Picture a hospital network's new CIO inheriting a stalled AI program. The patient chatbot sits mostly unused. A model that predicts readmissions produces scores the doctors don't trust. The organization spent real money and has little to show for it.
That hospital is a composite, not a client. But the pattern is real, and anyone who has sat in an AI steering committee will recognize it.
In 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, rising costs and unclear business value (Gartner).
The common thread is a technology-first approach. Teams start with what the tool can do and go looking for a place to put it. A better starting point is the work people are actually trying to get done.
Jobs to Be Done
Clayton Christensen and his colleagues at Harvard Business School put it simply: customers don't buy products, they "hire" them to do a job (Harvard Business Review, 2016). As the old marketing line goes, nobody wants a quarter-inch drill. They want a quarter-inch hole, and behind that, a shelf on the wall.
The same shift works for AI. Nobody inside an organization wants "an AI initiative." They want something specific to get easier, faster or more reliable, and AI is one candidate they might hire to do it. We first wrote about this in Implementing AI with Jobs in mind; this post puts it to work.
Reframing the hospital's problem
Go back to the composite hospital. Suppose the CIO stops asking where AI could be used and asks three different questions:
- What critical jobs are our people struggling to get done?
- Where do current solutions fall short?
- Where could AI close that gap better than the alternatives?
Ask the head of emergency medicine, and AI probably isn't on their list at all. What they want is patients assessed faster and the sickest ones seen first. That's the job. A general-purpose chatbot doesn't do it. A tool that helps flag high-risk patients at intake might.
The readmissions model fails for a related reason. It was built to produce a prediction, but the doctors' job is deciding what to do for a specific patient before discharge. A score they can't interpret or act on doesn't help with that job, so they ignore it. Bring the doctors in to define the job, and the tool has to show its reasoning and fit the moment they make the call.
A real case: the job versus the claim
This isn't just a thought experiment. In 2021, an independent validation of Epic's widely used sepsis model at Michigan Medicine found it missed about two-thirds of sepsis cases while flagging nearly one in five hospitalized patients (JAMA Internal Medicine). The job clinicians needed done was catching sepsis early without burying them in alerts. The model was sold on a different measure of success, and nobody checked it against the real job until outside researchers did. We walk through that and other cases in Pick the Right AI Bets, Then Prove They Work.
Beyond healthcare
The same lens explains AI that works in other industries:
- Investing. Robo-advisors like Wealthfront are hired for a job most people would describe as "invest sensibly without having to watch my portfolio." Wealthfront now reports more than $100 billion in platform assets, including cash (Wealthfront, October 2026).
- Retail. Home Depot's customers aren't really hiring it to buy power tools. They're hiring it to finish a project with confidence. In 2025 it launched Magic Apron, generative AI tools in its app and website that answer project questions, help pick tools and estimate materials (Home Depot). It hasn't published sales or satisfaction results, so the real test is whether more projects get finished.
- Wind power. Siemens Gamesa aimed at a narrow job, keeping turbines from stopping unexpectedly. It uses models of turbine behavior to spot failing parts early, in one example weeks before a fault would have stopped the turbine (Siemens Gamesa).
Applying Jobs to Be Done to your AI initiatives
Five steps, and each one maps onto work I/O psychologists have done for decades:
- Interview for the job. Talk to the people doing the work about what frustrates them and what a good day looks like. Listen for the emotional side, like the worry about missing something, not just the functional tasks. This is job analysis by another name.
- Map the surrounding jobs. One job rarely stands alone. Faster triage connects to staffing, bed management and handoffs. Map them before you pick one.
- Prioritize by impact. For each candidate job, weigh four things: the value of doing it better, the risk if the AI gets it wrong, the friction it adds or removes for the people using it, and whether the action it should trigger is clear and visible. High value with low risk and low friction goes first.
- Design for the job. Let the job's requirements pick the technology, not the other way around. Sometimes the answer isn't AI at all.
- Measure the job, not the model. Define success as the job getting done better, then check it against real outcomes. Accuracy on a test set isn't the same as shorter waits or fewer missed cases. In I/O terms, that's validation against the criterion that matters.
The human side
Jobs to Be Done keeps people at the center of AI adoption, because it starts with what they're trying to accomplish. That's the same starting point I/O psychology has always used: understand the work and the people doing it before you change either.
The goal isn't replacing people with AI. It's giving them better tools for the jobs they already have, and proving the tools actually help.
Looking ahead
As AI gets more capable, the pull toward the latest breakthrough will only get stronger. Jobs to Be Done is a good antidote to shiny object syndrome. AI isn't magic. It's a tool, and the job tells you whether it's the right one.
If you're weighing where AI fits in your organization, start with the jobs. Our AI Hiring Validation work helps you prove that the tools you pick actually do them.