While the promise is immense, many agentic AI projects fall through the gap between PoC and production. Harnath Babu, Partner & CIO, KPMG India discusses why most agentic AI projects get stuck at the PoC stage. He also outlines the playbook for successfully scaling from pilots to production.
Edited Excerpts:
Q. Industry reports indicate that a vast majority of agentic AI projects fail to scale beyond pilots? Why is that?
Most agentic AI projects fail to scale as organizations mistake demonstration for deployment. Pilots are usually undertaken in controlled environments with clean data, tolerant stakeholders and human-in-the-loop safety nets. However, when they move to production, harsh realities around fragmented data and weak integrations are exposed. The operating models are traditionally not designed for autonomous systems. Another key factor is unclear accountability for when an agent takes the decision and
Agentic AI requires rethinking how decisions are made, owned, audited and improved at enterprise scale. And, organizations are still catching up on that.
Q. What are the key success factors for scaling up agentic AI projects from pilot to production?
Programs that are built for scale from the start are more likely to be successful. Critical factors that determine project success include decision ownership, solid data foundations, clearly defined autonomy boundaries, production-grade governance, monitoring, auditability, backup plans and cost controls. Instead of being retrofitted these must be baked in right at the outset.
Q. What would you suggest as the roadmap for effectively scaling up agentic AI projects?
Internally facing agents are a good point to start as that helps with building trust, and enabling observability and control. Once the guardrails have matured, their scope can be broadened further.
Step-wise, it may look something like the following -prove value, move towards hardening the platform, institutionalize governance, and finally scale autonomy across higher‑stakes workflows. This creates a playbook that can be replicated and adapted in the future.
Q. Can you share some high-impact agentic AI use cases that are already live at KPMG India?
We are currently developing agentic Al solutions across a few high-impact domains such as Contract Analysis Agent, Risk Management Orchestrator and Hiring AI Agent.
The focus is on augmenting expert judgement, enabling faster and consistent decision-making, and creating a structured auditability for critical processes.
While Contract Analysis Agent reviews contracts, extracts key obligations and identifies risk exposures, the Risk Management Orchestrator synthesizes signals across systems to identify and prioritize emerging risks. The Hiring AI Agent streamlines candidate screening, skills mapping and interview workflows.
Q. Going forward in your agentic AI journey, which are the areas you see strong potential?
Looking ahead, we see strong potential in an integrated GRC agent that is capable of interpreting policies, monitoring control effectiveness and initiating compliance workflows autonomously. We are also looking at coordinated agents across HR, finance, client-facing teams and technology, which can help move from isolated task automation to a more connected decisions ecosystem.
In the long-term, we are looking to embed multiple specialized agents across enterprise functions, aimed at enabling faster execution, stronger governance and resilient operating models.
Q. How should businesses go about choosing and prioritizing the right use cases to set up agentic AI for success?
The decision is based on a combination of factors such as business impact, technical readiness and risk posture. The use cases must be evaluated on measurable outcomes, data maturity, integration feasibility and governance readiness.
Mature leaders are moving from experimentation to purposeful investment, starting with a clear articulation of where it creates differentiated value, and not just efficiency.
The strategy is to tie agentic AI to business capabilities, not technology. Decisions that are repeatable, data-rich, time-sensitive, and currently constrained by human bandwidth are prime candidates.
