Watch a day in the life before you scope the AI

An operations manager at a home-services company had a tidy plan for an AI scheduling assistant. She had the requirements written up. The tool would look at the day’s jobs, the technicians’ locations, and the traffic, and suggest an optimal route and order. On paper it was clean. Everyone who reviewed the spec agreed it made sense.

Before signing off, she did something the vendor did not expect. She sat next to her lead dispatcher for a full day and watched him work. No interviewing, no survey. Just a chair beside his desk and a notebook.

The day rearranged everything she thought she knew.

The real workflow was full of things no spec mentions

The spec assumed the dispatcher started from a clean list of jobs and assigned them. What actually happened was that he spent the first hour on the phone. A customer needed to move to the afternoon because of a school pickup. A technician texted that he was stuck at a job that turned out bigger than quoted. A parts supplier was running late, which meant two installs could not happen before noon no matter what any algorithm said.

The dispatcher was not solving a routing problem. He was solving a constant stream of human exceptions, and the routing was the easy part he did in the gaps. The tidy spec had captured the five percent of his job that was mechanical and missed the ninety-five percent that was judgment.

She saw him do things he would never have thought to mention in an interview. He always put the newest technician’s last job of the day close to home, because that kid got flustered when he ran late and it kept him calm. He knew one particular customer would complain no matter what, so he scheduled her for mid-morning when the crew was fresh. None of that was written anywhere. All of it mattered. An optimizer that ignored it would produce routes that were mathematically better and operationally worse, and the dispatcher would quietly override it until he stopped using it at all.

A data lead at a hospital system told me a version of this from a different industry. Her team built a tool to flag patients for follow-up calls. It worked in testing. In the clinic it went unused, because the nurses did their triage from a paper list they annotated by hand during rounds, and the tool lived on a screen they never opened at that hour. Nobody had watched a day of rounds. The tool fit the workflow the team imagined, not the one that existed.

Shadowing shows you the friction and the workarounds

A day in the life gives you three things a document cannot. You see the exceptions, the moments where the neat process breaks and a human improvises. You see the workarounds, the sticky notes and side spreadsheets people built to survive the official system. And you see the timing, the actual moment a decision gets made and whether anyone would even be looking at your tool then.

Those three things decide whether an AI feature gets used or gets ignored. You will not find them in a requirements meeting, because the people in that meeting describe the process they think they follow, which is the clean version. The messy real version only shows up when you watch.

The operations manager rescoped the whole project after that day. The AI stopped being an optimizer that owned the schedule and became an assistant that handled the mechanical routing in the background while surfacing the exceptions to the dispatcher faster. It fit his day instead of fighting it, and he actually used it.

How we approach it at Density Labs

Part of every AI Opportunity Assessment, our two-week fixed engagement at $2,500, is time spent beside the person whose work the AI will touch. We watch a real shift, note the exceptions and the workarounds, and map the moment the decision actually happens. It is unglamorous and it is the single best predictor we have of whether a feature will get used. The spec tells you the process someone wishes they ran. The day tells you the one they run.

Before you scope an AI feature, spend a day in the chair next to the person it is for. The workflow you build against should be the real one, not the one on the slide.