All articles
Deploying Agents Into a Business That Was Not Designed for Autonomy Produces Instability at Machine Speed
Anuj Nanda, Co-Founder of Neural Nomad, explains why agent sprawl is an architecture problem rather than a tooling problem and why pilots designed for controlled environments break at production scale
Once you start to look at the overall process and register these agents, it becomes a lot more manageable and a lot more governed. You're not just governing sprawl now. You're governing what the agents are producing as well.
Microsoft's 2026 Work Trend Index puts the number at 19%. That is the share of enterprises considered AI-ready. The other 81% are doing what most organizations do with new technology: inserting it into existing processes and hoping. With autonomous agents, it does not hold. A loan operation built to monitor 100 human-processed applications per day cannot absorb 50,000 agent-processed ones without a fundamentally different architecture for monitoring, escalation, and constraint.
Anuj Nanda, Co-Founder of Neural Nomad and former Director of Data Management and Division CIO at Kroll, spent 16 years at JPMorgan Chase across data governance, program management, and process reengineering. He now focuses on what he calls Agentic Business Reengineering (ABR). The premise is that organizations should not begin with the agent. They should begin with the business process the agent will operate inside: the decisions it may influence, the data it may use, the constraints it must follow, the humans it must escalate to, and the outcomes it is expected to improve.
"Once you start to look at the overall process and register these agents, it becomes a lot more manageable and a lot more governed," Nanda says. "You're not just governing sprawl now. You're governing what the agents are producing as well," says Nanda.
Shadow agents are shadow IT at machine speed
Nanda sees agent sprawl as shadow IT's faster, harder-to-detect successor. “If you ask a CIO how many agents they have, the answer might be 200,” Nanda says. “But the real number could easily be five times that, because business teams can now create agents without going through traditional technology channels.” The risk is not simply that agents exist outside governance. It is that they are producing outputs, triggering actions, and influencing decisions across business processes that were never designed for autonomy.
That is why Nanda connects agent sprawl to Agentic Business Reengineering: companies need to redesign workflows, decision rights, controls, and accountability before agents scale. Neural Nomad's Sovrain Control Plane then becomes the operating layer that registers agents, applies constraints, monitors outputs, and gives leadership visibility into autonomous activity across the enterprise.
Constraint design is the new control model
Nanda draws a distinction between controlling agents and constraining them. Control limits how many exist. Constraint defines what each can do within a decision architecture tied to a business objective.
His example: if a company wants to generate $2 million in new revenue through agent-enabled growth, the work does not sit inside a single agent. It touches sourcing, sales, HR capacity, onboarding, servicing, and financial oversight. Instead of each team building independently, the workflow is designed as a single process with registered agents, defined decision points, and budget constraints enforced across the chain. "You're not just creating agents. You're designing the process for agentic operations," Nanda says. "That's what creates operational coherence at the enterprise level."
Pilots break because they were never built for production
Nanda identifies a failure pattern he has seen repeatedly: pilots designed in controlled environments that collapse at scale. "Those pilots are designed to work at a pilot level, not from a production standpoint," he says. "When companies scale, the pilot breaks because exception cases were not thought through." A two-developer test of Claude for code generation gets scrapped entirely when it does not perform perfectly rather than being redesigned as a production-ready workflow with agents testing, reviewing, and escalating alongside humans.
The downstream bottleneck is equally dangerous. An agent that processes 50,000 loans feeds into a review step still handled by a single human. The backlog grows until the process fails. "You have to design the overall process to say, under this constraint you're good to go, approve them," Nanda says. "Only escalate to a human using a specific business rule. Then the process works the way you intended."
The thesis is consistent throughout: the business process has to be redesigned before agents are introduced into it. "If you deploy agents into a business that wasn't designed for autonomy, you don't get efficiency," Nanda says. "You get instability at machine speed."
