The Override Tax

The thirteen years, in full

UPS started in 2003 with package tracking: the foundation that put package, address, and route data in one place so routing software could later run on top of it. The routing came after, and the gap between those two things is the whole story.

The routing worked in the lab early. In practice it produced routes drivers would not drive, because it solved each day from scratch and a customer with a standing daily pickup got a different time every morning. UPS went back and rebuilt it around how drivers actually work rather than around the math. A version that balanced the best route against keeping the day recognizable was working by roughly 2007. Testing then ran for years rather than months — prototypes at a handful of sites through 2010 and 2011, trial sites in 2012, a first large rollout to about 10,000 routes in 2013, then acceleration. By December 2015 more than 35,000 of the 55,000 US drivers were on it. Full deployment finished in 2016, the year it won the Franz Edelman Award.

The part nobody can write down

Somewhere at UPS, a long time ago, a man with a stopwatch decided a driver's steering wheel could have five and three-quarter inches of play in it. Not six. Not five and a half. You can still go and read the number, in a document called the 340 Methods that walks through a driver's day one step at a time. This is the most carefully written-down workforce in America, and has been since before the Second World War. The same company spent $250 million and thirteen years finding out it had no idea the school on the corner was locked until three.

The reason is worth getting exactly right, because it is the same reason your own handbook is no use for this. The 340 Methods describe what every driver should do identically — that is what a written procedure is for. The knowledge the routing system needed is the opposite kind. The gate on the corner shuts at ten. This dock takes twenty minutes in the morning and four at night. That customer signs for anything. None of it repeats. Every piece applies to exactly one route on one person's day, which means the return on writing it down is close to zero for everybody except the one driver who already knows it. So it never got written down. Not through carelessness — because it was never worth anybody's time.

Was it even a good bet

In 2003 somebody at UPS had to put their name on this: a quarter of a billion dollars, ten years of work, and not one package moving faster until year eleven at the earliest. Do it properly. Spread the $250 million across ten years. Start the savings in 2016 and hold them at $350 million a year. Discount it back at 10%. The spending is worth about $154 million in today's money. The savings are worth about $1.01 billion. Now put that against the only question that ever mattered — whether the thing would work at all.

That is the finding underneath the finding, and it is worth more than the routing lesson. Almost everybody spends their effort on the model, because a model is something you can work on at ten at night — and adoption is not.

Ordinary software on top of two things nobody else had

Working out the best order to visit a list of stops has a name — the travelling salesman problem, or in its operating version a constrained vehicle-routing problem. Mathematicians have been chewing on both since the 1930s and the algorithms are in a library book. Nobody at UPS was short of math. The rare part sat underneath, and it was two things. One was the data layer, built since the late 1990s on twenty years of scanners and handhelds. The other was the route rules — what time each gate shuts, which dock is slow in the morning, which customer needs a fixed slot — which never existed anywhere but in drivers' heads, and getting it out took years of engineers riding in trucks.

The logbook problem

In 2009 UPS put more than two hundred sensors on a package car — GPS, the handheld, seatbelt use, whether the back door was open, how long a driver stood at each address. The drivers' union told members to start keeping their own logbooks: write down your load, your route, anything unusual, so when management turns up with their version you have got yours. Read that as an owner rather than as a union story. The first instinct when people got measured was to start a second set of books. They were not trying to change what they did. They were trying to stop being the only party without proof.

Put consequences on a number and the number changes without the behavior changing. What people know goes further underground than it was before you started, and the board turns green while the business gets worse. Which is why the sentence on Monday morning is not a nice touch — it is the whole thing, and it is the one part of the four weeks that cannot be handed off, automated, or skipped. Anyone who cannot say nobody is in trouble and mean it should not start, because a clean list of what people think you want to hear is worse than what is there now. It looks fixed.

The math never worked until about 2009

Divide $250 million by 55,000 routes and it comes to about $4,545 a route. That is an allocation, not a quote — UPS spent most of it on fixed costs that would not shrink much for a smaller fleet, which is exactly the point. Run the allocation across other sizes and watch the idea die.

