5 tips for building an AI tool to address worker shortages

Caterpillar's chief digital officer outlined five strategies the company used to build the Cat AI Assistant, an AI agent deployed in under a year to help address a 37,000-per-year shortage of skilled technicians in the U.S. automotive and heavy equipment industry. The company consolidated data across nearly 50,000 service documents and about 1.5 million parts onto a platform called Cat Helios before launching the assistant, which the executive said should be scoped like a job description rather than a search engine, with defined rules for what it decides versus escalates.

The U.S. automotive and heavy equipment industry has a 37,000 shortage of skilled technicians https://www.nada.org/nada/issues/service-technicians each year, even after accounting for technical school graduates. In construction, experienced hands are retiring, leaving increasingly complex machines in the hands of increasingly inexperienced workers. Manufacturing and field services tell the same story about an aging workforce, constrained talent pipelines, amid strong and growing demand. In these industries, missteps could mean a downed power line, machine failure on a critical concrete pour, or a collapsed trench during excavation. As tenured talent retires, there’s a loss of expertise that can’t be replaced by hiring https://www.fastcompany.com/section/hiring alone. AI https://www.fastcompany.com/section/artificial-intelligence agents can help close that gap. Like the relationship between an apprentice and a mentor, AI agents can convey a wealth of knowledge to less experienced team members. They can also handle routine tasks, freeing technicians to focus on what only people can do: judgment, validation, and oversight. Building an agent that works in these industries requires a fundamentally different approach. These are five strategies we used at Caterpillar when creating and deploying AI technology to help bridge the worker shortage. 1. Start by defining the problem. Many AI deployments move too quickly. Businesses have an issue and jump to an AI fix like an LLM. This often results in failures when it’s time to produce and deploy at scale. Start with friction, like the moment a decision stalls because the expert isn’t available. Take the time to truly understand the cause and the impact. Then, select the right tool. Your answer could be an AI agent, or it could be something else, like heuristic algorithms, machine learning solutions, or simply better training. For example, at Caterpillar, our diverse customer base spans power generation, construction, and mining industries, requiring a large ecosystem. We have nearly 50,000 service documents and about 1.5 million parts. When we built the Cat<sup ®</sup AI Assistant<sup TM</sup , we started with “How can we make this easier for our customers?” In many companies, data is fragmented across new and legacy systems. It’s stored in different structures with different formats and must follow a variety of governance requirements. This isn’t new to any data or analytics team, but the importance of unifying the data—no matter how tough—can’t be understated. When I stepped into the role of chief digital officer at Caterpillar, we started a massive project to clean all our data and move it to a data platform called Cat Helios <https://caterpillar.sharepoint.com/:w:/r/teams/DESDCommunicationsGroup/Shared%20Documents/2026/2026%20Media%20and%20External/Fast%20Company%20Impact%20Council/Fast%20Co%20Op%20Ed AI%20Teammates%20 1 .docx?d=w6466ea7dfc3d45b8bc0bd808045c13ff&csf=1&web=1&e=hWkfQr . It was a behemoth undertaking. But this foundation, which now powers all our digital applications, enabled Cat AI Assistant to go from concept to launch in under a year. The data foundation is the foundation of trust. How organized your data is determines how far the solution extends. An agent with partial information, like service history but not parts availability, won’t be effective. Similar to how people build expertise, AI agents tend to have strong expertise in some domains and less in others. A single agent covering the full breadth of a complex environment will have uneven performance—and in a high-stakes context that leads to lost trust, fast. Stop designing AI agents like search engines and start designing them like a job description. What’s its role? Who is accountable for its work? Create a defined scope, explicit rules of engagement, a clear hierarchy of what the agent decides versus what it escalates, and accountability for output quality. Multiple specialized agents can each handle a narrow function reliably; the right agent will handle the right problem. This architecture is what makes both depth and breadth possible. Cat AI Assistantis a collection of agents, each with its own purpose and job description. It exists to help users be more effective at their existing tasks. Just like that apprentice relationship, it provides access to knowledge that would have taken years to acquire . Its role is to be the trusted coach, not the technician. In all AI, accuracy improves over time. In high-stakes industries, however, there is a chasm between good enough for a demo and good enough for the field. Think about the standard for an excavator operator—high accountability, no margin for error. That’s the bar AI agents operating in analogous roles must meet. Tomorrow’s operators will run machines and orchestrate entire job sites, intervening when human judgment matters most. Achieving that requires architecting the system to retrieve information, validate outputs, and communicate uncertainty. Organizations that deploy AI solutions at 70% end up with trust issues on the frontline. Once trust is gone, it doesn’t come back easily. Nail your why at rollout. Our goals are augmentation and empowerment. We’re striving to make everyone more effective at what they already do—faster learning, more efficient, and more accurate work. This message must be clear. Adoption comes down to understanding the benefits. What’s the value to their day-to-day lives? Maybe it’s shortening the learning curve, so they look good to their manager, wrap up faster and get home, or stand out from competition. The demographic math isn’t reversing and retirements aren’t slowing. Organizations that extend the expertise of their existing workforce through well-designed AI agents will build a sustainable talent pipeline compared to those who will spend a lot of money by treating this as a plug-and-play technology problem. Anyone can connect an LLM. Building a digital assistant that meets the real-world standards required by a field technician or a clinical operator that meets real accuracy standards, sits on a real data foundation, and fits into the actual workflow is hard. It’s also exactly the work that defines who will lead these industries on the other side of this labor transition. Ogi Redzic is the chief digital officer at Caterpillar.