An AI Route Gave a Home Health Aide Trips No Bus Could Make

Renee kept a notebook in the outside pocket of her work bag. The left side of each page held the visits assigned by the scheduling app. On the right, she wrote the buses she needed, the wait between them and how late she reached each client.

One line appeared often: 18 minutes in the app, 54 minutes in the notebook.

Renee, a composite based on home care workers in Midwestern suburbs, earned $15.25 an hour helping older adults and disabled clients remain at home. She heated meals, helped with bathing and recorded changes in mobility. Most weeks, she was paid for about 28 hours.

Travel between clients was unpaid unless the company classified it as part of a continuous shift.

She did not own a car. Her suburb had bus service along its commercial roads, but many clients lived beyond those lines. A trip of six miles could require one bus toward the city, a transfer near a shopping center and another bus back into the suburb. Missing a connection added 30 minutes.

Before the company introduced the optimizer, a coordinator assembled routes. The coordinator knew Renee rode the bus and tended to keep her near two main roads. There were gaps in the day, and some available visits never fit. Still, Renee could usually look at a route and know that it was possible.

The new system filled those gaps.

The route on the notebook page

During one week that fall, the app placed a short personal-care visit between two clients who lived 12 miles apart. The system allowed 18 minutes for travel. Renee copied that number into her notebook, checked the transit planner and wrote 54 beside it.

The first client needed help getting out of bed and preparing breakfast. Leaving early would have meant skipping part of the visit. Renee stayed until the assigned work was done, walked to the bus stop and watched the transfer leave while her first bus waited in traffic.

By the time she reached the second home, the client had been waiting for medication help and a meal. The client called the agency. Renee marked herself present in the app, which recorded her location and arrival against the scheduled start of the visit.

Her phrase for the mismatch was short.

“It says 18 minutes. I write 54.”

The optimizer was not consulting the public bus timetable and making a poor choice among transit routes. It was doing something different. It predicted travel time from past trips recorded through workers’ phones, then used those predictions to arrange more visits within the available day.

Most of the workers in that data drove. Their phones moved between homes along direct road routes, and the system learned that a six-mile trip usually took about as long as a car trip. Renee’s phone also traveled on roads, often at traffic speed while she was on a bus, so location data alone did not reliably identify her travel mode. The company profile had no setting that made bus schedules a firm constraint.

A conventional schedule could have contained the same bad estimate. The difference was scale and repetition. The optimizer generated a new sequence whenever visits changed, using learned travel predictions to rank thousands of possible combinations, and it kept selecting narrow gaps that a coordinator familiar with Renee would have rejected.

It also treated the prediction as a standard for her work.

When a prediction becomes a score

The worker dashboard showed Renee an overall performance score built from several signals, including whether she accepted assignments and whether she arrived within the expected window. She could see the components but not how heavily each one counted. Over six weeks, her displayed score fell from 91 to 78.

Declining an impossible visit protected the client from waiting, but the refusal could lower her acceptance measure. Taking it exposed her to a late arrival. Calling the scheduling staff sometimes helped, though assignments changed outside the periods when a coordinator was available, and another optimized route could appear the next morning.

The score mattered even without an automatic firing rule. Workers with stronger records were considered first for preferred hours and steadier clients, according to the explanations Renee received from her manager. When her score dropped, she noticed that open visits disappeared from her app sooner. She could not tell whether another worker had claimed them or whether the system had ranked her lower.

Her notebook became a second record of the job. The app had a scheduled arrival, a location stamp and a late mark. The notebook had the missed transfer, the 30-minute wait and the bus stop that sat nearly a mile from the client’s home.

The system could learn from errors only if the error was labeled as a travel prediction problem. Renee’s late arrival did not necessarily provide that label. From the company’s view, the planned trip took 18 minutes and the worker arrived late. Her bus ride was ordinary location data surrounded by a performance outcome, which made it possible for the system to treat the failure as hers rather than as evidence that its travel model was wrong.

That distinction changed her work from caregiving into continuous route repair. She checked assignments against a transit app before accepting them, saved screenshots when the predicted trip differed from the bus journey and asked clients whether a visit could move. None of that appeared as paid care.

The hours the system found

Renee did not want the old schedule back in full. The optimizer found visits that a busy coordinator might have missed, and during its first month her paid care rose by about five hours a week. A full day displayed in one place was easier to understand than a string of calls about openings.

For workers with cars, the travel estimates were often close enough. The company also filled more visits after cancellations, which mattered to clients who otherwise went without help. Renee could see why her manager kept using it.

The extra hours did not all become extra money. During one pay period, she spent $46 on rides after two routes broke down between buses. At $15.25 an hour, that expense consumed about three hours of gross pay.

On other days, she refused the ride, arrived late and watched the score fall.

Her manager eventually added a note saying she used public transit. The note was visible to scheduling staff, but the optimizer did not convert it into bus routes, transfer buffers or a smaller travel area. A person could read the note after a problem appeared. The machine continued to build the route from its own estimate.

For a while, Renee developed a workaround. She accepted visits near the main bus line and delayed decisions on the rest until she could compare them with her notebook. That made her acceptance record look hesitant, but it reduced the number of clients left waiting. She also learned which assignments were risky from their mileage and direction, even before checking the transit planner.

Eight months after the system arrived, the company began testing a broader travel buffer for workers who did not drive. It was a blunt adjustment rather than a transit model. A predicted 18-minute trip might receive additional time, but the system still did not know that one missed connection could add half an hour.

The change helped on routes with frequent service. It did little for clients beyond the last useful bus stop.

Renee kept the notebook. On a later page, the app allowed 22 minutes between two visits. Her transit plan required 47, including a transfer. She declined the second visit, wrote both numbers on the same line and left the space for an arrival mark empty.

Questions people ask

How can an

AI route be possible by car but impossible by bus?

A route optimizer may predict travel from phone location histories or past driving times without using transit schedules as constraints. If most recorded workers drive, the model can learn car-like travel times even for an aide whose phone moves along the same roads on a bus.

Why would refusing an impossible assignment affect a worker’s score?

In Renee’s dashboard, acceptance and punctuality both contributed to the displayed performance record. Declining protected her from a late mark but weakened one measure, while accepting exposed her to another, because the scoring system did not record an infeasible transit connection as a separate cause.

Did the scheduling system give her more work?

At first, it added about five paid care hours a week by fitting visits into gaps. Some gains were offset by unpaid route checking and $46 in rides during one pay period, so a fuller schedule did not always produce the increase shown on the screen.

What record showed that the travel estimate was wrong?

The app preserved its prediction and her arrival location. Renee’s notebook preserved the missing context: 18 minutes allowed, 54 minutes by transit and a 30-minute wait after a missed transfer.