One of the most consequential questions I face as a CEO is this: When work needs to get done, should we hire a person, deploy an AI agent, or redesign how the work gets done?
We usually describe AI transformation through its most straightforward moments: the initial technology investment and the following productivity gains. Far fewer people pay attention to what happens between those milestones, as workflows, roles, institutional knowledge, management practices, and customer expectations collide.
I know this firsthand because I’m leading an established, historically people- and process-intensive services business through that messy middle. I believe that if an AI strategy doesn’t include the necessary steps of redesigning jobs, retraining managers, and changing how decisions get made, then the company isn’t transforming around AI—it’s just adding software.
Redesign the workflow, not just the task
McKinsey’s 2025 State of AI survey found organizations generating the most value with AI were nearly three times as likely as others to have fundamentally redesigned workflows.
That requires separating tasks from roles. Rather than asking which jobs AI can replace, I start with the required outcome. Then I break the work into three components: execution, exceptions, and accountability.
From there, we can determine what an AI agent can handle, where a person needs to step in, and whether the work should be organized differently altogether.
In addition to separating the task and role, it’s important to create a reinforcing learning system, which we applied to employment-certificate requests. Previously, a specialist manually copied employee information into a template, created a PDF, signed it, and sent it to the employee. The process took at least 20 minutes.
Now, an AI-powered workflow uses a contract ID to retrieve the relevant information and generate a completed certificate within seconds. A person still reviews, signs, and sends it.
We automated the repetitive work, not the accountability. Keeping a human in the loop helps us catch issues and improve the system over time.
Turn institutional knowledge into infrastructure
Redesigning a workflow often reveals a deeper problem: The organization’s knowledge was never designed to function as a system.
In a services business, experienced employees routinely compensate for fragmented information, incomplete documentation, and conflicting interpretations. The ability to handle ambiguous scenarios is what separates companies with high customer satisfaction scores from those that simply automate. Users dreaded previous generations of chatbots precisely because they couldn’t navigate that ambiguity.
An AI agent can’t automatically inherit that context. When you make an employee’s knowledge accessible to the AI agent, each error and exception becomes an opportunity for the system to learn and give customers better answers.
An employee in Germany asked for guidance after taking a sick day. The initial response incorrectly requested a doctor’s note. Under German law, employees only need a note when an illness lasts longer than three consecutive calendar days. We corrected the underlying guidance and changed the governance process so the correction would carry into future interactions. Our experience team now validates policy changes, resolves ambiguity, and determines when new guidance enters the system.
Correcting one response improves one interaction. Correcting the underlying information, assigning ownership, and updating the source of truth improves every interaction that follows. Our experienced employees answer difficult questions and teach the system to recognize and resolve them too.
Prepare and hire people for good judgment
Alongside technologists, managers also determine whether redesigned work succeeds. They decide when employees can trust an AI-supported answer, when they should challenge it, and when a case requires escalation. They must also evaluate contributions differently: When polished output becomes abundant, identifying a flawed assumption or unusual case may be more valuable than producing more work.
During an internal hackathon, an onboarding specialist—someone who helps employees get set up on our platform—examined a manually maintained tracker containing hundreds of rows. She built a dashboard that flags missing information and alerts the team when an onboarding appears complete, but banking information is still missing. The value was in making exceptions visible early enough for people to act.
Managers therefore need more than training in how to use AI. They must learn to evaluate the reasoning as well as the output, empower employees to challenge AI-supported answers, and give customers a clear route to a person.
The same emphasis on judgment should shape hiring and retraining. Companies can give candidates and employees imperfect AI-generated recommendations and ask what they would verify, which assumptions they would challenge, and when they would escalate. The goal is to develop people who recognize when a convincing answer may be wrong.
Build an organization that can keep up with AI
The question I began with—whether to hire a person, deploy an AI agent, or redesign the work—doesn’t have a permanent answer. As AI technology improves, the boundary between human and machine responsibility will keep moving.
Companies that build organizations capable of shifting with that boundary while improving services, strengthening customer relationships, and keeping accountability clear will get better at transforming themselves.
Francoise Brougher is the chief executive officer of Pebl.