Agentic AI in GCCs: Governed Automation That Survives Past the Pilot
India's Global Capability Centres are increasingly where enterprise AI work gets built. The Nasscom-Zinnov 2026 report puts India at about 28% of global GCC AI talent, second only to the US. That makes GCC leaders natural sponsors of agentic AI: systems that plan and take actions toward a goal rather than only answering prompts.
It also makes them natural candidates for disappointment. Gartner has forecast that over 40% of agentic AI projects will be cancelled by the end of 2027. This article looks at why, and at what a GCC can do differently.
What the evidence says
Gartner's June 2025 prediction attributes the expected cancellations to escalating costs, unclear business value and inadequate risk controls. Its senior director analyst Anushree Verma described most current projects as early-stage experiments or proofs of concept, driven by hype and often misapplied. Gartner also warned of "agent washing", where vendors relabel assistants, chatbots or robotic process automation as agentic, and estimated that only about 130 of thousands of vendors offer genuine agentic capability.
The same forecast is not pessimistic about the technology. Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from 0% in 2024, and 33% of enterprise software applications to include agentic AI by 2028. Its advice is to pursue agentic AI where it delivers clear value or return on investment, and often to rethink workflows from the ground up rather than bolt agents onto existing ones.
Two caveats: these are forecasts, not measured outcomes, and the full Gartner report is behind a paywall. Treat them as informed warnings rather than certainties.
Why pilots succeed and production struggles
The following is analysis, consistent with the Gartner themes but not drawn from a specific study. Pilots usually run with narrow scope, curated data and close human supervision. Production adds messy inputs, legacy systems, unclear ownership and cost at volume. A demonstration that worked on 50 cases may behave inconsistently on 50,000. When nobody has defined who is accountable for an agent's actions, the project loses sponsorship the first time it makes an expensive mistake.
Choosing use cases that can survive
Screen candidates against five questions:
- Is the value measurable? Define the business outcome (time saved, errors reduced, revenue protected) before building.
- Is the task bounded? Clear inputs, clear outputs and a defined set of permitted actions.
- Is the cost of an error tolerable? Start where mistakes are cheap to detect and reverse.
- Is a simpler tool enough? Many use cases labelled agentic can be handled by conventional automation or a well-designed assistant, which are cheaper to run and easier to govern.
- Is the data ready and permitted? Quality, access rights and privacy obligations, including India's DPDP Rules, must be settled first.
Governance that holds up
Governed automation is less a single control than a set of habits that survive scale. Consider building in:
- →Human oversight by design: define which actions an agent may take alone, which need approval, and which are never permitted.
- →Access control: give agents the minimum permissions required, with credentials managed like any other privileged service.
- →Evaluation: test against realistic, representative cases before release, and keep testing afterwards because behaviour drifts as inputs and models change.
- →Logging and traceability: record inputs, decisions and actions so that errors can be investigated and audited.
- →Cost monitoring: track compute, tooling and support costs per outcome, not only total spend.
- →Clear accountability: name a business owner and a technical owner for each agent.
- →Incident handling: treat agent failures as incidents, with runbooks and post-incident review.
The operating model matters as much as the model
The Nasscom-Zinnov 2026 report highlights AI governance as part of the maturity path for GCCs. In practice that points to an operating model with a small platform team that provides shared tooling and standards, product-aligned teams that own specific agents and outcomes, and a review forum linking engineering, risk, security and business leaders. A centre that only runs headquarters' AI requests can deliver well, but a centre that owns AI outcomes and governance is the one positioned to move up the maturity ladder.
A phased approach
Phase 1: Assess
Identify candidate workflows, measure the current baseline, and rank by value and risk.
Phase 2: Prove in a controlled setting
Build one or two bounded agents with human approval for consequential actions, and measure against the baseline.
Phase 3: Industrialise
Introduce shared platform components, evaluation pipelines, monitoring and access controls before scaling usage.
Phase 4: Extend
Add use cases only when the earlier ones show sustained value at acceptable cost and risk.
An illustrative scenario (hypothetical)
A GCC finance technology team considers an agent that drafts reconciliation explanations for analysts. It starts with read-only access, human approval before any posting, and a baseline of analyst hours per case. After three months, results are compared against the baseline before any decision to expand permissions. This scenario is illustrative and not a client case.
Where advisory support fits
Lakhlani International's capability includes assessing opportunities to establish AI engineering capabilities, agentic AI solutions, AI-enabled teams and governed automation workflows aimed at measurable business outcomes, supported by governance, security controls, human oversight and operational accountability.
Technology, Product Development & GCC Setup and Scaling
AI engineering capability, agentic AI solutions and governed automation workflows with measurable outcomes.
Explore the capabilityBook a consultation
To explore whether and where agentic AI makes sense for your centre.
Start a conversationWhy India GCCs stall after launch
The maturity stages, why centres plateau, and why owning AI outcomes moves a centre up the ladder.
Read the maturity guideTechnology blueprint for a new India GCC
Where data and AI capabilities sit in the build sequence, and how DPDP readiness shapes them.
Read the blueprintContact Lakhlani International
For a governed automation question, start with a conversation.
Go to contactFrequently asked questions
What is agentic AI?
Agentic AI refers to systems that can plan and take actions toward goals with some autonomy, rather than only responding to a prompt. Gartner defines it as models that can autonomously perform tasks such as planning, learning and goal achievement.
Why does Gartner expect many agentic AI projects to be cancelled?
Gartner points to escalating costs, unclear business value and inadequate risk controls, and notes that many projects are early experiments driven by hype and misapplied. It also cites vendors relabelling existing products as agentic.
Does that mean GCCs should avoid agentic AI?
No. Gartner's advice is to pursue it where it delivers clear value or return on investment. The point is to select use cases carefully and build governance, not to avoid the technology.
How do we decide whether a use case needs an agent at all?
Test whether conventional automation or a simpler assistant meets the need. Choose an agent only where the task requires planning or multi-step action and the value justifies the added cost and risk.
What governance controls matter most?
Defined permissions and human approval points, minimum-privilege access, evaluation before and after release, logging for traceability, cost monitoring and named accountable owners.
How does DPDP compliance affect AI agents in India?
Agents that process personal data must fit within the data fiduciary's obligations on notice, consent, security and breach reporting, which phase in around mid-May 2027. Confirm applicability to your data flows with legal counsel.
References
- →Gartner prediction, covered by Outlook Business (June 2025). https://www.outlookbusiness.com/artificial-intelligence/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-says-gartner
- →CRN India, Gartner agentic AI forecast. https://www.crn.in/?p=72052
- →OODA Loop, Gartner brief. https://oodaloop.com/briefs/technology/over-40-of-agentic-ai-projects-will-be-scrapped-by-2027-gartner-says
- →Nasscom-Zinnov, GCC Landscape in India 2026. https://media.zinnov.com/wp-content/uploads/2026/05/zinnov-nasscom-india-gcc-landscape-2026-report.pdf
- →Business Today, 7 May 2026 (India share of GCC AI talent). https://www.businesstoday.in/amp/technology/news/story/the-big-boom-how-indias-gcc-story-hits-2030-target-4-years-early-530265-2026-05-07
- →Government of India (PIB), DPDP Rules 2025. https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf