Your first AI workflow should be the most boring job in the building
Ask a founder where AI should go first and the answer is usually the most exciting problem in the business: forecasting demand, answering every customer question, writing the quotes. Those are real problems. They are also the worst place to start.
The better first project is the job nobody wants to talk about. The one your team does the same way every day, grumbles about, and could explain in five minutes. Boring is not a compromise here. Boring is the qualification.
Why the exciting projects stall
Exciting problems are exciting because they are hard. They involve judgment, exceptions, and knowledge that lives in one or two people’s heads. That is exactly the material AI handles worst when nobody has written it down.
When a first AI project tries to automate judgment, one of two things happens. The tool makes confident mistakes that a person has to catch and fix, which erases the time it was supposed to save. Or the team quietly stops trusting it and goes back to the old way. Either result teaches the business the wrong lesson: that AI does not work here.
The problem was not the technology. It was the choice of job.
Four tests for a good first workflow
Before you point any AI tool at a task, run it through four questions. A strong candidate passes all four.
- Is it written down? If the steps exist in a procedure your team actually follows, the tool has something to follow too. If the steps live only in someone’s head, write the procedure first. You can’t optimize what you haven’t defined, and you can’t automate it either.
- Does it repeat the same way? Look for tasks that happen daily or weekly with the same inputs and the same shape of output: sorting incoming requests, drafting routine follow-up emails, summarizing a daily report, pulling the same fields out of the same kind of document.
- Is the data clean and in one place? A task that draws on one organized source is a good candidate. A task that requires hunting through inboxes and three versions of a spreadsheet is not ready yet, and no tool will fix that for you.
- Is a mistake cheap and easy to catch? Start where an error costs a few minutes of rework, not a lost customer or a wrong order. A person should be able to glance at the output and know whether it is right.
Most businesses find two or three tasks that pass. Pick the one your team complains about most. Relief is the fastest way to earn trust in a new tool.
Keep a human in the loop, on purpose
A good first workflow is not “set it and forget it.” The AI drafts, sorts, or summarizes. A named person reviews and approves before anything leaves the building or changes a record.
That review step is not a sign the tool failed. It is how you learn where it is reliable, with low stakes and real examples. Over a few weeks, the reviewer sees which outputs are consistently right and which ones need correcting. That record tells you whether to widen the tool’s job, adjust the procedure it follows, or leave it where it is.
It also protects your team’s knowledge. If the tool goes down or the vendor changes its terms, the people who do the work still know how to do it, because the procedure is written down and they have been checking the output all along.
What a good first project leaves behind
The real payoff of starting boring is not the task itself. It is what the business builds on the way.
To pass the four tests, you had to document a procedure, put its data in one place, and decide who owns the review. Those are the same foundations every larger AI project needs, including company-specific AI: an assistant that answers from your own approved documents. The first workflow becomes a small, working model of how the business adopts technology: define the process, structure the data, then automate with a person watching.
That sequence also keeps you in control. When your procedures and your data live in files you own, the AI tool is replaceable. You can switch tools later without rebuilding what the business knows.
Where to start this week
Walk through a normal day and list every task someone does the same way more than twice a week. Run each one through the four tests. Cross off anything that fails the first test, then write that procedure down anyway, because it is your next candidate.
What remains is your shortlist. Pick the most boring item on it, name the person who will review the output, and start there. The knowledge your team writes down is what your own AI runs on, and the most boring job in the building is where you prove the foundation is ready.