Over the past three years, Ford has hired about 350 veteran engineers, some of them former employees, to find the failure points its automated quality systems and AI inspection tools were missing. It needed them because, as its vice president of vehicle hardware engineering has since admitted, the company had assumed that introducing artificial intelligence and feeding it design requirements would be enough to produce a high-quality product. It was not.
Ford is not an outlier. Klarna announced in February 2024 that its AI assistant was doing the work of 700 customer service agents. By May 2025, it was recruiting human agents again, its CEO conceding that the company had focused too much on cost while letting quality slip. The data tells the same story. A survey of 1,000 C-suite and senior leaders at medium and large organizations found that 39% had conducted layoffs as a result of AI—and 55% of those admitted to making wrong decisions about the redundancies. And the reversals are expected to multiply: Gartner predicts that by 2027, half of the companies that attributed job cuts to AI will be rehiring people to do similar work under different titles.
When firing is closely followed by rehiring, something has clearly gone wrong. The tempting explanation is that the technology was not ready. But look at what Ford’s veterans are doing now that they are back: training the next generation of engineers and retraining the AI systems that were meant to replace them. Passing on what they knew was one of the things their jobs were for, and it never appeared in any task list, so nobody saw it go. Jobs have been eliminated without a proper understanding of what those jobs entailed and what purpose they served. This is, as I argued recently, a systems and people problem, not a technology problem. Redesigning work is one of the hardest parts of the AI transformation. Here’s a process for doing it successfully.
What a job is for
A job exists because the organization needs several things from it. Those are its goals, and they are the starting point for redesigning work in the age of AI: Understand what a job is for, then work out how people and machines can be combined to fulfill it.
Goals and tasks are different kinds of things. A goal is an outcome; a task is an activity. Some goals are met simply by completing the right tasks. Others are not. One goal of a customer service role, for example, might be to give the customer accurate information about the product, sales conditions, and refunds. The work of meeting that goal breaks down neatly into tasks. Another goal might be to give a customer having trouble the sense that someone has taken responsibility for the problem. The work of meeting that goal is much harder to decompose, because a customer does not feel that someone has taken responsibility just because every item on a checklist gets ticked.
A redesign that starts from the tasks will see the first kind of goal and miss the second. One that starts from the goals sees both. The five steps below are built to do the latter, and they can be adapted to most contexts.
1. Describe the goals. Begin with what the organization needs the role for, not what the person in it does: the output required, the relationships that need to be built and maintained, the information that needs to move through the system. A role might also exist partly to develop the people the organization will need down the road—the goal Ford lost sight of. Write them all down, including the ones that never appear in the job description.
2. Ask which goals need a person. For each goal, ask one question: Would it still be met if every task involved were performed by a machine? Resolving a balance query is a goal the tasks alone meet; a customer feeling that someone has taken responsibility for a stolen card is not. This is a question about the outcome, not about how capable the technology is.
3. Decompose the rest. For the goals the tasks alone meet, break the work into its tasks. Tag each as “automate,” “augment,” or “not yet automatable,” based on what the technology can do in your context rather than what a vendor says it can do. “Not yet automatable” means the technology cannot do the task well enough today. It is different from a goal that needs a person, and the two should not be filed together, because one will change with the technology and the other will not.
4. Choose the lens. The first three steps involve analysis. The fourth is a decision: What lens will you use for the redesign? A cost lens automates everything in step 3 that can be automated and staffs the step 2 goals as thinly as they can be staffed. A quality lens keeps a person wherever the output is better for it. A human-centric lens asks a further question: What does the redesign do to the people on each side of the work? That means the customer, who is a person before they are a ticket, and the employee, whose own goals for the job, such as learning to perform their role better, count alongside the organization’s. Under this lens, the organization may keep a person in a position even when a machine might be able to do it as well or better.
5. Recompose what remains. After step 3, you are left with goals requiring human input and tasks the technology cannot yet handle, all spread across the old roles. The aim is to rebuild this remainder into whole roles. Pool the remainder across the function rather than within each job, then group by what the pieces have in common: the same judgment, the same relationships, or the same point in the workflow. Each group is a candidate role. The same remainder produces different roles under each lens. Under a cost lens, recomposition means finding the fewest people who can hold the remaining goals. Under a human-centric lens, it also asks about the people who will hold the roles: whether the recomposed job is one someone would want, whether it develops the individual, and how wide the entry points are through which the next generation will acquire the judgment the roles now depend on.
Redesigning customer service
Imagine running the process for a customer service team that today is a single tier of agents handling everything from balance queries to stolen cards. Step 1 identifies four goals: resolving routine queries; giving a customer in trouble someone who takes responsibility; feeding what customers say back to product and risk; and producing the experienced agents who will handle the hardest cases in three years.
Step 2 finds that only the first and third are met by the tasks alone; the second and fourth need a person regardless of what the technology can do. Step 3 breaks the work of the first and third into tasks. Most of the routine-query tasks tag as “automate.” The feedback to product and risk tags as “augment,” because the patterns still need a person to read them. In step 4, you choose a human-centric lens. Its consequences show up in the final step.
Step 5 pools what is left: the calls where a person needs to take responsibility, the development of future agents, the disputes and edge cases the technology cannot yet handle, and the review of the automated conversations. The lens then adds two things a cost lens would not: a route to a person that is visible to the customer, because the customer is a person before they are a ticket; and a share of routine queries reserved for new agents, even though the AI could handle them more cheaply, because that is how agents develop. Grouped by the judgment they need, these fall into two roles: A senior agent owns the high-stakes cases, reviews samples of the automated conversations, and carries what they see back to product and risk, while a junior agent handles the reserved routine queries and supervised escalations, and works on review alongside the senior.
The junior role is where the lens shows. A cost lens would still have to staff the development goal, but at the minimum possible level: a handful of trainees fed the thinnest possible diet of escalations. The human-centric lens makes the junior role a job worth having, with enough real work in it to produce the senior agents of the future. There are still fewer people than before, because the routine volume is largely gone. But the two roles together form a development pathway, and the entry point stays open.
Understand a job before you eliminate it
To redesign work for the age of AI, you don’t just need to understand AI—you also need to understand work. You need to know what a role is for before you can break its work into tasks and redistribute them between people and AI. That’s why Atlassian, Moderna, and Lumen have each put their head of HR in charge of AI transformation.
As mentioned above, Gartner expects half of the companies that cut jobs for AI to be rehiring for similar work by 2027, though under different titles. That is a redesign done twice—the second time under pressure, leaving the remaining workforce in disarray and the entry points closed. The alternative is to do it once, in the right order: Know what each job is for, decide which of its goals needs a person, and staff those goals as though you will need the people who hold them. Because you will.
