The Route Board Sent Tow Crews Into Floodwater. She Overrode It.

The dispatcher kept a notebook beside the route board. She had started it for ordinary reasons, mostly to remember which repair shops accepted late arrivals and which apartment lots required a smaller truck. During a week of heavy rain, it became a record of roads the software could not see.

She worked nights for a regional towing service covering several counties. After 14 months on the job, she earned $22.40 an hour and could manage most shifts without calling her supervisor. The calls ranged from dead batteries to crashes requiring a flatbed, and the company’s routing system ranked available drivers according to who it predicted could arrive soonest.

The system had been useful in dry weather. It could compare a driver finishing one job with another driver who appeared closer but still had a vehicle loaded. It matched truck type to the call and recalculated as crews moved. Before the company installed it, dispatchers relied more heavily on radio updates and their own memory of where each driver had gone.

During the storm, the board assigned a driver to a disabled car on the far side of a low crossing. The route passed through a road that had taken on water. The driver reached a barricade, reported it, and turned around.

The dispatcher rejected the route and moved the call to another truck approaching from a different direction. Within the same shift, the board proposed the flooded road again.

She opened the notebook and wrote down the call type, the suggested road and what the driver found there. She left out customer names. By the end of six weeks, the notebook held 17 weather-related overrides.

What the route board predicted

The board did more than display a map. A machine-learning model estimated which driver could complete each call with the least delay, using recent GPS positions, past travel times, truck availability and the kind of equipment attached to the job. Its rankings changed as vehicles moved and new calls entered the system.

That prediction was the reason the story could not be reduced to a bad spreadsheet. The board was not repeating one stored route. It kept generating new rankings from live vehicle data, then placing the highest-ranked assignment where a dispatcher was expected to accept it.

The model had plenty of evidence that the low crossing was usually fast. Months of completed trips showed trucks moving through it without trouble, while the longer road added distance and often passed through heavier traffic. What it lacked was a dependable local feed showing temporary flood closures. Reports from drivers arrived by voice or message, and those reports did not become structured road data for the next prediction.

Sparse traffic could also look less alarming to the system than it did to a person. When most drivers avoided a flooded road, there were fewer fresh GPS traces from that stretch, but the board did not translate that absence into a closure. Historical travel times still made the road attractive.

The display did not show how much of its recommendation came from old trips or whether it had enough current traffic to support the estimate. It offered a ranked route with an arrival prediction. To the dispatcher, the certainty was part of the problem. A road shown at the top looked settled even when the useful fact was missing.

She began checking county closure notices and driver messages before accepting routes near creeks and underpasses. That work took place outside the board. On quiet nights it added little. During storms, each call required her to compare the machine’s prediction with information scattered across another screen and her notebook, while drivers waited for an assignment.

The score behind the board

The company also used the routing data to score dispatchers. Her dashboard combined several measures, including how often she accepted recommended assignments and whether predicted arrival windows were met. An override counted as a departure from the board even when a driver had already reported water or a barricade.

Her score had usually stayed near 94. After the storm period, it fell to 81. The company considered scores below 85 during shift reviews and when deciding who received preferred schedules, though managers could read notes attached to individual calls.

That created a choice inside nearly every disputed route. Accepting the recommendation preserved her adherence rate but could send a driver toward a road she believed was unsafe. Rejecting it created a record that looked inefficient until someone opened the call notes and compared them with the weather.

She described the choice as taking the bad number. The number belonged to her, while the route belonged to the system.

Her manager initially saw the overrides as a pattern before seeing what connected them. The dashboard grouped rejected recommendations in one measure. It did not place a closure notice beside each rejection, and it did not distinguish between a dispatcher who preferred her own route and one responding to a hazard reported from the road.

The notebook did. One entry matched a driver’s GPS path approaching the crossing, stopping short and doubling back. Another lined up with a county notice issued after the board had calculated the route. The pages gave her manager a way to read the lower score against events the performance dashboard had excluded.

There was another layer to the record. Drivers knew the company could see truck locations for dispatch, but the same GPS traces also became evidence about route compliance, delays and stops. The dispatcher’s clicks were logged as well. A location signal gathered to answer where a truck was could later help judge whether the driver or dispatcher followed the system’s preferred plan.

No one had to accuse her of making a bad call. The score moved first.

A partial change

After the notebook was reviewed, the company stopped treating documented weather overrides as ordinary rejections during severe storms. A supervisor could mark a group of affected calls so they did not carry the same weight in the dispatcher’s weekly score. Her earlier score was reviewed, though the dashboard did not erase every decline.

The routing tool also gained a temporary exclusion that allowed dispatchers to keep a reported road out of recommendations for the remainder of a shift. It was a narrow change. Someone still had to learn about the hazard, decide that the report was credible and enter it before the model generated another route.

The dispatcher kept using the board. On normal nights it found combinations she might have missed, particularly when several drivers finished jobs close together and the nearest truck was not the right truck. She did not want to return to holding the whole service area in her head.

She also stopped treating the first-ranked route as the answer. During rain, the notebook stayed open beside the screen, with each override written on a new line. The board continued to calculate from the data it had. She worked from the part it did not.

Questions people ask

Why did the routing system keep choosing flooded roads?

The model ranked routes from historical travel times, live truck locations and job requirements. It did not receive dependable local closure data, so a road that was normally fast remained highly ranked. With few vehicles traveling through the flooded section, there were also too few fresh GPS traces to correct that history.

How did rejecting an unsafe route lower the dispatcher’s score?

Her performance measure included acceptance of the board’s recommendations. The dashboard initially counted a safety-based override much like any other rejection, unless a manager opened the notes and reviewed the surrounding events. During six weeks of storms, her score fell from about 94 to 81.

Was GPS data used only to locate tow trucks?

No. The company used truck locations to predict arrivals and assign calls, but those traces also showed route deviations, stops and delays. Combined with the dispatcher’s logged choices, the same data helped measure whether workers followed the automated plan, even when they departed from it for safety.

Did the company fix the route board?

It added temporary road exclusions and changed how some documented weather overrides affected scores. The tool still depended on a person to report a closure before the next recommendation. On rainy nights, the dispatcher continued recording each disputed route in the notebook beside the board.