PostNord Boosts Parcel Address Reads to 80% with Lockheed Martin's Minerva AI

A shifting landscape

Like other mail and parcel operators around the world, PostNord is dealing with changing and challenging market conditions, navigating multiple forces at once. Mail volumes are falling 12–15% a year, pushing the company to consolidate terminals and find efficiencies. Parcel volumes, meanwhile, are climbing around 10–12% in 2026, with long-term growth expected at 7% annually.

"We have to keep increasing parcel capacity - both through new terminals and by upgrading what we already have," says Magnus Lundin-Maraboli, Head of Machines and Maintenance at PostNord AB. “Part of that efficiency improvement is being able to ensure the safe delivery of post and parcels, particularly around last mile – and address recognition is key to this.”

Enhancing automated reading accuracy for imported parcel labels

The original trigger to investigate new ways to improve automated label recognition occurred almost a decade ago. Christian Edfjäll, Senior Technical Project Manager, tells the story. “During 2016–2017 we saw a rapid increase in import volumes from China, in particular with small parcels presenting address labels that traditional OCR systems struggled to read correctly. This coincided with a new Swedish value-based tax on parcels, placing significant pressure on our existing manual and semi-automated processes,” he says.

The company realised it needed to respond swiftly and immediately started investigating ways to increase reading rates for import volumes to enable a more automated and scalable handling process. “We quickly realised that conventional OCR technology could not solve this challenge,” says Edfjäll.

Traditional OCR limitations

Traditional Optical Character Recognition (OCR) systems have long been used in postal sortation to interpret (read) handwritten and printed addresses and codes on letters and parcels. However, conventional OCR systems often still struggle with recognition and/or accuracy. Handwritten text, for example, especially when contrast is poor (for instance black ink on a red envelope) may be unrecognisable when scanned (binarized), requiring the image to be inspected by a human operative to manually key in the correct details.

Printed address labels, as in the case of packages from China, also pose problems. They may be crumpled, be very small, have poor print quality, include unusual foreign or stylised fonts, or appear shiny or smudged. Similarly, items like magazines that are wrapped in clear plastic are also notoriously difficult for OCR systems to interpret: the system cannot correctly identify the relevant address label area against the visual ‘noise’ of the package.

Explains Edfjäll, “The automated address recognition use case we had identified pointed to using AI-enabled technology, but at that time no vendor was developing solutions for our sector. Following discussions with Lockheed Martin, we jointly agreed to embark on a collaborative project to pioneer the step-change new AI-enabled technology required.”

A pioneering approach with AI

Within seven months of agreeing the scope and business case, The Lockheed Martin team had developed, integrated and implemented a working production environment prototype for Minerva, creating one of the very first AI-enabled OCR enhancement sortation solutions. The latest Minerva solution is now delivered to the market using a Software-as-a-Service model.

Over the years, close collaboration and a real partnership approach have been fundamental to enabling this technology to continually improve and evolve. 

Dave Woodward, Systems Solutions Programmes Manager, Lockheed Martin, observes: 

Significant cost savings

Over the past decade, Minerva has delivered significant cost savings for PostNord. These savings have been achieved by reducing the need for manual video coding and keying, minimizing mis-sorted mail and parcels that loop within the system, and automating data capture and sharing for parcel customs information.

Before Minerva was implemented, PostNord’s automated OCR address read rate for parcels was 54%, with an error rate of 7%. Today, the read rate has increased to around 80%, while the error rate has been reduced to just 2%.

A decade of collaboration

Says Lundin-Maraboli, “We started with, and successfully solved, one specific problem. But Minerva’s developers have always been open and receptive to new ideas and ways of working. Over time, working together, we’ve been able to expand and enhance Minerva capabilities to resolve other operational challenges. With a strong understanding of how we operate, the Lockheed Martin team proactively suggests new capabilities and improvements. No other generalist AI team could offer us this level of specific sector knowledge, insight and performance capability.”

Practical advice

What lessons has PostNord learned from live usage of Minerva? “Overall, our experience has been very positive,” says Edfjäll. “As with all AI based systems, there are some limitations - and there will always be edge cases that can be distracting - but ultimately, we have consistently focused on the optimisation of system performance as a whole, which has been delivering huge benefits, for many year.”