How to choose the first workflow to automate with AI

Most AI projects that stall do not fail because the technology is weak. They fail because the first workflow was the wrong one: too broad, too risky, or impossible to measure. Choosing the first workflow well matters more than choosing the model.

In our projects we usually start with the process, not the tool. Here is a simple method for narrowing a long list of ideas down to one.

Start with real, repeated tasks

Ask each team to list the tasks they repeat every week, described in concrete terms: "copy supplier invoices into the ERP", "answer WhatsApp enquiries about availability after hours", "build the Monday report from five spreadsheets".

A goal such as "use AI in customer service" is not a workflow yet. A workflow has a trigger, a set of steps, a person who does it today and a result you can see.

Five questions for every candidate

• Volume. How often does it happen? A task done hundreds of times a month repays automation far sooner than one done twice a quarter.
• Rules. Can someone explain how it is done? If an experienced employee can describe the steps and the usual exceptions, a system can be built to follow them. If every case is a judgement call, it is not a first project.
• Data availability. Are the inputs already digital and reachable, such as emails, PDFs, a CRM or a shared drive? If the information lives in people's heads or on paper, the first project quietly becomes a data project.
• Risk. What happens when the system gets something wrong? A drafted reply that a person approves is low risk. Changing records or sending money without review is high risk. Start at the lower end, with a human approval step where it matters.
• Measurable outcome. Can you count the result? Hours spent, response time, error rate, backlog size. If you cannot measure today's situation, you will not be able to show what changed.

A simple scoring approach

Score each candidate from 1 to 5 on each question, giving a higher score for lower risk. Add the five scores for a total out of 25.

For example, a team comparing three ideas might score them like this (illustrative scores, not a benchmark):

• Supplier invoice entry: volume 5, rules 4, data 4, risk 4, measurable 5, total 22.
• After-hours WhatsApp enquiries: volume 4, rules 4, data 3, risk 4, measurable 5, total 20.
• Contract review: volume 2, rules 2, data 3, risk 2, measurable 3, total 12.

The numbers are not precise, and they do not need to be. The value of the exercise is that it forces the team to discuss each factor openly. When two candidates score close together, choose the one with a clear owner and the shortest path into the systems you already use.

Warning signs to watch for

• Nobody owns the process end to end.
• The workflow depends on systems that do not yet connect.
• Success requires the AI to be right every single time.
• The people who do the work today were not asked.

Your next step

Once you have a winner, write down the baseline and the success measure before anything is built, and keep the first version narrow. A focused pilot typically runs 8 to 12 weeks and ends with something running in production, not a slide deck.

If you would like a second opinion on your shortlist, book a free 30-minute consultation, or describe the challenge in the AI solution designer on our homepage to see how we would approach it.