Can AI help revive the 21st Century Post Office before it’s too late?

 The key role that the US Postal Service played in advancing Artificial Intelligence in the late 1980s and 1990s may seem surprising today as the historic institution struggles to maintain relevance in a changing marketplace.  Michael Kubayanda and Robert A. F. Reisner look back at its pioneering role.       

AI’s Surprising Catalyst: Part I

“The US Postal Service has been in the news recently for many of the wrong reasons. In March, Postmaster General David Steiner warned a Congressional panel that after losing $9 billion in 2025, the agency could run out of cash next year and would have to stop delivering mail to every household and business every day. The crisis eased when the carrier and its regulator acted to stave off the cash crisis for the time being. But Postmaster General Steiner has asked Congress for additional help. If the USPS is going to gain support from policymakers and stakeholders for longer term solutions, the venerable carrier will have to show that it can deliver value in the future, not just rely on subsidies to an outdated business model.

The critics will argue that the Service is not capable of adapting to the Age of Artificial Intelligence. But in fact, the USPS already has a record of using innovation to overcome major challenges, and in doing so, it helped to spur the artificial intelligence revolution itself. The USPS helped to accelerate AI three decades ago, investing billions in optical character readers (OCRs) to read addresses on letters and packages. Today, OCR technology is ubiquitous, and the USPS itself is a user of predictive and generative AI applications. But in the late 1980s and early 1990s, the USPS’ transformative investment was highly innovative. Before the nation gives up on the USPS, it’s important to see that it has faced a turnaround imperative before and prevailed.

There is growing interest in government and the postal sector in identifying new AI applications, including in agentic AI. There is value therefore in zeroing in on the earlier USPS innovation success and seeing that its key elements still exist today.

How and Why the USPS Adopted AI More Than 30 Years Ago

 This key chapter in the history of AI and the USPS started when the Postal Service faced a basic problem: reading handwriting to make sure that mail arrives at its intended destination. Even as the USPS invested in automation, handwritten addresses were an overwhelming problem which threatened to undermine those investments and limit efficiency, especially during the holidays when many customers addressed cards by hand.

In 1988, the USPS published a new automation plan to guide investment in improving efficiency throughout the mail processing network. The plan envisioned remote encoding centers, where employees would look at pictures of addresses and apply barcodes to help sort and route letters to their destinations. USPS tasked leading technology companies to develop OCR systems to read addresses automatically. But even with OCRs, handwriting was still a problem.

In 1989 and 1990, AI pioneer Yann LeCun and researchers at Bell Labs used handwritten samples provided by the USPS to test using a neural network that would provide the computer intelligence to facilitate interpreting addresses. LeCun published seminal work describing the neural network’s success. Bell Labs continued fine-tuning this approach for USPS optical character reading in the 1990s. These systems were among the earliest large-scale uses of machine learning and helped to push machine learning to the forefront as the dominant AI approach.

In machine learning, algorithms analyse training data and then determine rules or patterns based on that data, using a bottom-up approach. Partly due to the scale of the work at the USPS, machine learning began to overtake other AI approaches (symbolic systems) that guide algorithms using logical rules in a top-down fashion. The neural networks developed by LeCun and colleagues underlie today’s Large Language Models (LLMs) and chatbots. The USPS contributed a large amount of data to this learning process before ubiquitous Internet, social media, and smartphone data even existed.

But neural networks require substantial processing power. Computing power has grown and costs have gone down in the last 30 years. But in the 1990s, costs remained an issue for scaling up an OCR solution based on neural networks.

Starting in 1994, a team at the University of Buffalo led by AI researcher Dr. Venu Govindaraju came up with a new method. The U. of Buffalo method took images of handwritten addresses and matched the data to a lexicon (or database of words) containing the elements of addresses – cities, street names, numbers, and ZIP codes. This method combined OCR with postal knowledge to interpret complete addresses. The system learned from the underlying data, relying on pattern matching algorithms to interpret addresses from millions of possibilities. This approach limited potential matches by searching only for selected street names and number combinations within the given ZIP codes.

This method required relatively little computing power compared to using neural networks, while achieving fast recognition and ever-increasing accuracy. Commercial automation providers also refined the system on behalf of the USPS.

In five years, this system improved from a 40 percent success rate in interpreting handwritten addresses to better than 95 percent. The USPS handled the small number of remaining addresses with people, by sending an image to a remote encoding site where employees read them. As the researchers improved the technology, the need for remote encoding decreased. By 2014, the USPS needed only one facility for interpreting addresses manually, down from 55 such facilities in 1997.

The work served as an early large-scale testbed for ideas that remain central to AI today: learning from data, managing uncertainty, estimating confidence, and using contextual knowledge to resolve ambiguity. Character recognition and pattern matching, which helped to facilitate USPS automation three decades ago, are now fundamental parts of AI-driven services such as creating automated meeting summaries, filling out prescriptions and providing doctor’s notes, and fraud detection in financial services.

A Public Sector Sandbox for Innovation

The USPS’ successful OCR investment may even offer insight into potential paths for future innovation. The USPS started with a narrow and tangible problem and created, in effect, a “sandbox” for teams from the private sector, universities, and government to experiment with new applications to solve the address interpretation challenge.

The USPS, and its partners, analysed large amounts of training data and facilitated decentralised learning through its partnerships. The sandbox participants learned over time rather than sticking with one initial approach. To facilitate experimentation, the USPS struck agreements with major labor unions that acknowledged an initial need for employee coders and anticipated reducing the number of hand coding sites as the machines learned. It also retrained and redeployed workers over time.

 The USPS’ 1990s Sandbox Model

 

Large amounts of training data	Decentralised, iterative, and inductive learning
Partnerships with universities and the private sector	Agreements with major labor unions

 

Today’s USPS continues to have access to the elements of its prior innovation sandbox. The USPS has access to contractor technology expertise to supplement its in-house capabilities, opens its platform to industry customers and third-party service providers (a critical part of e-commerce and the traditional mailing industry), and uses price incentives and contract agreements to provide access to its hybrid physical-digital platform.

About the authors: Michael Kubayanda is the former Chairman of the Postal Regulatory Commission and Robert Reisner is the former Vice President for Strategic Planning of the USPS and the author of “When a Turnaround Stalls” Harvard Business Review, 2002. The authors would like to thank Dr. Venugopal Govindaraju of the University of Buffalo, and Steve DeNunzio and Dr. Vince Castillo of The Ohio State University for their input and feedback.