Anand Naidu is a seasoned development expert with a deep mastery of both frontend and backend architectures, specializing in the intricate coding languages that power modern enterprise systems. With years of experience navigating the friction between legacy ERP software and cutting-edge automation, he provides a pragmatic perspective on how businesses can bridge the gap between rigid data silos and the fluid demands of the modern marketplace. His insights are particularly vital today, as organizations move beyond basic digitization toward the complex, often misunderstood world of agentic AI and autonomous reasoning.
The following discussion explores the hidden vulnerabilities of current enterprise systems, the technical hurdles stalling AI adoption, and the necessity of building “memory” into digital workflows. We delve into why 70% of transactions often require manual intervention and how leadership can move past the high failure rates of current AI pilots to achieve genuine bottom-line impact.
ERP systems are often built for optimal paths, yet a vast majority of transactions require manual overrides or patches. How do these “shadow workflows” create hidden financial risks for an organization, and what specific steps can leadership take to identify these invisible processes before starting an AI implementation?
In my experience, ERP systems are notoriously optimistic; they are designed for the 30% of transactions that follow a perfect, “happy path,” leaving the other 70% to be handled by human intervention. When employees use spreadsheets, group chats, or email chains to patch these gaps, they create shadow workflows that are completely invisible to leadership and unaudited by the system. This creates a massive financial risk because a single corner case that slips through the cracks can lead to payments posted to the wrong ledger, shipments that bypass crucial compliance controls, or inventory records that diverge from physical reality. To fix this, leadership must conduct a deep operational audit to uncover where the “real” operating system of the company lives—usually in the manual workarounds—before layering AI on top of a broken blueprint. It is about acknowledging that these workarounds exist because the ERP failed to adapt to changes, like new payroll rules or equipment loan programs, that occurred long after the initial rollout.
Agentic AI relies on documented workflows, but many businesses operate through informal group chats and manual spreadsheets. In what ways does automating only the visible processes increase operational complexity, and can you share an example of a “corner case” that usually breaks these autonomous systems?
Automating only the visible, documented processes is like paving a road that leads directly into a swamp; the AI will execute its task perfectly until it hits the undocumented reality, at which point it creates a new layer of complexity rather than solving the old one. If an agentic AI automates a procurement chain but doesn’t see the informal approval process happening in a WhatsApp group, you end up with two parallel systems that eventually clash and corrupt your data. A classic “corner case” that breaks these systems is an invoice that arrives in an unexpected format or a tax regulation that shifts mid-quarter. While a human clerk might instinctively know how to handle a vendor discrepancy based on a ten-year relationship, a standard agent without that context will either stall or process the data incorrectly, forcing a manual cleanup that is often more expensive than the original task.
Most enterprise AI deployments lack persistent memory, meaning the reasoning used to solve a problem dissolves once the task is finished. Why is this “amnesia” particularly damaging for long-term institutional knowledge, and how can a system be designed to retain context across different quarters?
This structural “amnesia” is perhaps the most frustrating barrier we face because it prevents AI from ever gaining the “seniority” we value in human employees. When an agent reconciles supplier payments for Q3 but forgets the specific reasoning and judgment calls it made by Q4, the organization loses the ability to compound intelligence over time. Without persistent memory, the agent might resolve a recurring discrepancy today, only to have to “re-reason” the same problem from scratch three months later, which leads to inconsistent outcomes. To bridge this gap, we need architectures that prioritize contextual memory, moving away from stateless execution toward systems that can store and retrieve the nuances of previous interactions. We have to move past simple summaries—which often confuse the agent with conflated problems—and build a modular framework where the AI can actually learn the “quirks” of a specific business.
Billions are being spent on generative AI, yet only a tiny fraction of companies currently achieve full production or measurable bottom-line impact. What technical hurdles consistently stall these projects at the pilot stage, and what metrics should executives prioritize to ensure AI actually improves the P&L?
It is a sobering reality that while $30–40 billion has been poured into generative AI, a staggering 95% of companies see no measurable impact on their bottom line, according to recent studies. Projects consistently stall at the pilot stage because the workflows are too brittle; they work in a controlled environment but shatter when they hit the thousands of “corner cases” found in real-world ERP data. Executives often make the mistake of tracking individual productivity gains, which don’t always translate to P&L performance, rather than measuring end-to-end process completion rates and error reduction in complex transactions. To move the needle, we must focus on the 5% of projects that reach full production by prioritizing modular augmentation—updating specific, rigid parts of the ERP rather than trying to overhaul the entire system with a single, unproven agent.
There is a growing need for AI architectures that prioritize contextual memory and learning over simple autonomous execution. How would an “agent alternative” differ from current models in its day-to-day reasoning, and what role does modular augmentation play in updating a rigid, decades-old ERP system?
An “agent alternative” differs by being “contextually aware”; it doesn’t just execute a command, it remembers the environment and the history of the data it is touching. Unlike standard agents that operate in a vacuum, this alternative builds an evolving understanding of the business, much like a seasoned employee who knows exactly why a certain vendor always submits late invoices. Modular augmentation is the secret weapon here because it allows us to plug these intelligent, memory-capable frameworks into decades-old, rigid ERP systems without requiring a total “rip and replace.” This approach generates rapid development of new functionality that can meet dynamically changing business processes, essentially giving an old system new “brains” that don’t forget what they learned yesterday.
What is your forecast for the future of AI-driven ERP transformation?
I forecast that the “hype phase” of autonomous agents will soon give way to a more disciplined era of “contextual transformation” where memory is the primary metric of success. We will stop seeing AI as a series of disconnected chatbots and start viewing it as a persistent layer of institutional knowledge that sits on top of our systems of record. By 2026, the companies that thrive won’t be those with the fastest agents, but those that have successfully digitized their “shadow workflows” and integrated them into a modular AI framework that learns from every transaction. The shift from “stateless” automation to “stateful” intelligence will finally allow ERPs to move beyond being simple silos and become the truly dynamic operating systems they were always meant to be.
