AI Is Making The Handoff Optional
Swapnil Upganlawar, Senior Leader in generative AI at Amazon Web Services (AWS).
gettySome of the boundaries that AI agents are beginning to cross are the same ones I spent years building enterprise software to manage.
Early in my career, I worked on enterprise service-management products built around ITIL, a framework for IT service management. Machine learning was already part of my work. These products supported interconnected business services, where changing one component could affect others.
Building for those environments meant thinking beyond whether an individual operation succeeded. Someone needed to understand its effect on the service, own the consequences and know how to recover when something went wrong.
That experience gave me respect for organizational boundaries. Different teams held different expertise and authority. Passing work between them could be essential. But it also made the dependency between technology and process hard to miss: When systems could not share information, people had to carry it.
As my ML work expanded into recommendation and personalization systems, I began addressing another part of that problem. We built ways to help people find relevant knowledge and make better decisions without searching through everything themselves. A useful recommendation still left a person with the work of acting on it.
My responsibilities also spanned program management and engineering. I saw the work between a good idea and a delivered result. Resolving dependencies, negotiating priorities and deciding which commitments could be kept were substantial parts of the job. So was the less rewarding work of collecting information from one group and translating it for another. Some of that coordination requires leadership. Some exists because information and systems are fragmented.
That distinction became more consequential when I led AI transformation across the employee life cycle for a workforce of 1.3 million people. The work spanned onboarding, ongoing support and departure. I led teams across science, engineering, product and program management, bringing responsibility for the employee experience and its supporting systems into the same leadership scope. We were now trying to resolve needs that crossed several departments.
Imagine an employee moving into a new role. The reporting line changes, but access permissions do not. The employee supplies the details to one team, then another asks for them again. Each department may be completing its part correctly. The person still cannot get started. That is why I think about the whole journey when assessing an AI improvement. A better answer is valuable. The employee also needs the organization to follow through.
Today, I lead a portfolio of five enterprise AI products spanning intelligent search, managed knowledge retrieval, business assistants and agentic workflows. Search helps locate knowledge scattered across systems. Managed retrieval helps ground AI responses in organizational information. Business assistants make that knowledge easier to use. Agents can connect it to actions, as long as they have the right tools and authority.
My job is to make those connections dependable. The goal is to let other teams build useful AI experiences without reconstructing the foundations for finding and using enterprise knowledge each time.
Recent agent releases combine persistent execution, reusable skills and context across applications. Those advertised capabilities need testing in each workflow, but they make it reasonable to ask whether another handoff is necessary.
We could have removed some of that work years ago. AI should not get credit for a process we never bothered to simplify. The more interesting opportunity involves requests that arrive incomplete, require interpretation or depend on several kinds of expertise.
A 2025 product-development experiment gives us one reason to examine those boundaries. Individuals using AI matched the performance of two-person teams without it and spent 16.4% less time than the control group. AI also helped participants combine technical and commercial perspectives. That is evidence about a particular task, not a staffing formula. It suggests that some work we habitually divide into successive contributions may be worth organizing differently.
The coordination can also reappear elsewhere. An October 2026 study inside an AI developer found that βmeta-work,β including managing agents and building supervision tools, rose from 10% to 22% of employeesβ interactive AI sessions. Meeting time among technical individual contributors increased 16% over the previous year. The findings describe one rapidly growing workplace and do not establish that AI caused the increase in meetings. Before calling a workflow simpler, I want to know where that work went.
Having led both program management and engineering, I expect some responsibilities we currently separate to come together. Someone who can use AI to investigate a problem, test an approach and coordinate routine execution may be able to own more of the outcome. That requires changes in authority and incentives. If people are rewarded only for completing their departmentβs portion, they have little reason to simplify the whole process. I would give the person responsible for the outcome the authority to challenge unnecessary handoffs and judge the result across the entire journey.
For employees changing roles, I would review each handoff with the people responsible for it. Reentering a verified start date is a reasonable candidate for removal. Authorizing access to sensitive records needs a different standard. The person performing a seemingly redundant check may know exactly which mistake it prevents. I would also keep someone accountable for the complete request, including exceptions. Otherwise, departmental queues could become agent queues, with the employee still chasing progress. The measures should follow that person: time to completion, repeated requests for information, corrections and whether they can actually do their new job.
My early experience building enterprise service-management products still shapes this approach. Ownership, reliability and recovery remain essential. What deserves reconsideration is every transfer that adds no expertise, decision or protection.
If AI makes every team faster but the employee still waits just as long, what have we actually improved?
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?