The Hidden Cost Of AI Agents And Why Enterprise Architectures Are Cracking

In an effort to secure a future rooted in AI, many enterprises are inadvertently rewinding back to the past. Before cloud transformation and SaaS eclipsed on-premise software and infrastructure, company data was often siloed by default because interconnectivity between departmental functions was either too cumbersome, or impossible.

Once cloud tools became dominant, however, point solutions by necessity morphed into point solutions imposed by vendor lock-in and inflexibility by design. In the pursuit of fast utility, many businesses set themselves up to incur increased costs down the road, with few options to roll back the tech sprawl that had become so ingrained.

For all of the infinite potential agentic AI has to offer, the current rush to deploy intelligent agents risks the very same information silos that created a generation of technical debt and enterprise-wide fragmentation. Enterprises are unknowingly walking into the same fragmentation issues that plagued the early SaaS era: disconnected systems, compounding technical debt, and compliance blind spots.

This warning comes from Manoj Mohan who joined CIO News for an interview to discuss the risks of siloed agent sprawl in the enterprise. Head of Engineering for Data Platforms and Insights at Intuit, Manoj has over two decades of experience on the front lines of automation building Enterprise Data & ML Platforms at companies like Apple, Meta, and Cloudera.

• Debt by a thousand agents: "In the early days of the cloud, speed felt limitless—until cost and complexity caught up," said Manoj. "We’re seeing the same playbook replayed with AI agents. You gain incredible velocity at first. But within 12–24 months, the architecture becomes brittle, integration gets expensive, and the compound interest of poor design hits hard." Without a unifying blueprint, what begins as rapid experimentation becomes architectural entropy. Agent silos proliferate, tech debt compounds, and operational resilience erodes, often silently, until it’s too late.

• Out of touch, out of compliance: "Imagine a customer requests ‘do not contact’ under GDPR or CCPA, but that data never reaches a siloed AI sales agent. That’s not just a technical oversight, it’s a compliance failure, a reputational risk, and a legal vulnerability in one stroke," explained Manoj. This is the hidden danger of unchecked agent sprawl: not just operational inefficiency, but systemic governance blind spots. The more agents you deploy without orchestration and context-sharing, the more fragmented your compliance posture becomes.

Manoj proposes a proactive architectural model anchored by a strategically empowered Enterprise Architecture function, tasked with designing and enforcing a unified AI blueprint that scales across the organization. The three-layer solve prevents agent chaos by first cataloging all agents in a shared registry for visibility, then standardizing their communication through a common protocol, and finally controlling their actions with automated governance-as-code. But the architectural blueprint cannot exist in hypothetical isolation; it requires reusable technical foundations for agents to interoperate.

• The common tongue: To avoid agent fragmentation, emerging standards like the Model Context Protocol (MCP) and Agent-to-Agent aim to establish a shared language for inter-agent communication, favoring flexible, event-based payloads over rigid APIs for adaptability and scale. "The great advantage of agentic workflows is their ability to loosely couple and exchange information across systems," he noted. "You don't want to tie that to a hard-coupled API. In the future, I see more systems moving towards flexible, semi-structured event payloads, which will allow them to evolve and adapt more easily."

• Build once: Manoj drew a parallel to workflows he saw firsthand at Intuit, where business units like QuickBooks and Credit Karma all needed common functions like payment processing. Instead of each building their own version, they built a single, configurable platform with fewer engineers. "That’s the model for AI as well," he said. "While the business context of each agent is unique, the technical implementation, the high-level design, and the architectural flows must be consistent and seamless." This efficiency reduced developer workload to the point of a task that previously required up to 25 engineers could now be accomplished with 10.

Even the most well-intentioned AI strategies eventually face a reckoning: internal plans must align with customer reality. Manoj framed the challenge as a dual mandate, serving both the developer persona and the customer persona.

• Productivity's two faces: For internal teams, agentic AI offers supercharged velocity. Developers can ship features faster, reduce toil, and iterate in real-time. But the real transformation comes from serving the "customer persona" with that newfound engineering productivity. In one use case, Manoj's team used agentic workflows to automate complex manual tasks that customers were previously forced to handle themselves. "If we can build automation for something like financial transaction reconciliation that was a manual burden, it will instantly make the customer happy." When groundwork is laid in a repeatable fashion, engineering teams benefit from efficiency while customer benefit from faster features being shipped. “AI done right isn’t just about building faster,” Manoj said. “It’s about building better at scale, deriving compelling value for customers.”

But he also offered a sobering prediction: most enterprises will stumble before they scale. “AI is evolving so fast, even seasoned technologists are struggling to keep pace. If engineering leaders are feeling the heat, how can we expect more traditional organizations to get it right on the first iteration?”

The pressure is immense, especially as prominent leaders like Salesforce’s Marc Benioff make bold claims about massive AI-driven productivity gains and human worker replacement. Manoj predicted many companies defaulting to vanity metrics to buy time. “They’ll chase quick wins and only address agent sprawl and architecture gaps after cracks start to show.”

It should be clear by now that AI for the sake of AI is not a viable strategy, but the obvious is not always obvious when complex externalities pressure businesses to act less than rationally. "The pragmatic answer, unfortunately, is many leaders are still trying to find how to prove AI as a core part of their strategy. But AI is not just part of your strategy. If you bring that lens, you're never going to get the best value. It has to be holistic," Manoj said. Proactively planning the chain of ownership, managing stakeholder communication, and getting the fundamentals of orchestration right from the outset is a recipe for minimal risk and maximum probability of success.