Jayant SwamyChief Enterprise Architect, Genpact is an experienced technology leader and expert in data and cloud technologies. He has more than 20 years of experience with enterprise architecture, AI, data engineering and digital transformation. Before joining Genpact, in 2024 he worked as the CTO of Xtrac8.Tech as well as as the CTO and founder of an AI company focused on technologies like generative AI. Swamy worked for Accenture for more than a decade, holding various leadership roles, such as CTO and Global Lead of the Institute of Applied Intelligence. Swamy also held the positions of Managing Directors and Global Business Leaders for the Data on Cloud and Chief Data Architect for Data Engineering and Data Innovation. He worked for more than 7 years with Fannie Mae as the Senior Principal, overseeing various technology initiatives and information related to credit loss and servicing.
Genpact It is a technology and business service company that focuses on helping businesses transform their complex operations using artificial intelligence, data analytics, process intelligence and domain knowledge. Its origins date back to a 1997 project within GE Capital. The company became independent in 2005 and listed on the New York Stock Exchange as of 2007. Genpact is a global company that works across many industries. These include banking, insurance and finance. They also work with supply chain companies, consumer goods and healthcare.
From enterprise architecture roles and data leadership at Accenture, Fannie Mae and Oracle to your current position at Genpact as Chief Business Architect, you have had a varied career. This experience has influenced your views on where AI autonomy is a true enterprise asset and where human judgement still remains important.
In my professional career, I have learned that innovation is not as important as scale, governance and integration. The experience of enterprise shows that even impressive technologies fail if they can’t be integrated into an organization. Startups reinforce the importance of moving fast, experimenting and maintaining architecture that is simple and flexible. AI autonomy works best when you combine both mindsets — innovating at velocity while maintaining strong foundations. AI has the potential to transform work flow in an organisation. Our findings recent research It is important to stress this point.
Models are only part of the equation. The surrounding architecture — data, systems, workflows, controls, and integrations — determines whether AI creates value at scale. What we’re talking about is “no artificial intelligence without process intelligence.” Genpact Labs’ applied AI is based on the same premise. We take new AI technologies and transform them into client-ready, production-grade solutions. Each capability has a clear business goal and is integrated into workflows, data and control systems.
AI can be used where there is a clear objective, reliable data, and well-defined boundaries. For decisions with high stakes or that require empathy and accountability, human judgment is essential. It’s not about eliminating humans from the system. Designing the system to allow everyone to know where AI is able act autonomously, when humans are required for visibility, and when human judgement must take place.
Most companies measure AI’s impact on customers primarily by speed, cost-savings, and containment rate. How can these metrics give a false impression of the quality of the service?
These are primary measures of performance. The metrics are a good indicator of how well a system performs, but they don’t necessarily reflect how the customers feel. A high containment rate can look good internally but hide bad outcomes when a client gives up or repeats questions, gets incomplete answers, can’t get through to a human, etc. The AI is able to keep the customer within the system by containing them, but not necessarily providing the service they requested. It is the same with cost-savings, as they can compromise 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. The final test for customer-facing artificial intelligence is whether the AI solves end users’ underlying problems accurately, fairly and with adequate human support.
What are the signals that an AI agent should use to identify uncertainty, frustration or urgency?
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. It can manifest as contradictory information, failed attempts or an agent who is going around 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.
System boundaries are also necessary. A different approach should be taken if a problem is not within the workflow or requires data that’s inaccessible, crosses a certain risk threshold, etc. 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.
What can businesses do to distinguish an interaction from one that’s merely complicated and those that require human judgement or empathy?
It’s not the difficulty of the task that makes a difference. The decision’s weight is determined by its judgment, consequences, and context. When setting up a system where machines do the processing and humans check it, you need to understand that human input will never replace machine input.
AI can perform complex tasks with ease when there is a clear objective, reliable data, a defined process, and measurable outcomes. The system can handle complexity without a human, if it understands its boundaries. When the intent of the customer is not clear, when there are several valid options, and/or when the decision has significant consequences in terms of financial, legal or medical implications, a human should be involved. Some interactions also call for reassurance, explanation, negotiation, or empathy — not just an answer. The human element is just as crucial as the technical aspect. This distinction should be reflected in the architecture. Rather than deciding on these points after the fact, they should be built in to the workflow.
What would an effective escalation procedure look like to ensure that customers don’t have repeat their interactions or begin them again once a human agent steps in?
