How Enterprise AI Teams Are Finding Their Strongest Use Cases In Previously Abandoned Projects

Most companies looking for AI projects start by asking what they could build next. A better place to look is the work they already tried and abandoned. Projects that got scoped, costed, and then killed for lack of people or budget are worth re-testing now, because the thing that blocked them may no longer be a problem. A shelved project also has a paper trail. Someone documented what it would take, and the first attempt showed how far that got.

Luke Hobson, EdD, is the Assistant Director of Instructional Design at MIT xPRO, the professional development division of the Massachusetts Institute of Technology. Instructional designers build the courses and training programs that companies and universities run, and they decide how the material is structured so people learn from it. Hobson has worked in the field for more than a decade, teaches in the University of Miami's doctoral program in education, and surveys other designers about how the profession is changing.

"My colleagues are going back and looking at past problems, seeing if AI can help with something we couldn't do alone at the time," Hobson says. MIT xPRO wanted a database showing what sits inside each of its roughly 60 course and program offerings, including the videos, readings, and case studies, so anyone on the team could see it without opening a course and clicking through the navigation. An earlier attempt ran for more than a year and produced something incomplete. The team came back to it with AI and had a working version in four hours.

Designers elsewhere are applying the same test to other stalled work. "Maybe that was a people power issue, where we didn't have enough folks working on the team at the time. Maybe it was a budget issue or some other constraint. Can AI actually help to solve that problem?" Hobson says.

Judgment tasks come first

Hobson's survey of 587 instructional designers, published in July, asked what they use AI for. The results show 73% using it daily or a few times a week. The top tasks are building assessments and activities at 64%, research and analysis at 62%, and writing learning outcomes at 62%. Those are the decisions that determine whether a course teaches anyone anything. "We're still following best practices and standards. It's not so much what you just read about on LinkedIn and X," Hobson says. A separate question found 73% saying the core skills of the job cannot be replaced, so frequent use has not meant handing the work over.

How much pressure designers feel to use AI depends on where they work. The survey put corporate designers at 52%, higher education at 33%, and government at 29%. In higher education, approvals go through committees and policy reviews, so a tool gets questioned several times before it reaches a course. Corporate teams usually start further along, with a leader who has already bought a product and a team that has to show a return on it.

Who checks the work

Instructional design has been around for decades and has its own standards for what a well-built course looks like, so an experienced designer can open one and tell within seconds whether it was made carefully or rushed. That training is what makes fast work usable, because someone has to catch a bad course before a student sits through it. "Any person within my field who has been doing this for a while can go inside of a course and immediately tell you if it's good or if it was made in a rush," Hobson says. Education has also heard a lot of technology promises before, including the virtual reality headsets that were going to change the classroom and never did, so the field tends to ask what a tool actually changes before adopting it.

The same standard applies to anyone presenting AI-assisted work. Someone who cannot explain how they reached an answer loses the idea the first time a colleague questions it, which is why the review step matters more than the speed that produced the draft. For data and infrastructure teams, that covers anything a model proposes about architecture, cost, or risk.

Designers also learn what works from each other more than from formal sources. The same research put social media first at 65%, colleagues and professional networks second at 57%, and academic journals last at 20%. A screen share and a working example carry more weight than a claim about the next big thing, so practitioners who find something useful tend to post the steps.

That shows up in what MIT xPRO gets asked to build. Requests have moved off generative AI as a broad subject and into narrower pairings, including leadership and AI, project management and AI, agentic systems, context engineering, and retrieval. Companies asking those questions want training embedded into workflows they already run.

Governance is next

Hobson tracks what people search for when they come looking for courses, and the requests are moving toward policy. Governance, security rules, and company-wide AI strategy are climbing the list, above the tool questions that came first. "Thinking about security policy, guidelines, governance, all of that is going to be coming down much sooner, I think," Hobson says. Those topics do not excite anyone the way a hands-on build does. They show up anyway, because companies hit a point where not having a policy is what stops the work.

The teams asking are usually already running AI against real data, where who signs off on an action has stopped being a theoretical question. Experimenting is how anyone learns what a tool can do, and the trouble starts when everyone experiments at once with information the company cannot afford to expose.

A recent HBR guide on generative AI for managers, by Elisa Farri and Gabriele Rosani, put a name to the approach Hobson had already been using. "You should be experimenting, but not your whole organization. Certain people should be figuring it out, especially with sensitive data and privacy," Hobson says.