For Nitin Garg, the potential of agentic AI goes well beyond generating content or answering questions. Drawing on more than two decades of experience in enterprise technology, business strategy, and product management, he sees an opportunity to fundamentally change how organizations manage complex workflows, connect disparate systems, and automate tasks that have traditionally required significant human involvement.
As a technical product management leader, Garg has developed and launched enterprise software products designed to address operational challenges at scale. His recent work focuses on agentic AI and workflow automation, including building capabilities that allow AI agents to interact with multiple business applications, access relevant information, and execute tasks on behalf of users. His combined experience in engineering, business operations, and product development informs his approach to identifying where automation can deliver measurable business value.
In this interview, Garg discusses what distinguishes agentic AI from traditional automation, the challenges of developing secure and reliable AI agents, and how businesses can determine which processes are worth automating. He also shares his perspective on the importance of trust, governance, and human oversight as organizations begin giving AI systems greater autonomy.
Can you tell us a little about your professional background and how your experience in engineering, business strategy, and product management led you to focus on AI and enterprise workflow automation?
I spent the first 8 years of my career with Tata Consultancy Services, a global leader in information technology services. At TCS, I honed my “technology” skills while dealing with deep rooted IT problems and scaling opportunities in enterprise systems for our partner Eli Lilly,
During my MBA, I had an opportunity to intern and then take on a full-time role with Deloitte Consulting. At Deloitte and subsequently with PE owned operating companies, I honed my business and strategy skills. This unique intersection of technology and business skills catapulted me towards a product management role. I have spent the last 6 years honing my product management skills and have focused on AI based projects for the past couple of years.
AI is currently reshaping and enabling us all to re-imagine every aspect of the business processes across varied industries. These business processes at the end of the day are small and large workflows, which are in-part automated (via underlying enterprise systems) and in-part managed by humans. My expertise in program management and process optimization ,with a PMP certification and a Six Sigma Black Belt certification, along with my deep product management skills have organically led me to the space of AI and enterprise workflow automation, where I believe I am uniquely qualified and positioned to make a big impact.
Agentic AI has become a major topic in enterprise technology. How does it differ from traditional automation and generative AI, and what can businesses accomplish with these systems that wasn’t possible before?
Traditional automation has always been heavily dependent on detailed standard operating procedures, translated into structured algorithms and code, that led to a deterministic outcome. These automations served us well and fueled two decades of technology advancement. The onset of generative AI enabled us to harness the powers of LLM to create new content and ideas, conversations, stories, images, videos, and most importantly software/code.
These models can now be trained to become experts at specific tasks, for example software development and coding agents, and can be grounded in an enterprise’s business context and years of data. Agentic AI now provides these AI systems the agency to take actions on behalf of the business user (e.g. by assuming their identity). Agentic AI does so via a combination of skills, actions, and connectors. In my mental model, this completes the full loop where an AI agent can take action on behalf of the user, while keeping the human In the loop, as configured. This was never possible before and we are just at the beginning of this exciting transformation journey ahead.
You’ve worked on developing AI products that can complete tasks across multiple business applications. What are some of the biggest technical challenges you’ve encountered, particularly when it comes to connecting different systems and getting AI agents to carry out tasks accurately?
One of the toughest problems while designing the connectors for AI agents that enable them to take action, was solving for the identify and access management (IAM) for the agents. For a connector to be able to take an action, we needed to resolve IAM policies across multiple organizations and enterprise systems. In addition, the actions taken by AI agents need to be “on behalf of the user so it can leverage the user’s identity and access in each of the enterprise systems to ensure trust, security, and traceability. At one time, this required chaining through six different services. It worked but was slow and it was limited in the sense that it did not scale universally to other AI applications.
As the technology evolved, model context protocol emerged as the leading open-source standard, eliminating the need for every AI application to have its own custom connector
With so many potential uses for agentic AI, how do you determine which business processes are worth automating? What factors do you consider when evaluating whether a particular use case will actually save time, reduce costs, or improve productivity?