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.
It is mostly an issue of architecture: If the AI sits only in the chat interface, it will be disconnected from CRM, ERP and knowledge sources. Workflows are also not integrated.
What factors, such as customer vulnerability, financial risk and regulatory requirements or potential negative consequences, should influence the escalation levels?
The rules for escalation should be based on the risk level. If there are serious consequences to getting it wrong, a person should get involved sooner.
It is possible to automate a password reset, but it would be better if disputed financial transactions were subject to tighter controls. For claims and collections, this might mean allowing full autonomy for routine updates to status, but mandating a human review when the case crosses an amount threshold, or there is a warning flag. More oversight is needed when the decision has a high level of regulation or significance. An especially vulnerable client may require earlier intervention by a human, while decisions that have irreversible effects should be made with lower autonomy levels.
It’s not a good idea to leave it up to agents to determine these thresholds on their own. The thresholds need to be built in to the architecture, orchestration layer and with clearly defined rules about what agents are authorized to do and what needs additional validation.
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. Businesses that treat every interaction with the same degree of autonomy can quickly become prone to costly mistakes.
What metrics can organizations use beyond task completion to determine if an autonomous system of customer service is delivering positive results for its customers?
Although completing tasks is helpful, it’s not nearly enough. It’s not the question that should be asked. “did the agent finish the task?” The answer is, “did the customer end up in a better place?” The difference is between being accountable and promising a result.
It means going beyond time spent on handling to include metrics like accuracy of response, resolution in the first contact, return contacts, escalated quality, satisfaction with customer service, fairness, consistency, and outcomes. Measuring beyond an individual encounter is also important. Agents can be successful at one stage of the journey, but create a problem further down. A case that is closed quickly can trigger a second call several weeks later. This will impact customer satisfaction and may require human intervention.
In addition, it’s important to determine whether or not the AI system is addressing the client’s core need. This goes beyond simply concluding the conversation and deeming the outcome a success. It’s important to be able to observe the entire workflow and not just the metrics on the AI interface. AI returns are best determined by measuring business and customer outcomes instead of activity.
How can companies prevent AI agents from optimizing their performance for operational goals, like reducing the call volume, handling time or customer needs, to the detriment of fairness or trust?
It is amazing how well AI systems can hit the target that they are given. It’s both a plus and a disadvantage: if you tell a system to only focus on speed, it will do everything it can to get there.
AI must be balanced in the business world. It should not only focus on efficiency but also include customer outcomes and risk. Accuracy, trust, and fairness. To achieve this, businesses must set clear guidelines for acceptable behavior. They should also test AI to ensure it is not biased and has unintended results. The governance of a system cannot be added after it has been in operation. The architecture should include governance from the beginning.
Customers’ context can be scattered across conversations, transaction, enterprise systems, and channels. How can AI agents be given enough context for better decision-making without increasing privacy, governance, and security risks?
It’s not about giving AI agents access to all information. The key is to give it controlled access to relevant information, at the appropriate time. Enterprises require a robust data layer and integration that allows agents to connect with authoritative record systems while controlling the information they can access and what actions each agent is allowed to take. It includes role and purpose-based data access, the minimum amount of information necessary, lineage tracking and audit logs as well as protection for sensitive information.
AI can improve its decisions by adding context, but it only works if the AI is governed. If you solve one problem, it can cause a bigger issue in terms of privacy and security. If handled correctly, the system can learn from prior exceptions and make future interactions even smarter.
When AI is used to interact with customers, what will be the impact on human workers and AI?
Humans will spend less time executing each step in a process and more time directing, monitoring, and optimizing the agents who are taking the workload. The human role will shift from executing every step of a process to directing, supervising, and improving a system of agents taking on the volume. The skills that businesses need will also change. AI fluency, critical thinking and domain expertise will be important. Agent orchestration, observability and the ability of a system to correct itself when it looks incorrect will also matter.
The challenge is not limited to the workforce. The architecture of the organization is needed to facilitate collaboration. This includes connecting agents with the correct data and systems and coordinating their work. Governance, human override and escalation are all things that need to built-in from the beginning.
As chief enterprise architect I believe it is my responsibility to ensure that the organization understands not only how to use AI but also when to question, trust, and control it. This is how architecture can become the gateway to AI with real accountability that builds trust and has real impact.
Please visit our website for more information. Genpact.
