How AI parcel volume prediction became part of Ampère’s daily planning
Ampère is growing fast. But how do you plan hundreds of collection routes a day when you know how many parcels you'll pick up, but not how much space they'll take up? Together with KickstartAI, Ampère developed a model that predicts parcel volume before collection. Today, the model is part of daily operations and has been used on millions of parcels, according to Ampère's Cas Beems and Thijs van Bergeijk.
For consumers, a parcel is just a parcel. In logistics, it's a different story. A hundred small boxes can fit comfortably in a van; a hundred large ones can throw an entire route off schedule.
Originally launched by bol, Ampère handles parcel collection and delivery for bol and its selling partners. Collection routes are especially challenging. Parcels need to be picked up from partners, processed through the network, and in some cases delivered the very same day.
"We enable premium logistics for selling partners," says Thijs van Bergeijk, Head of Planning & Forecast at Ampère. "A parcel can be collected in the morning and delivered that same afternoon. That means planning has to be right. If you're too tight, you end up making adjustments throughout the day. If you're too conservative, you're driving around with empty space."
Cas Beems, Product Owner & Business Analyst, puts it more simply: "If you're expecting two roll containers and there are ten waiting for you, your whole plan falls apart. If you're expecting ten and there's only one, you're transporting air."
From gut feeling to data
From the start of Ampère, these estimates were largely based on experience. That worked, until the operation grew bigger. "We had plenty of developers and technical expertise," says Cas. "What we didn't have yet was experience in turning data into predictive models and building the right data pipelines around them."
That's where KickstartAI came in.
"We started by mapping the operational process together," Cas explains. "What exactly happens when a collection order comes in? What data is available? What assumptions are planners making without even realizing it? Once we understood that, the question became: can we replicate that decision-making process using data?"
That approach led to the development of Quantile-Assured Volume Estimation, or QAVE: a data-driven model that uses historical parcel measurements to predict the expected collection volume for each selling partner.
"The data comes from dimension scanners above our sorting belts," says Thijs. "They measure the length, width and height of every parcel. By analysing those measurements, we can understand the average parcel size per partner, how that changes over time, and how much extra space is needed because parcels never stack perfectly inside a roll container."
Embedding into daily operations
What started as a pilot has become part of Ampère's day-to-day operations. Since September 2024, QAVE has been used structurally across the business and has generated predictions for millions of parcels. In practice, the model has proven highly accurate: on average, its prediction differs by only 13 litres from the actual collected volume.
Just as importantly, planners have embraced it. "Planners are naturally cautious, and for good reason," says Cas. "Their decisions have a direct impact on the operation. If you've been working a certain way for years, you're not going to blindly trust a machine learning model just because someone tells you to."
There was plenty of scepticism in the beginning. "Some planners literally said: if the model gets it wrong, we're the ones who have to clean up the mess," says Thijs.
To build confidence, Ampère and KickstartAI rolled the model out gradually. It started with a single depot, a single proposition, and the option to switch back if needed. Planners also received an override function, allowing them to intervene manually in exceptional situations. "That button mattered," says Thijs. "Not because people used it a lot, but because it gave them confidence."
Today, the conversation has changed. "If planners see something unexpected now, they're more likely to ask: can we improve the model? That's when you know people trust it."
Moving from one model to a full data science team
The project started as a way to improve volume predictions, but it ended up triggering something bigger. Ampère now has its own data science team, developing and improving predictive models across the organisation. "This project really was the spark," says Cas.
The ideas behind QAVE are now being applied to a range of other challenges. The team is working on models that predict parcel weight, estimate stop times in last-mile delivery, and forecast how many additional parcels will enter the network later in the day.
"We sometimes create a plan at ten in the morning while orders can still come in until five in the afternoon," says Thijs. "In that situation, you want to use historical patterns to estimate where a partner is likely to end up by the end of the day."
Much of the infrastructure built during the KickstartAI project is still being used today. Not just the technology itself, but also the way of working that came with it.
What problem are you actually trying to solve? What data do you need? How do you get people to trust a model? And how do you make sure it becomes part of the operation rather than remaining an interesting experiment?
"We were talking about improving QAVE again just last week," says Thijs. "That tells me we genuinely believe in it. And we have already built another model to measure weight based on our model (QAVE). So I feel there's still a lot more value to unlock."
