AI Agents create value when the work is first redesigned, not when tools are added.

Most organizations have now adopted some form of AI tools, but many are just beginning to adopt AI Agents. Far fewer can show an effect on earnings. In McKinsey’s 2025 State of AI survey, 39 percent of respondents attributed any enterprise-level EBIT impact to AI, and about 6 percent reported at least 5 percent of EBIT attributable to AI and described its value as significant.

One reason is that most business cases measure a single source of value: cost savings. That lens is necessary but incomplete. In my experience, AI Agents should be measured against three lenses, and the workflow should be redesigned with all three in view before anything is built.

Among the organizations I work with, those that apply all three lenses to a redesigned workflow, with the baseline agreed with finance before the build, set themselves up for a material return. The three examples below are drawn from organizations I work with through Ampliwork across the globe. Each was chosen because it illustrates one lens clearly.

Cost savings. A global professional services firm relied on a large outsourced group to run its tax operations. It has replaced much of that group with AI Agents, each performing the work of five contractors, and improved the operating margin of its tax operations by 25 percent relative to its previous level.

Cost avoidance. A global asset manager faced slow client onboarding, with frequent errors and costly rework. It has redesigned the process around AI Agents, reducing processing time by 75 percent and errors by 60 percent, and with them most of the rework.

Revenue generation. A national train operator could put tickets on sale only once each new timetable had been checked by hand. AI Agents now validate timetables faster and more reliably, extending the ticket retail window from 12 to 32 weeks and generating millions in new revenue.

One further discipline separates material returns from marginal ones: redesigning the workflow rather than the task. An agent added to an unchanged process typically moves the bottleneck rather than removing it.

AI Agents now act inside live systems. They update records, route cases and prepare decisions that affect customers and money. That capability is what makes them valuable, and it is also why they must operate under human oversight. In my view, a controlled environment overseen by people is the only responsible way to deploy agents at scale. It is the model I advocate and implement with the organizations I work with.

Oversight must be real rather than nominal. A person asked to approve thousands of outputs a day will, over time, approve what they no longer have time to examine. A controlled environment therefore rests on five elements.

1. Defined boundaries. The agent operates within the organization’s own environment, with specified systems, data and permissions, and nothing beyond them.
2. Decision rights set by consequence. Routine, reversible actions run within agreed limits. Actions that affect customers, safety or money require human approval.
3. Escalation by design. When the agent’s confidence is low, or a case falls outside its rules, it routes the case to a named person with the context needed to decide.
4. A complete record. Every action and the reasoning behind it is logged, so that decisions can be audited and the system improved.
5. Human accountability. The agent is tested before it goes live, monitored in operation, and a named person owns the outcome.

Independent research supports this model. Stanford’s study of 1,500 workers found that equal partnership between person and agent was the preferred arrangement in 47 of 104 occupations studied. Harvard Business School’s field experiment at Procter & Gamble found that individuals working with AI performed at the level of two-person teams without it. Neither finding supports removing people from the work; both support redesigning what people do.

The first effects of AI on employment are falling on those starting their careers. Stanford’s 2025 research using payroll data found a 13 percent relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, while more experienced workers in the same roles were unaffected. The decline was concentrated where AI automates work rather than augments it.

Entry-level roles have long been where people learn judgment, by doing routine work under supervision. As agents absorb that work, graduates must arrive able to do what the first years of a career once taught: identify where an agent belongs, measure what it returns, test whether it can be trusted and supervise its work. These capabilities are built through practice, not lecture.

Universities must therefore integrate the practical application of AI Agents into the curriculum, and some are leading the way. The University of Miami Business School is one of them, and I developed and teach its first agentic AI course for business undergraduates. Students learn to decide what is worth building, to design and build AI Agents, and to test whether their work can be relied upon. The aim is to graduate AI practitioners, not only informed observers. I regard this as one of the most important contributions I can make to the next generation.

Employers carry a parallel obligation: to be explicit about where agents will redeploy capacity and where they will reduce it, and to design new routes to expertise for those entering the workforce.