Tuesday, 22 September 2026
Universities around the world are rapidly adding artificial intelligence courses, degrees and executive programmes, but one of the bigger educational opportunities may lie outside both advanced computer science and high-level AI strategy. At Ai4 2026 in Las Vegas, University of Southern California lecturer Mike Lee argued that the next stage of AI education should concentrate far more heavily on domain specialists who understand how real organisations operate and can redesign those workflows around AI agents. Lee described the challenge through what he called the “90% problem”. The percentages were not presented as a formal study of university education, but as a way of illustrating what he sees as a gap between highly specialised AI researchers, managers learning strategy and the much larger population of professionals who understand real-world problems but are not trained as AI engineers.
That distinction is becoming more important as generative AI moves beyond prompting and document creation towards agent-based workflows. Knowing how to ask a chatbot for an answer is very different from designing a system in which several AI components carry out separate roles, use different information sources, remember relevant context, communicate with external software and escalate decisions to people when necessary. Lee’s argument was that universities need to teach this second level of capability to non-engineers rather than leaving agent development largely to computer scientists. The objective is not to turn every student into a machine-learning specialist, but to give people enough technical understanding to redesign processes within fields they already know.
The approach Lee outlined at Ai4 starts with problems rather than models. Instead of beginning with a programming language or asking students to demonstrate a generic AI technique, students are encouraged to identify a problem they understand personally and then determine how a combination of AI agents, software integrations and human oversight might solve it. One classroom example involved a student who wanted music to respond automatically to mood. A user could send an image through an existing messaging interface, an AI system would interpret the context, recommend music and connect with a streaming platform to generate an appropriate playlist. The importance of the example was not the consumer application itself, but the way multiple components could be linked into a single workflow without requiring the user to navigate a separate application.
Another student project addressed the problem of people losing access to public healthcare benefits because they fail to complete administrative renewal requirements. The proposed system used policy information as a controlled knowledge source, evaluated cases according to different levels of risk and then selected different forms of communication. Lower-risk cases could receive automated messages, while more complicated situations could be escalated to human intervention. A further prototype examined healthcare appointment scheduling, where insurance information and a patient’s request could be processed by multiple specialised agents before the system identified an appropriate appointment and interacted with a calendar. These were classroom demonstrations rather than clinically validated healthcare products, but they illustrated an important principle: AI workflows can be designed around multiple decision points, with automation used selectively rather than as a replacement for human involvement.
This decomposition of work may become one of the most important AI skills in business. Traditional enterprise technology has long been organised around workflows, permissions, databases and integrations. Agentic AI introduces software components capable of interpreting information and making limited decisions inside those workflows. Domain specialists may therefore become particularly important because they understand which steps require judgement, which can be automated, what information is relevant and when a system should stop and ask a human. As foundation models become more accessible, competitive advantage may increasingly depend less on owning the underlying model and more on combining it with proprietary processes, data, institutional knowledge and customer understanding.
That could change the value of non-technical expertise. AI development has often been discussed as a competition for computer scientists, data scientists and machine-learning engineers, but organisations also need people who understand how industries actually work. Knowing how a property transaction is structured, how an insurance claim is processed, how a hospital manages patient administration or how a supply chain responds to disruption can become as important as understanding model architecture. The opportunity lies in turning that domain knowledge into workflows that AI can support without losing the judgement and accountability required in real operations.
There are limits to the idea that every domain specialist can become an AI developer. Production systems handling medical information, financial decisions, personal data or critical infrastructure still require professional engineering, cybersecurity, legal review and governance. Low-code and AI-assisted development can make prototyping dramatically easier, but ease of creation does not remove the need for testing, access controls, monitoring and accountability when a system is deployed in the real world. The distinction is therefore between democratising the ability to design AI workflows and eliminating technical expertise. Universities can give students enough understanding to identify opportunities, construct prototypes and communicate requirements while specialists remain responsible for the deeper infrastructure needed to operate consequential systems safely.
This could resemble the development of the internet, when widespread adoption did not require everyone to become a network engineer but did require almost every profession to understand how digital tools could change its work. The same principle may now apply to AI. Teaching only AI strategy may be insufficient because strategic presentations alone do not redesign processes. Transformation happens when somebody who understands an existing workflow can determine which parts should be automated, what data should be available, where AI should make recommendations and where people should retain authority.
For employers, that creates a different type of workforce requirement. Companies may increasingly look for professionals who combine domain knowledge with enough technical understanding to build and supervise AI-enabled workflows. A finance specialist might design an agent that assembles information for underwriting while understanding when an analyst must intervene. A marketing professional could coordinate several systems across research, content and campaign management. A logistics manager might use agents to handle exceptions across warehouse and transport systems. The technological foundation could come from external platforms, but the quality of the resulting system would depend heavily on the people who understand the business process.
Commercial real estate provides a particularly clear example. Property companies already possess large numbers of workflows involving leases, due diligence, valuation, tenant communication, building operations, development documentation, invoices and investment analysis. Much of this information remains fragmented between documents, email and specialist software. The professionals best positioned to determine where AI could improve those processes may not be AI researchers, but asset managers, valuers, leasing professionals, engineers and investment analysts who understand how the work is actually performed. A leasing professional could help design an agent that reviews lease obligations and identifies upcoming actions, while a property manager could define a workflow connecting maintenance requests with building data and contractors. An investment team could create specialised agents for market research, financial analysis and due diligence while retaining human approval over assumptions and transactions.
This also raises questions about how quickly universities can adapt. Traditional curricula often change more slowly than commercial technology, while agentic AI is evolving rapidly enough that some of the most valuable programmes may need to operate more like laboratories. Students can experiment with current tools while learning principles that are likely to remain important even as individual platforms change: workflow design, data quality, security, verification, permissions and human accountability. USC’s wider emphasis on applied computing and student entrepreneurship fits that approach, with Lee’s teaching and startup work focused heavily on practical implementation rather than purely theoretical AI.
The larger educational question is whether AI becomes another specialised technology subject or something closer to a general professional capability. If AI becomes embedded across most industries, universities may eventually need to teach its practical use in the same way they incorporated spreadsheets, databases, digital communications and internet research into disciplines far beyond computer science. Researchers capable of advancing the underlying technology will still be essential, while executives will continue to need strategic understanding, but much of the practical economic value could come from a much larger population that knows how to redesign work inside individual industries.
That would shift AI education away from simply asking students how the technology works or which tools they can operate. The more important question becomes whether they can identify a real problem, divide it into appropriate tasks, determine which decisions can be delegated to machines and build a workflow in which AI and people operate together. If that becomes a standard professional skill, the most consequential AI workforce may not consist only of people with AI in their job titles. It may be the much larger group of accountants, engineers, healthcare workers, marketers, property professionals, entrepreneurs and managers who learn how to turn their existing expertise into AI-enabled systems.
Source: CIJ.World Research & Analysis Team
