Stop Writing Prompts. Start Designing Workflows.

A team lead tells you: "We tried ChatGPT for three months and gave up, it just wasn't that useful." Nine times out of ten, here is what actually happened: someone typed a question, got an answer, closed the tab, and never came back. That is not use, that is a single test drive. And it explains most of the disappointment we hear when we start a training with a new client.

Most people treat AI like a smarter search engine. One question in, one answer out, and they judge the whole technology on that single exchange. But the value of AI rarely lives in one prompt. It lives in the sequence: draft, check, refine, reuse. A workflow, not a transaction.

We see this constantly in workshops. Someone asks Claude or ChatGPT to "write a client proposal" in one go, gets something generic, and concludes the tool is shallow. The people who get real value do something different: they break the task into steps, feed the model context at each stage, check the output against something they know is true, and only then move forward.

Most people conclude AI is inconsistent or overhyped. What actually matters is that AI is a step in the process, not the complete process. The gap is not model capability, it is workflow design: knowing which steps to hand to the model, which to keep for a human, and where to place a checkpoint before something wrong ships.

Four practical implications for an SME or a leadership team.

Map the task before you open the tool. Write down the steps a competent employee would take to do this work well, by hand. That sequence is your workflow, AI slots into individual steps, not the whole task.

Determine who is responsible for the task and when it’s controlled. Put a human checkpoint after the step with the highest cost of being wrong, not at the end. If a mistake in step 2 of 5 sends the wrong instructions to step 3, that is where the checkpoint belongs, not after step 5.

Save what works as a reusable prompt sequence, not informal knowledge. If someone on your team develops a sequence that reliably produces good client emails, write it down as a repeatable process with the context it needs. Otherwise, that knowledge leaves when they do.

Judge value over ten uses, not one. A single test tells you almost nothing about whether a workflow is sound. Ask a team to run the same workflow ten times over two weeks, then evaluate.

This is squarely a Knowledge problem in the MAKIA framework: not whether teams have access to a tool, but whether they know how to sequence it into real work.

The question worth asking at your next AI review is not whether the model is good enough. It is whether you designed a workflow, or tested a chatbot. Most teams that give up on AI never actually left the second category.

Try this week

Pick one recurring task you already do with AI in a single prompt: a summary, an email, a report. Write down the 3 to 5 steps you would take doing it by hand. Rebuild the AI version as that same sequence, with one checkpoint in the middle. Compare the output to your usual one shot result.