The test

Not every big-company play comes down. Three conditions decide it, and routing passes all three. One: the expensive part was a one-time cost of writing something down or computing it, rather than the ongoing cost of doing the work — one-time costs collapse, ongoing ones do not. Two: headcount-driven cost is small at your size, because change management costs the same per person whether you have twelve or fifty-five thousand, so twelve is where it is cheap. Three: currency is a light monthly loop rather than a standing engineering function — twelve rules reviewed once a month is a different animal from a system that needs somebody watching it full-time. Most plays fail the first test. Run any large-company case through all three before spending a month on it.

Service routes bleed harder than parcel routes

UPS runs 100 to 160 stops a day at roughly two minutes each, and the drive between them is most of the cost. Service routes are the opposite: twelve to twenty-five stops, twenty to ninety minutes of work at each, and durations that vary from visit to visit. That matters for three reasons, and all three run against the service operator. Job time dominates: when work at the stop is ten times the drive to it, a schedule slips through overrunning jobs rather than bad driving, and routing software optimizes the smaller half. Variance compounds: a filter change takes forty minutes or ninety and nobody logged which, so every job is another chance for the day to slide. Access is binding: a parcel can be left at the door, but a pool that needs the gate open or a vending machine inside a building that locks at five either happens in the window or happens tomorrow.

If you buy businesses for a living, this one hides in plain sight. You will get the financials and the loss will not be in them — no line item, no accrual. It arrives as jobs per tech per day sitting slightly below what the fleet should manage, and every comparable carries the same drag, so it reads as the industry rate. Then you will sit across from a dispatcher who could tell you exactly where it goes, and nobody in the room will ask her. So ask, in the management meeting: how many times a week does somebody change the plan after it is built, and how many jobs sit behind a typical change? Nobody has it written down, which is precisely why asking is informative.

The math, and where we guessed

The equation is how often × how deep × 2.6 minutes, giving minutes lost per tech per week; money is minutes × techs × 50 weeks ÷ 60 × $32 an hour. Depth is worked out as (stops − 1) ÷ 2, on the assumption that an override lands on average halfway through a route. Three of those numbers are ours rather than anyone's measurement. 2.6 minutes — what one scrambled job costs — prices extra driving only, not the phone call or the second trip, which is why we think it is more likely low than high. Depth as half your stops is a convenience; if your overrides cluster at the start of the day, your real tax is larger. $32 an hour is a round figure for wage plus taxes plus truck; yours is knowable in five minutes and you should use it instead.

What to buy, and what not to

Buy nothing in month one. That is the actual recommendation and the one most likely to be ignored. The four weeks run on the platform you already have, or on no platform at all. If you already own scheduling software, the boxes you need almost certainly exist and are empty — arrival window, service day, skill tag, preferred tech, default job time. Open them before you shop. Do not buy standalone routing software before the four weeks are done. It is the most expensive way to run the wrong rules faster.

Four ways this fails

Your routes are too thin. Under fifteen notes in week one means there is nothing behind your overrides to scramble; stop, and run the same four weeks against parts ordering or quoting. You count and punish. The team stops telling you rather than stops overriding, the wall chart goes green, and you lose both the minutes and the reason for them — the failure that looks like success. You turn field complaints into rules. A job booked at 40 that always takes 90 is a wrong number, not a rule; make it a rule and the wrong number becomes official. Dana never checks the rules. Skip the ten minutes and you have a black box she has no reason to trust — exactly what UPS spent three years and 700 trainers undoing.

What UPS proves: that the routing knowledge in a field operation lives with the people doing the work; that software optimizing against incomplete rules produces output the field will not follow; that extracting those rules was historically expensive enough to require enormous scale; and that adoption, not mathematics, decided the outcome. What it does not prove: that the four weeks works in a business the size of yours. Nobody has run it there, which is a real gap and the reason there is no reader case in here. Run it and you will know something we do not.

Axiom Research · The Operator's Brief · No. 01 · The Workup