AI workflows, lesson one: get the process clear before automation

Many people come to AI with “automate this for me.” Press a little further, and it often turns out the task has three ways of being done, two exceptions, and a final step that depends on one colleague's memory. The problem isn't that AI isn't smart enough — it's that the organization hasn't yet made its own work clear.

A process isn't a diagram drawn for managers

A process that can actually be run has to answer at least five questions: when the work begins, where the input comes from, who makes the call at each step, what to do when an exception arises, and where the result is kept. Leave one out, and people fill the gap with unspoken understanding; hand it to a system, and that understanding doesn't turn itself into rules.

Find the “usually”

Don't try to handle every case at once. Pick out the most common sixty to eighty percent and write it up as a main path that runs end to end; then flag the exceptions one by one: incomplete data, an amount over the limit, a customer with a special arrangement, another department that hasn't replied yet. The main path is the first version of automation; the exceptions decide when to stop and hand off to a person.

Automation doesn't speed up chaos; it forces a team to admit where things have quietly been held together by people all along.

Separate rules from judgment

“Send a reminder after three days without a reply” is a rule; “whether it's the right moment to chase this particular customer” is a judgment. The first can be handed to a system; for the second, AI can supply context and suggestions, but someone still has to own the call. Keeping the two apart is what stops you from dressing up relationships and risks that shouldn't be automated as just another conditional.

Documentation pays management back too

Once a process is written down, new hires know what to do, managers can see the bottlenecks, and departments can align on who is responsible. Even if AI is never adopted in the end, the act of sorting it out makes the work better. That's also why we're in no hurry to start from a tool: tools can be swapped out, but a clear way of working stays.

Start with one task that happens every week and involves no more than two or three roles. Walk through it on a single page, then discuss whether to add AI. The pace of adoption often starts to pick up right from this step.