Who Orchestrates The Orchestrators? Enterprise AI Scale Requires Agentic Operating Models

Agentic AI’s scaling problem starts with a question regulated enterprises can’t avoid: Who orchestrates the orchestrators? As banks and fintechs move from pilots to production, hundreds of agents can’t simply be unleashed across fragmented platforms, data environments, and business units. They need an operating layer that coordinates reuse, enforces guardrails, and keeps autonomous work tied to enterprise value. Without it, agentic AI creates motion. With it, governance becomes the architecture that makes scale possible.

Vikas Krishan is Chief Digital Business Officer and Head of UK and EMEA at Altimetrik, an AI-first digital engineering firm of 11,000-plus employees serving regulated enterprises. With two decades of leadership across financial services, consulting, and global outsourcing, Krishan has watched governance move from a back-office compliance concern into an operating architecture question that determines whether AI investment produces enterprise value.

"Everyone wants to deploy hundreds of agents, but that has exposed how unprepared some organizations are to execute across fragmented technology stacks," Krishan said. "If you only address the technology stack, you’ll never get the efficiency you’re looking for, and it’ll never work the way you want." For Krishan, the harder bottleneck sits underneath the agents themselves. Pilots stall when institutions try to industrialize without first fixing the stack and the org chart around them.

• Governance as velocity, not friction: Krishan pushed back on the framing that compliance slows AI rollout. "Governance isn't a hindrance. What AI governance means is it gives you guardrails, security, and validation around your model, and it also avoids duplication," he said. Enterprises do not need two agents performing the same job. Without orchestration discipline, the same agent should execute across multiple processes and business units. Governance becomes the framework that makes that reuse possible at scale.

• Where agent pilots break: The breakdown surfaces when institutions try to scale agents across functions. Krishan pointed to the finance function. Business unit controllers in regulated firms typically run their own agents on their own platforms with their own goals. A mature model centralizes that work, with shared agents executing across business units and orchestration agents coordinating above them. "All of the various levers around process and around people and cultural alignment haven't happened yet," he said. That gap, more than any technology limitation, slows enterprise-scale agentic deployment across banking operations.

• The middle management adoption gap: Krishan described a pattern that enterprise leaders are starting to recognize. Junior staff adopt AI as natives. Senior leaders approach it as veterans of prior technology cycles. The middle layer struggles. "It's the middle levels that are finding it difficult to adopt," he said. Their work spans people management and individual execution in ways that do not map cleanly onto agent-assisted workflows. Reorganization, rather than retraining, is the work most enterprise leaders have not yet started.

Underneath the adoption question sits a deeper architectural argument. Enterprises have not redesigned work itself around AI. They treat agents as faster versions of existing tasks, which caps the value those agents can create across data architecture and decision authority.

• Humans at the center: Krishan reframed the workforce model dominating most agentic conversations. "At the moment, we're using people as a guardian function. What we are going to start seeing when we reorganize is having people at the center and AI doing that work," he said. Engineering teams would not need separate QA and developer roles. Finance teams would not need siloed controllers. The human role shifts from execution to orchestration, with agents handling the work around them and governance agents enforcing the guardrails.

• The agent marketplace: Krishan described a near-term model that resembled a corporate app store. "In the future, individuals will be able to pull down agents from that store, which will then be able to perform jobs and tasks," he said. Coordinating agents and orchestration agents would sit above that layer, applying outcomes from one process into insights in another. Regulation would shape the architecture from the start. The EU AI Act has already established explainability requirements pushing financial services firms toward tighter model controls, while vendors are beginning to extend agentic platforms into corporate banking workflows under those same constraints.

The harder admission inside regulated enterprises is that AI exposes years of deferred architecture work. Krishan recalled when bank CIOs spent four-year tenures building monument platforms. That era has ended. "The CIO role is becoming much more nuanced and difficult. That role is actually beginning to modernize and develop and build an architectural vision that will enable that AI future for an organization," he said.

Modernization, governance of digital coworkers, and operating model redesign now run in parallel. Agent counts keep climbing. The institutions that turn that volume into enterprise value rather than activity will be the ones that orchestrate the orchestrators before the orchestrators outpace them.