The AI Post Office: Rediscovering USPS’s Forgotten Digital Future

 In the second instalment of their feature on AI and the future of the 21st Century Post Office, Michael Kubayanda and Robert A. F. Reisner look back at the pioneering role the US Postal Service played in artificial intelligence research during the late 1980s and 1990s — and consider what that forgotten history might tell us about its future.

Applying Innovation to Today’s Challenges

The contemporary USPS, like its previous generation, faces a daunting, large-scale efficiency challenge. Today it runs a growing national network as the mail and package volume which pays for the network declines. The USPS has to provide affordable universal service. And it needs to design transportation and processing networks to do all this, while reducing costs.

The USPS announced the Delivering for America plan (DFA) in 2020, a massive investment to modernize and centralize the postal network by upgrading outdated infrastructure; reducing operating costs, in some cases at the cost of slower mail delivery; and increasing prices, while transforming the USPS into primarily a modern e-commerce company.

Few of the hoped-for benefits of the DFA materialized in the first few years. Some of the early struggles might be chalked up to delays in obtaining a return on investments in infrastructure. That wouldn’t explain, however, why some results were the opposite of what the USPS aimed for – productivity declined, new showcase facilities malfunctioned, and its lightly regulated e-commerce package business might have actually become less competitive in some ways.

Postmaster General Steiner pivoted quickly after taking over last year, reemphasizing industry partnerships to cut costs and speed up delivery while keeping the DFA title and its focus on large, centralized processing plants.

  It’s Hard to Model the Huge Variety of Issues Which Affect Postal Performance

The USPS hasn’t yet resolved concerns about how it redesigned the processing and transportation networks in recent years. The USPS used powerful analytics software, but leaned on deterministic models, meaning it did not account for variation in the network.

In real life, the postal system varies a lot. The US has major geographic differences in package and mail volume, population density, terrain, and weather. There is more mail during the holidays than in June, and mail volume affects processing plants in cities differently from facilities in rural areas. Both pre and post-DFA, the plants have different layouts, operating plans, and schedules. These differences affect costs, efficiency, and quality of service, and under its universal service mandate the USPS is not allowed to ignore challenging or unprofitable areas to construct a uniform network.

The Postal Regulatory Commission, among others, noted that USPS should use stochastic modeling techniques, which produce a range of possible outcomes weighted by their probabilities. The USPS appeared worried that it did not have the capacity to find every relevant variable, weigh them appropriately, and analyze a large number of combinations. This could be a long process, if handled manually, and automation is not an easy solution. Companies often need highly trained specialists to deal with the logistics software as well.

But in recent years, AI is being used to address such issues. Logistics and supply chain planners in other organizations have used LLM chatbots as a plain English interface with analytics software, prompting LLMs with questions that lead to more technical queries for the underlying mathematical models. Similarly, a USPS manager could ask an LLM what variables would affect the time packages will arrive at a processing plant. January weather in Alaska and traffic in metro Atlanta might be two examples of complexities that could be addressed in this way.

Technology leaders at the USPS, like others throughout the logistics world, have begun to voice similar aspirations at industry forums. Will they have time, and can the USPS update its practices? Critics will argue that the traditional USPS would lack the capacity to create “modern” solutions.  But the case of OCR might offer more promise than people might think is possible for the USPS.

With its databases, logistics software, and human experts who run the network every day, the USPS has the data and models showing how variables have affected service and costs, for example, in all areas of the country. Accurate data on these variables could support a stochastic analysis with guidance from human experts. AI-supported analysis should be faster and would help incorporate large combinations of variables into a plan. The USPS might also use this approach to adjust to fast-emerging bottlenecks, such as dynamic rerouting to deal with the traffic jams that formed when it opened new facilities in implementing the DFA.

Digital Twins: A Data Platform for Learning Before Deploying Real-World Changes

Creating a modern-day sandbox to facilitate investments in reinvention might not require rebuilding the postal network to test innovations, given the tools currently available. The USPS could use “digital twins” at the heart of an effort to use LLMs and other technologies to improve postal operations. A digital twin is a replica of a real-world system, updated with the data from the system. Digital twins are a key part of AI strategies in improving physical systems, including logistics and distribution networks.

Digital twins allow for a wide spectrum of experimentation on a twin without interrupting the operations of the real postal network. Digital twins lower the hurdles, and potentially the risks, for experimentation by government agencies. The twins can also help with documenting the case for updating a strategy based on new evidence, if used with this purpose in mind.

Analysts armed with machine learning tools and accurate data can use a digital twin to run more analyses on factors that could affect service and costs. Postal executives could take this more detailed approach without excessive delays in making adjustments to DFA, and without making real-world changes that are not fully understood. They would still need to explain how the new tools helped them reach their conclusions.

Technology Governance and Risks

The USPS is a government service, not an investor-backed startup. As its leaders design and implement changes to the network, they need to anticipate the pitfalls of AI and other technologies, even beyond the headline-grabbing doomsday scenarios that are becoming increasingly common.

In contrast to their ability to deal with many complex scenarios, AI applications often have problems explaining how they reach their results. From procuring new AI tools to putting them to work, executives will have to take care to build models in stages, ensure explainability, and direct the AI tools to show their work and cite sources. Policymakers, stakeholders, and oversight agencies expect this type of transparency at a minimum from public sector operators. Where an AI tool cannot explain or reproduce its analysis, managers need to rely on human experts using software with transparent reasoning.

The USPS faces legal, regulatory, and stakeholder requirements that are common for the government, utilities, and highly regulated industries. These requirements are often qualitative rather than quantitative. Meeting those commitments (often shifting and conflicting) will continue to require conscientious humans prompting the AI systems and interpreting the results, rather than AI-only optimization for its own sake.

The USPS will also need to address AI-related issues of governance, ethics, safety, and privacy. The USPS was one of the first major organizations to have a Chief Privacy Officer in 2000. Privacy enhancing technologies, such as anonymizing data in a digital twin, should be viewed as basic and foundational steps rather than “nice to haves”.

But the experience of the USPS in contributing to the creation of AI should provide a source of inspiration that the challenges of adaptation to an uncertain future are not insurmountable.

Subject to evolving standards and guardrails, the USPS will have more opportunities to use AI than outlined here in these modest suggestions. There are already 35 approved “use cases” at the USPS, according to a recent USPS Office of Inspector General study. But to scale from pilot to enterprise applications, there is a need to understand the elements that were present in previous USPS success. And even then, the world of AI is moving much faster than it was 30 years ago and the USPS needs to be ready to build on its knowledge and success, working with others while fulfilling its universal service obligation. By building on past success, the USPS can earn the time to reinvent itself and reinforce its value. The USPS will need to earn critical time from its stakeholders and policymakers and demonstrate the capacity for successful innovation.

Despite the obstacles to using new technology to help reinvent the postal system, the USPS’ historic role in AI suggests that it has the capacity for reinvention.

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.

To read the first part of this feature see .