It is a quarter past eight in the evening at a regional distribution centre. The web orders for tomorrow have just dropped into the transport plan, two stores have cancelled their deliveries and a driver has called in sick. The new route optimisation AI produces a plan within a minute. The logistics manager looks at it, frowns, and starts moving stops around by hand, because the plan sends a heavy truck down a street it cannot enter and promises a delivery window that one customer stopped accepting months ago.
Nothing is wrong with the algorithm. It did exactly what it was asked to do with the data it was given. The problem is that much of what makes a route work never made it into that data. It lives in the head of the planner, in a spreadsheet on a shared drive, or in a note from a driver.
That gap between what the optimiser knows and what the planner knows is the main reason route optimisation so often delivers less than the demo suggested. The MetroMinds Library shows where the gap comes from, and what you can do about it before you sign the contract.
The highest value and the lowest feasibility in one place
The Library contains 60 use cases labelled routing and network optimisation, spread across all 13 industries, from freight and public transport to healthcare and government. Every one of them scores the maximum of 5 on business value. That is rare: across all 4,980 use cases, the average business value is 3.82.
The same 60 use cases also score the maximum on risk, because a wrong route costs money, time and sometimes safety. And none of them scores higher than 2 on feasibility. The average is 1.58, against 3.30 for the Library as a whole. Very few types of AI use case combine such a clear prize with such a demanding delivery.
The prize is real. Eurostat reported that in 2020, 20 percent of road freight vehicle kilometres in the EU were driven by empty vehicles [1]. Every percentage point of empty running that better planning removes is fuel, driver time and capacity that comes back. That is why route optimisation looks so attractive in a business case, and why it deserves a careful look at what stands in the way.
Your data is reachable, so why do the plans fail?
Here the Library contains a surprise. The data side of these use cases scores well. Data access and trust reaches 5 out of 5 for routing and network optimisation, higher than the Library average of 4.26. Integration readiness is around average, at 4.03. Order systems, vehicle data and traffic feeds are usually available and can be connected.
What makes these use cases hard is the work itself. Decision complexity, exceptions and ambiguity, and context volatility all reach the maximum score of 5. Collaboration intensity jumps from a Library average of 2.89 to 4.85, because a route decision often involves the warehouse, the customer service team and sometimes compliance.
So the systems hold plenty of data, just not all of the right kind. The optimiser can read the orders and the trucks. It cannot read the planner's knowledge that a particular loading dock closes early on Fridays, that one customer only accepts deliveries from the side entrance, or that a new driver should not get the route through the old town.
Planning data is more than master data
The use cases in the Library are explicit about this. For a typical multi-stop delivery route, the AI needs "accurate addresses, service windows, asset availability, travel times, and capacity limits", and the Library warns that "missing constraint data creates infeasible recommendations". The business case, it adds, "gets harder when planners keep constraints in notes, spreadsheets, or personal memory".
Addresses alone can make a large difference. A research brief from the Universal Postal Union found that "address quality emerges as one of the most significant determinants of first delivery attempt success rates", and that poor address quality can reduce delivery success rates by as much as 14 percent [2]. The same logic applies to every constraint a route depends on: time windows, vehicle restrictions, driver qualifications and customer preferences.
So when you prepare a route optimisation project, the most important inventory is not your systems. It is the list of rules your planners apply that no system knows about. In many organisations nobody has ever written that list down.
What missing constraints cost you
Besides producing bad routes, missing constraint data quietly lowers the benefit in two ways, both of which show up in the business case.
First, every route with incomplete constraints has to be rebuilt by hand, so it delivers little or no time saving. Second, planners who see infeasible plans stop trusting the optimiser and start checking every route, including the ones that were fine. The example below shows how quickly the benefit shrinks.
This is also why the time saving should be the headline of the business case. The Library treats fewer wrong routes and higher planner throughput as separate value streams that need their own baseline, because the rework they remove overlaps with the planner minutes already counted. Our article on AI business cases that count the same savings twice explains why that matters.
How to get your planning data ready for route optimisation
You can start well before any optimiser goes live, and most of the work does not require technology.
- Sit with your planners for a week and write down every rule they apply that is not in a system: access restrictions, customer preferences, driver qualifications, dock hours.
- Check address and location quality for your most frequent delivery points, and fix the ones that fail geocoding.
- Decide which constraints are hard rules and which are preferences, because an optimiser treats them very differently.
- Agree that every override in the new system gets a reason. Recurring reasons are your list of missing constraints.
- Pilot on routes where the constraint data is already good, so the first experience builds trust instead of damaging it.
The override log deserves special attention. In the Library use cases, the AI never releases a route on its own. The logistics manager publishes, revises or returns the plan, and the reason for each change is recorded. Over time, that log turns the planner's personal knowledge into data the optimiser can use.
Seeing the planning gap before you buy
In MetroMinds, every use case carries its , and profile, so you can see early that a route optimisation idea is hard because of the work, not because of the data. The value and feasibility matrix makes the trade-off visible: high value and low feasibility usually means a that needs preparation before it can scale.
The AI Readiness Scan then shows how ready your organisation is for the portfolio you want to execute, including data, governance and the people side. For route optimisation, that is often where the biggest gap sits.
Is your planning data ready for the optimiser you are about to buy? Take the AI Readiness Snapshot
Sources
- Eurostat, "A fifth of road freight kilometres by empty vehicles", news item, 10 December 2021 (2020 data).
- Universal Postal Union, "Enhancing e-commerce customer experience: exploring the hidden role of addressing", UPU Think Tank Brief No. 2/2025.
- MetroMinds Library, 4,980 AI use cases with benchmark ranges based on real-world implementations, August 2026.
