Jayant Swamy, Chief Enterprise Architect at Genpact, is a technology and data leader with more than two decades of experience spanning enterprise architecture, artificial intelligence, data engineering, cloud technologies, and large-scale digital transformation. Before joining Genpact in 2024, he served as CTO and Chief Architect at Xtrac8.Tech and as CTO and Co-Founder of an AI startup focused on technologies including generative AI and large language models. Swamy previously spent more than a decade at Accenture, where his leadership roles included CTO, Global Lead and Global Managing Director of the Institute of Applied Intelligence, Managing Director and Global Business Lead for the Data on Cloud business, and Chief Data Architect and Global Lead for Data Engineering and Data Innovation. Earlier in his career, he spent more than seven years at Fannie Mae overseeing technology and business information initiatives related to servicing and credit loss management, following several years as a Senior Principal at Oracle (ORCL ).
Genpact is a global technology and business services company focused on helping enterprises transform complex operations through artificial intelligence, data, process intelligence, and domain expertise. The company traces its origins to a 1997 initiative within GE Capital, became independent in 2005, and went public on the New York Stock Exchange in 2007. Today, Genpact works with organizations across industries including banking, insurance, finance, supply chain, consumer goods, healthcare, and technology, with a growing emphasis on generative AI, agentic systems, automation, and AI-enabled enterprise operations.
Your career has taken you from enterprise architecture and data leadership roles at Oracle, Fannie Mae, and Accenture to startup environments and your current position as Chief Enterprise Architect at Genpact. How has that experience shaped your view of where AI autonomy creates genuine enterprise value and where human judgment remains essential?
Throughout my career, I’ve learned that scale, governance, integration, and risk matter as much as innovation. Enterprise experience shows that impressive technology still fails if it can’t work within the organization; startups reinforce the value of moving quickly, experimenting, and keeping architecture simple and adaptable. AI autonomy works best when you combine both mindsets — innovating at velocity while maintaining strong foundations. Thay way, AI can truly transform how work flows in an organization. The findings from our recent research reinforce that point clearly.
The model itself is only one part of the equation. The surrounding architecture — data, systems, workflows, controls, and integrations — determines whether AI creates value at scale. That’s what we mean by “no artificial intelligence without process intelligence.” It’s also the premise behind our applied AI work in Genpact Labs, where we take emerging AI capabilities and turn them into production-grade, client-ready solutions. Every capability is tied to a measurable business outcome and wired into the workflow, data, and controls needed to operate.
AI should get room to act where the objective is clear, the data is reliable, and the boundaries are well understood. Human judgment remains essential for decisions that are high stakes, ambiguous, subjective, regulated, or require accountability and empathy. Autonomy isn’t about removing humans from the system. It’s about designing the system so everyone knows where AI can act independently, where humans need visibility, and where human judgment must take over.
Many companies evaluate customer-facing AI primarily through speed, containment rates, and cost savings. Why can these metrics create a misleading picture of whether the technology is actually improving the customer experience?
The issue is that those are primarily measures of efficiency. They show you how the system performed, but not necessarily how the customer experienced it. Take containment: a high rate may look successful internally, but it can hide poor outcomes if a customer gives up, repeats themselves, receives incomplete answers, or can’t reach a person when needed. Containment shows the AI kept the customer inside the system, not that they got what they needed. The same applies to cost savings, which can come at the expense of trust, satisfaction, or resolution quality.
AI systems optimize around the signals you give them, so enterprises need to design measurement cues into the system from the start — connecting front-end metrics like speed and containment to downstream signals like issue resolution and whether the outcome created problems elsewhere. Ultimately, customer-facing AI should be measured on whether it solves the end user’s underlying problem accurately, fairly, and with the right level of human support.
What signals should an AI agent use to recognize uncertainty, frustration, urgency, or a situation that falls outside its established workflow?
There’s no single magic signal — what matters is giving the system enough context and observability to recognize when its confidence, authority, or information is no longer sufficient. That can show up as conflicting information, repeated failed attempts, or the agent going in circles. It can also be a shift in the customer’s language — a clear signal of frustration or urgency — or the task itself requiring information, access, or authority the agent doesn’t have.
The system also needs to know its own boundaries. If an issue falls outside its designed workflow, needs inaccessible data, or crosses a defined risk threshold, that should trigger a different path. The goal isn’t for the agent to handle every possible scenario — it’s for it to know when to stop, escalate, or ask for help.
How can companies distinguish between an interaction that is merely complex and one that genuinely requires human judgment or empathy?
The dividing line isn’t how difficult the task is. It’s how much judgment, consequence, and human context the decision carries. Setting up agentic operations where machines process and humans validate requires understanding where human value can never be replaced.
AI can handle very complex tasks when the objective is clear, the data is reliable, the process is defined, and the outcomes are measurable. Complexity on its own doesn’t require a human if the system understands the boundaries within which it is allowed to operate. A human needs to be brought in when the customer’s intent is unclear, the situation requires discretion, there are multiple valid outcomes, or the decision carries significant financial, legal, medical, or reputational consequences. Some interactions also call for reassurance, explanation, negotiation, or empathy — not just an answer. Human involvement there is as important as technical accuracy. The architecture should reflect that distinction. Those intervention points should be designed into the workflow upfront, not decided after something goes wrong.
