AI Belongs in the Workflow

Here is a pattern to consider. A company buys AI licenses, runs training, and a few months later usage has quietly dropped by everyone except a handful of enthusiasts. The tool worked. People used it during the pilot. Then it faded.

The usual reason is that AI was added next to the work instead of becoming part of the workflow. That means the AI was not well suited to supporting that particular type of work, or it was cumbersome to use. The cumbersome point happens more often than people realize. If using AI means leaving the software system where the job happens, opening another window, pasting something in, reading the answer and pasting it back, the benefits of AI might very well be erased by the extra steps required. People do not stop using it because they dislike it. They stop because they don’t see a benefit in that situation.

The alternative is to start with a workflow rather than a tool. Take one process that matters, such as quoting, onboarding a new customer, or closing the month’s books. Write down every step, who does it, what triggers it and where it waits. The waiting is usually the most useful part, because delay is where cost accumulates in most businesses.

Then decide, step by step, where AI belongs. It will not be everywhere. Analysis of the companies furthest along with this found that even in workflows built around AI, AI accounted for only 18 percent of the steps.¹ The rest ran on ordinary automation: rules, approvals, system connections and moving data between applications. Reserving AI for the steps that genuinely need reasoning also cost 71 percent less to run than routing every step through a model.¹

That is important. The companies getting the most out of AI are using less of it, more deliberately, inside processes they took the trouble to redesign.

There is a trade-off at play here. Redesigning a workflow takes longer to show a return than just picking up a new AI tool. That’s because redesigning processes can be difficult, and efficiencies often emerge over time. But when the new system is up and running, the processes and the AI tool are designed for each other. The system works fine.

Pick one process. Map it end to end. Mark every point where the work waits on a person, and start there.

Sources

¹ Zapier, AI Workflow Index, 2026, based on analysis of the top 25 percent of companies ranked by adoption of AI workflows.