I recommend starting with strategic alignment and organizational fit. The process should support the broader company goals and list the business processes that have the potential to yield the maximum benefit to the enterprise (e.g. hours saved etc.). Then you have to assess the feasibility of the automation based on the AI applications you plan to use. You should also factor in the acceptable error rate to know the tolerance levels the AI application needs to operate in. You can then organize these business processes in a quadrant and I recommend starting with high potential and high feasibility projects first to get the maximum return on your efforts.
One of the challenges with enterprise automation is that important information is often spread across different applications, databases, and documents. How have you approached building products that allow AI agents to access that information and use it to complete more complicated workflows?
Honestly, this is exactly the problem that Agentic AI is good at solving. The AI agent can use the varied Model Context Protocol (MCP) connectors to tap into each of those distributed systems and databases. This enables the agent to bring in all that data and knowledge into the agent’s context and memory, so it can yield significantly better results. Also, it can chain through various MCP connectors to not only read information but also engage in writing actions to make updates to other sources. This enables these AI agents to connect to the varied business tools and systems and complete tasks end to end.
As companies begin allowing AI agents to make decisions and take actions with less direct human involvement, what concerns should they be thinking about? How do you build safeguards into these products without limiting the benefits of automation?
Trust and traceability are the most important pieces to get right. An audit trail of all actions taken by the AI agent should be readily available in any well architected Agentic AI system. When an AI agent causes financial harm, security breach, or a system outage, the system logs should be able to trace a specific sequence of actions the agent took. An AI agent may make mistakes by optimizing for a sub-goal in ways that violate the broader, implicit business policies, so building the necessary human in the loop checks are required. As the technology keeps improving and the agents can learn from their knowledge, over the next 3 years to 5 years, these HILO controls and thresholds can be adjusted to ensure the appropriate governance is applied.
You’ve been involved in bringing new enterprise products from the initial concept through development and launch. Can you share an example of a feature you’ve worked on, the problem it was designed to solve, and some of the decisions you made along the way? What were the most important lessons from that experience?
One of my projects was to launch a rules feature in an enterprise-wide headcount management and hiring system. These rules were highly configurable and would enable orgs to set the required guardrails to define the “size of the org” (e.g. the number of people) and the shape of the org (e.g. the right location mix, job level mix, job family mix etc.). Such guardrails would restrict hiring managers from opening roles that violate the underlying hiring rules for the org. One of the key challenges was that these rules could be applied at any level within the company hierarchy. So at the time of rule application, we needed to traverse up the organizational hierarchy to find all overlapping rules and apply only the highest order rule. As I was designing this, based on my customer discovery sessions, I also added features to carve out exceptions to the rule (e.g. apply Rule A to all except those roles at Location = X) to accommodate for edge cases, I also added an approval chain to request exceptions. This set of features enables the hiring managers to open roles that are aligned with the prg’s hiring guardrails but request approvals for exceptions adding flexibility. The feature scaled quickly after launch, and has improved the hiring guardrails for our organizations across the company.
Looking ahead, where do you think agentic AI will make the biggest difference in enterprise operations over the next few years? What needs to improve before companies can confidently rely on AI agents to manage more of their everyday business processes?
I think the technology is ready and available already to make a significant impact. It’s now up to the organizations and their leadership’s willingness to adopt. Invariably, there will be organizations that will adopt early and those that do not. What needs to improve now is human trust and acceptance, which in turn will drive adoption. A
Agentic AI will continue to improve and now we are capable of giving it memory and create feedback loops that enable it to learn from its mistakes so with time these AI agents will only get better. Tech organizations are actively experimenting and already embracing this technology with over 90% of new software code now being written by AI at several leading tech corporations. The rest of the industries are just getting started in leveraging this new wave of innovation and am excited for the times ahead.
