What should you automate before introducing AI?

Map the current process

Start with a real task: perhaps an email or document whose details someone later copies into a spreadsheet or another system. Follow it to the decision or action it supports. Note where information is copied, who is responsible and what happens when details are missing or conflict. This gives AI a clear job instead of asking it to fix an unclear process.

Make routine work consistent

Make routine steps consistent when they already follow clear rules. Decide where records are kept, which fields are required, who makes each decision and who receives the next task. A fixed rule and action may be enough; AI is unnecessary when the input and result are predictable.

Make information usable

Check whether people can find and use the information the task needs. Decide which system holds the current record, who keeps it up to date and when changes reach other systems. If records disagree, decide when to stop the process or ask someone to check. AI cannot know which record is correct unless you set that rule.

Give AI one clear task

AI can help when a task involves free-form information, such as reading an email. State what it should read, what result it should prepare and what to do when the information is unclear. Also decide where approved changes should be saved and who makes the final decision.

For example, a fixed rule can route a complete website inquiry to the right owner. AI may help classify a free-text request that does not fit a fixed category. If the inquiry lacks required details or conflicts with an existing record, route it for review instead of guessing.

• Use AI to extract or classify information that a person would otherwise review manually.
• Keep the result tied to a specific next step in the workflow.
• Define observable conditions for pausing or escalating uncertain cases; do not rely on a confidence score alone.
• Avoid using AI when a clear rule or ordinary integration solves the problem more simply.

Keep responsibility clear

Decide who reviews the result, what needs approval and how an error can be corrected. Start with a limited workflow, compare outputs with the intended task and adjust the process before relying on it more widely.