision is hard to undo?
Decision point checklist
- Which decisions in this workflow wait on numbers nobody can pull in under five minutes?
- Which numbers are pulled regularly but never inform a decision?
- Where does data quality silently degrade between steps?
- Which dashboards exist for confidence, not for choice?
- Where does a person re-key data from one system into another?
Reporting checklist
- Which reports are manually assembled from exports?
- Which dashboards are viewed but not used to make decisions?
- Where does reporting lag the operating moment it is supposed to improve?
Repetitive admin checklist
- Which steps generate text from structured inputs?
- Which steps classify, route, or prioritize?
- Which steps summarize long-form content for downstream readers?
- Which steps draft a first version a human reviews?
- Which steps wait on synthesis across multiple systems?
- Which steps require pattern recognition across history?
Revenue leakage checklist
- Where does follow-up slow after a qualified lead or customer signal?
- Which revenue opportunities disappear because ownership or timing is unclear?
- Where does poor data quality make pipeline or expansion reporting unreliable?
Retention signal checklist
- If the senior operator stopped doing this for a week, what would break?
- What is the cost of a single bad output here?
- Where would automation lock in a process you would not redesign by hand?
- What does retraining and adoption actually take if you change this step?
- Where is institutional knowledge undocumented and at risk?
- What happens in this workflow at 2x volume?
AI opportunity scoring
- Which redesigns must happen before any AI is applied?
- Which AI changes generate the most operator time back?
- Which AI changes the customer can feel?
- Which changes are cheap to reverse?
- What would you want to be true about this workflow in 90 days?
What to do with what you find
Cluster the answers into three buckets: redesign first (workflows that will become worse under automation), automate next (well-understood, stable, high-volume), and leave alone (low-leverage, high-context, or soon-to-be-deprecated). Then sequence by impact and reversibility — not by which vendor called last.