What should an effective escalation process look like so that customers do not have to repeat themselves or start the interaction again when a human agent takes over?
A good escalation isn’t just transferring the customer — it’s transferring context, so the next person can move straight toward resolution. When a human takes over, they should already have the conversation history, the customer’s identity and profile, and what the AI has already tried — enough to avoid asking the customer to start over.
This is largely an architecture issue: if the AI only sits in the front-end chat experience, disconnected from the CRM, ERP, knowledge sources, workflows, and orchestration layer, a seamless handoff becomes impossible.
Should escalation thresholds change depending on factors such as financial risk, customer vulnerability, regulatory requirements, or the potential consequences of an incorrect decision?
Escalation rules should depend on the level of risk. The greater the consequences of getting something wrong, the sooner a human should be involved.
A password reset can run fully automated, but a disputed financial transaction should have tighter controls. In claims or collections, that might mean full autonomy on routine status updates, but a mandatory human check when a case crosses a dollar threshold or a hardship flag comes up. The more regulated or consequential the decision, the stronger the oversight should be. A vulnerable customer may warrant earlier human intervention, and decisions with irreversible consequences should have lower autonomy thresholds.
Those thresholds also shouldn’t be left to individual agents to interpret on the fly. They need to be built into the architecture and orchestration layer, with clear rules around what the agent is authorized to do, what requires additional validation, and what must be escalated.
The opportunity is to let autonomy scale with risk — full speed on the low-stakes work, tighter control where it counts. At Genpact, we’ve built that tiering directly into how we design agentic workflows — letting routine cases run autonomously while routing anything that crosses a defined risk or dollar threshold to a human, automatically and without the agent having to decide to escalate. Enterprises that give every interaction the same level of autonomy quickly turn convenience into costly errors.
Beyond task completion, which metrics should organizations use to evaluate whether an autonomous customer service system is producing positive outcomes for customers?
Task completion is useful, but it is nowhere near enough. The question shouldn’t be, “did the agent finish the task?” It should be, “did the customer end up in a better place?” It’s the difference between promising an outcome and being accountable for it.
That means looking beyond handling time to metrics like response accuracy, first-contact resolution, repeat contacts, escalation quality, customer satisfaction, and fairness and consistency of outcomes. It also means measuring beyond an individual interaction. An agent can appear successful in one part of the customer journey while creating a problem downstream. For example, a case that closes fast can still trigger another call weeks later, impacting customer sentiment and eventually requiring human remediation, so it’s worth tracking whether it’s still resolved 30 days later.
It’s also important to know whether the system actually resolves the customer’s underlying need, rather than simply closing the interaction and considering it a win. That requires end-to-end observability across the workflow, not just metrics from the AI interface. Measuring customer and business outcomes, rather than activity, is what reveals AI’s real ROI.
How can businesses prevent AI agents from optimizing for operational targets, such as reducing call volume or handling time, at the expense of fairness, trust, or the customer’s underlying needs?
AI systems are remarkably good at hitting whatever target they get assigned. That’s the upside and the catch: tell an agentic system that speed is the only goal, and it will sacrifice everything else to achieve it.
Businesses need to give AI a balanced set of objectives where efficiency matters, but it also sits alongside customer outcomes, risk, fairness, accuracy, and trust. This is achieved by putting clear guardrails around acceptable behavior, testing for bias and unintended outcomes, maintaining human override mechanisms, continuously monitoring decisions, and making outputs traceable and auditable. Governance cannot be something you tack on after the system is already running. It should be built into the architecture from the start.
Customer context may be spread across previous conversations, transactions, channels, and enterprise systems. How can companies give AI agents enough context to make better decisions without creating new privacy, security, or governance risks?
The answer to fragmented context isn’t giving an AI agent access to everything. It’s giving it governed access to the right information at the right moment. Enterprises need a strong data and integration layer that connects agents to authoritative systems of record while controlling what each agent can see and do. That includes role- and purpose-based access, minimum necessary data, lineage tracking, audit logs, and protection of sensitive information.
More context improves AI decisions, but only if it’s governed. Otherwise, solving one problem can create a much bigger privacy or security issue. Handled well, context compounds as the system learns from previous exceptions, making future interactions smarter without creating new risks.
As customer-facing AI becomes more autonomous, how do you expect the relationship between AI agents and human employees to evolve, and what capabilities will organizations need to build now to make that collaboration successful?
The human role will increasingly shift from executing every step of a process to directing, supervising, and improving the system of agents taking on the volume of work. People will spend more time setting objectives, handling exceptions, exercising judgment, validating high-risk decisions, refining workflows, monitoring agent behavior, and improving the architecture and governance behind those agentic systems. That changes the skills companies need as well. AI fluency will matter, but so will critical thinking, domain expertise, agent orchestration, observability, and the ability to challenge an AI system when something looks wrong.
This isn’t only a workforce challenge. Organizations need the architecture to support that collaboration, connecting agents to the right data and systems, coordinating how they work together, and giving people visibility into what they are doing. Governance, human override, and escalation all need to be designed in from the start.
In my role as chief enterprise architect, I see it as my duty to make sure the entire organization knows not just how to use AI, but when to trust it, question it, and take control. In this way, architecture becomes the doorway to accountable AI that drives trust and real impact.
Thank you for the great interview, readers who wish to learn more should visit Genpact.
