You built the best routing system for a logistics company. The math is perfect. The system calculates the shortest path, least fuel, fastest delivery. In the morning, you hand drivers their routes.
By afternoon, nobody followed the system.
"I've known this road for ten years," says one. Another left without a word and took their own route. And here's the interesting part: most of the time, they're right.
Tacit Knowledge — What Can't Be Put Into Words
This is called tacit knowledge in the literature — implicit, unspoken knowledge. The kind that accumulates in a driver's mind after years of working a specific area. It can't be verbalized or formalized, but it's extraordinarily valuable.
This knowledge includes:
- A narrow street that appears on the map but a truck can't pass through
- A dead end — the GPS knows it, but it closes at certain hours
- A steep hill — the shortest route means climbing 150 steps, but entering from above is 3 times faster
- Parking — that street is packed at 9am, the driver knows this
- The customer won't be home at that hour — knocking is pointless
- Seasonal closures, construction sites, school dismissal times
None of this exists in the optimization model. Because it isn't data yet. It lives in the driver's mind — not inside the model.
MIT researchers put it clearly: "Experienced drivers know which roads are difficult, when traffic is bad, where parking can be found. For reasons that are hard to incorporate directly into an optimization model, they deviate from theoretically optimal routes."
Türkiye Makes It Even More Complex
Unplanned urbanization. Irregular street layouts. Changes that don't make it onto maps. In countries like Türkiye, tacit knowledge becomes even more critical. What an Istanbul driver carries in their head is far richer than their counterpart in Europe's orderly city networks — and far harder to formalize.
What Did Amazon and MIT Do?
In 2021, Amazon put this exact problem on the table. Together with MIT, they launched a major competition: the Last Mile Routing Research Challenge. The goal wasn't to replace drivers. It was to incorporate the tacit knowledge experienced drivers had built up over years into the model.
9,184 real Amazon driver routes were provided as data. 2,285 researchers from around the world participated. What did the winning team do? They hybridized mathematical optimization with driver data.
What Did UPS Do?
UPS's ORION system processes millions of real delivery data points using graph-based optimization algorithms. Over 200,000 route options are evaluated for each driver.
But the most critical part wasn't technical. Drivers didn't trust it. UPS didn't force the system — they made it transparent. Drivers could see why a particular route was being suggested. As trust built, adoption followed.
The result: a daily reduction of 6–8 miles per driver. 100 million miles and 10 million gallons of fuel saved annually. $300–400 million in yearly gains.
We're Not Discarding Optimization — We're Combining It With Learning Systems
There's a critical point here: pure optimization is indispensable. It carries mathematical guarantees. But it's static — it doesn't know what it doesn't know.
Hybrid approaches that combine the two are becoming inevitable. And the prerequisite for this combination: clean data. A consistent, accurate historical record with anomalous days filtered out. A machine will learn from dirty history too.
Would you like to discuss this topic? You can reach me at zuhalkartal.com