By Shafreen Anfar, Associate Director & Architect, WSO2
Every enterprise conversation about AI in India today circles back to one question: why isn’t it working the way we expected?
The instinct is to blame the technology. Maybe the model isn’t accurate enough. Maybe the use case was too ambitious. But that’s rarely the real story. According to EY India’s AIdea of India Outlook 2026 survey of enterprise leaders, integration and data readiness is the single biggest barrier to scaling AI in Indian enterprises today. 78% respondents named integration challenges as a top barrier, and 53% called it a “severe” problem, specifically when moving from pilot to scale.
This mirrors what we consistently see across enterprise AI deployment. Pilots perform well in a controlled environment. The difficulty surfaces once they need to operate inside the business, connected to live systems and real workflows.
A dispatch agent, and where it breaks
Consider a common scenario in India’s logistics and roadside-assistance space: an AI agent built to coordinate between a customer, a dispatcher, and a service provider, say, when a vehicle breaks down and needs a tow truck or technician. On paper, this is exactly the kind of task AI should be good at.
In practice, the agent needs to talk to several systems at once: routing and provider-availability data, the customer’s service history, SMS or WhatsApp for real-time updates, and an approval step if the repair cost crosses a certain threshold. An agent built without this integration can perform convincingly in a demonstration but often fails to hold up in a live enterprise environment, where systems are inconsistent and workflows are less predictable. The cause is typically not the agent’s capability, but the absence of proper integration.
Integration has to change shape, not disappear
This doesn’t mean integration goes away in the AI era. It means it must work differently.
Traditional integration was built for predictable, rule-based systems: define the workflow once, and it runs the same way every time. AI agents don’t work like that. They reason through a task in real time, deciding what to do next based on the situation in front of them. The systems connecting them need to support that: dynamic rather than fixed and built around giving agents the right tools and context, not just an endpoint to call.
Some of the few things that matter most are, first, how agents are triggered needs to broaden. Enterprises are used to structured triggers, an API call, a scheduled job. Agents increasingly need to respond to unstructured ones too: a chat message, or increasingly in India, a voice call, especially relevant as voice-first interfaces gain ground for customer service and support use cases here. Second, agents need real context: memory of past interactions and access to relevant knowledge, not decisions made in a vacuum. Third, they need a human in the loop for anything high-stakes. An agent approving a large repair cost or an insurance payout should escalate, not decide alone.
Two more pieces are easy to skip in the rush to deploy, and both matter more in India than the talk circuit usually admits. Agents need governance and audit trails built in from day one. With the DPDP Act now shaping how personal data can be processed and by whom, an AI agent making autonomous decisions using customer data isn’t just a technical risk, it’s a compliance one. Every agent needs a clear identity, defined permissions, and a record of what it accessed and why, the same discipline enterprises already apply to human employees handling sensitive data.
And because agents are inherently unpredictable, doing things slightly differently each time even with the same input, enterprises need real visibility into what an agent actually did, not just whether the task succeeded. Without that observability, a failure is nearly impossible to diagnose, and trust in the system erodes fast.
The fix isn’t a better model. It’s better plumbing.
None of this is about picking a smarter AI model. The frontier models available today are genuinely capable. The gap is what happens between the model and the rest of the business: whether the agent can see what it needs to see, act where it needs to act, and be held accountable when it does.
Indian enterprises have an advantage here that’s easy to overlook; decades of experience building complex, high-scale integration layers for banking, logistics, and retail systems that few other markets have had to solve at this volume. That expertise doesn’t become irrelevant in the AI era. It becomes the foundation the AI era is built on. The question worth asking in the boardroom isn’t “which AI model should we use.” It’s “is our AI agent actually connected to our business or is it just sitting on top of it.”
