How to choose your first AI automation workflow

Make a shortlist of complete tasks

Ask the team for three repeated tasks that interrupt useful work. Describe each from the incoming request to the completed result. “Use AI for operations” is too broad. “Prepare a customer brief and draft reply for a salesperson to review” gives you something concrete to examine.

Include the awkward steps: finding records, asking a colleague, checking a draft and updating another system. A task that takes ten minutes in conversation may take much longer once you count those handoffs. Use observed examples rather than the fastest case someone remembers.

Apply readiness checks before comparing savings

Can the system access the required information? Does somebody know what a correct result looks like? Can mistakes be noticed before they affect a customer or a business record? If those answers are unclear, the first project may be organizing the process or data rather than adding AI.

Name a person who can review exceptions and maintain the rules. Start with a draft or proposed update when a wrong action would be difficult to undo. A promising time estimate is not enough if nobody can own the workflow after launch.

• Input: identify the actual records, documents and permissions needed.
• Output: describe what counts as a complete, acceptable result.
• Exceptions: decide who handles missing information and conflicting records.

Compare net weekly time with the same calculation

Use a simple estimate: weekly task volume × minutes removed per task − weekly maintenance minutes. The time remaining with AI must include preparation, review, corrections and exception handling. Measure those inputs during a trial; before then, label them as assumptions.

Here is an illustrative comparison, not a client result. Meeting recaps: 10 per week, 15 minutes today, 7 minutes with AI, plus 20 minutes of weekly maintenance. Estimated saving: 10 × (15 − 7) − 20 = 60 minutes a week.

Quote preparation: 40 per week, 12 minutes today, 6 minutes with AI, plus 30 minutes of maintenance gives 210 minutes saved. Invoice entry: 80 per week, 5 minutes today, 4 minutes with AI, plus 40 minutes of maintenance gives 40 minutes saved.

On these assumptions, quote preparation is the stronger time-saving candidate even though invoice entry happens more often. Change the assumptions and the ranking may change. Quality, access and ownership still have to pass the readiness checks.

Check whether a simpler automation can do the job

Some steps need a fixed rule, a form or a connection between systems. Use AI where interpreting varied information adds something useful. For example, software can route a request using a known customer identifier while AI prepares a summary of the customer’s message.

Anthropic’s guidance on building effective agents recommends beginning with the simplest useful solution and adding complexity when it improves outcomes. I would apply that principle before deciding that the first project needs an autonomous agent. The goal is a dependable workflow the team can operate.

Choose a first version with a clear boundary

Define one trigger, a limited set of information sources, a proposed output and an approval point. For the quote example, the first version might prepare the customer brief and draft from approved records. A salesperson checks it before any quote is sent.

Agree on what would stop the trial: missing inputs, excessive correction time or an unresolved ownership problem. Keep the existing process available while learning. Once you have chosen the task, use a measured pilot to test whether the estimated saving survives ordinary work.

Give the recovered time a purpose

Decide what the team should do with the time it gets back: handle complex requests, follow up sooner or reduce an overdue queue. That makes the project easier to judge than a count of AI-generated drafts. Also record if work simply moves from one employee to another.

My ambition is to remove at least half of suitable repetitive office work. That is a target to test, not a selection rule that makes every smaller improvement worthless. Choose a useful first workflow, check the complete result and expand from evidence.

Bring three candidate tasks, a few real examples and the systems involved. That is enough to begin choosing what to build. A clear first decision is often more valuable than a long list of tools nobody has fitted into the work.