
New Delhi, Nov. 17 -- The more AI agents an enterprise deploys, the harder it could become to predict how they will interact. A system of 15 agents can have 105 possible pairs of interactions. With 1,000 agents and 200 agentic systems, the number rises to nearly 720,000, according to a new report published by IT consulting firm Capgemini on Tuesday.
The challenge is not simply one of scale. As agents interact across applications, data environments and business functions, one system's actions could trigger unintended consequences elsewhere. For CIOs and CTOs, the task is to ensure these interactions remain visible, controlled and accountable.
"With AI agents acting autonomously, AI is reaching a turning point," said Pascal Brier, chief innovation officer at Capgemini. "As we look ahead to 2027, the main challenge for organizations will be how to scale autonomy without losing control, which will require managing interactions between agents, containing unexpected outcomes and responding to threats at machine speed."
The shift raises questions beyond model performance: which agents are operating across enterprise systems, what data and tools can they access, who is accountable for their actions, and how can their activities be monitored or stopped?
The foundations beneath AI The challenge begins with the enterprise technology environment. Fragmented data, legacy applications and weak integration can undermine deployments even when models perform well in pilots.
"The biggest issue is rarely the AI model itself. It is the enterprise environment around it: fragmented data, inconsistent infrastructure, weak integration, poor observability, and security controls designed for traditional applications," said Romanus Prabhu Raymond, director of technology at ManageEngine, Zoho Corp.
A customer-service chatbot, for instance, may perform well in a pilot but struggle in production because customer records sit across incompatible CRM and legacy systems, he said.
Raghavendra Chinhalli, global CIO of Hinduja Global Solutions, has also identified fragmented data, legacy applications and inconsistent processes as barriers to enterprise AI adoption. In an earlier Mint CIO Circle interview, he stressed the importance of building digital foundations to operate AI at scale.
The problem becomes more consequential when AI systems are expected to take action rather than simply generate information. Agents working across procurement, finance or customer operations depend on multiple applications, data sources and approval workflows.
"When an AI service fails, the model may not be the problem. The root cause could be stale data, a broken API, network latency, or infrastructure saturation," Raymond said. Full-stack observability can help teams trace failures across data pipelines, applications, model endpoints, networks and infrastructure.
Enterprises need not complete a wholesale technology overhaul before scaling AI. Modernisation can progress alongside adoption, with investment focused on the systems and data supporting high-value use cases, he said.
The governance challenge As agents gain access to enterprise applications and perform tasks autonomously, identity and permissions become central to risk management.
"AI agents require an identity of their own. Every agent should have defined permissions, least-privilege access, strong authentication, action-level logging, and clear accountability," Raymond said.
A procurement agent, for example, could compare suppliers and prepare a purchase order but require human approval before committing expenditure above a specified threshold. Its identity, tools, decisions and actions should be recorded.
"We need to move from asking, 'Who accessed the system?' to 'Which human or agent made this decision, using what authority, and why?'" he said.
Governance, however, need not mean centralising every decision. Sunit Parekh, CTO at Thoughtworks India, has argued that as AI decisions spread across business units, organisations need shared governance, platform guardrails and workforce upskilling.
Capgemini proposes controls at three levels: individual agents, the systems in which they operate and an enterprise-wide control plane. These include an agent registry, defined identities and permissions, monitoring of outcomes and resource consumption, visibility into interactions, and mechanisms to limit or stop activity.
For CIOs, the priority is to connect AI ambition with operational discipline: strengthen data pipelines and observability, define boundaries for agent access, retain human approval for high-impact decisions and establish repeatable processes for moving successful pilots into production.
The next test may not be how many agents an enterprise can deploy, but whether it can manage their interactions and remain accountable for their actions.
Published by HT Digital Content Services with permission from MINT CIO CIRCLE.