How do enterprises move from agent sprawl to measurable business outcomes?
AI is moving from copilots to systems that execute work. Today's AI agents don't just answer questions. They interact with code, data and enterprise systems — autonomously, across systems, with humans in the loop only when judgment is genuinely required.
Most organizations are already living this reality. They've deployed agents — dozens, in some cases hundreds or thousands. KPMG’s Q1 2026 AI Pulse shows more than half of the organizations are now running AI agents in production. Progress in pilots is real. Enterprise impact is not. The gap has shifted from AI capability to enterprise architecture and control.
Sprawl is the organizational equivalent of water flowing downhill. One team builds an agent for customer support. Another deploys one for invoice processing. A third automates balance sheet reconciliation. Each works in isolation to solve a specific use case. None are built as part of a system.
This is just the latest version of a very old technology problem.
For example, financial institutions could have deployed AI agents running across business units. But they would have challenges in naming half of them, tracing what they decided or explaining why inference costs had doubled in a quarter. Why? There was no shared model for control, governance or visibility. Security teams find access paths nobody approved. Compliance can't trace outcomes. Token spend rises faster than results. Tokenomics and tokenmaxxing emerge. And brownouts become the bottleneck: responses still come, just slower, weaker and less sure.
The problem is structural. Agents treated as features or use cases to implement, not systems to operate.
Most agent strategies assume scale comes from building more agents. That assumption is wrong. Superagents invert that model.
KPMG takes the opposite approach: a small, governed and trusted set of Superagents designed to deliver end-to-end outcomes. A Superagent is an orchestrator. It takes a complex set of instructions and intents and turns it into action — assembling context, invoking reusable agent skills, coordinating across enterprise systems, enforcing rules and managing workflows from request to result.
Consider vendor onboarding, for example. What may have been an eighteen-day compliance chain with sanctions checks, beneficial ownership review, contract history and regulatory exposure, could become a six-hour parallel execution. The compliance officer sees exceptions instead of paperwork.
It operates on a work surface, a shared live environment where agents and humans act on the same outcome simultaneously, rather than passing context, memory and artifacts between disconnected systems.
Another example could be an M&A deal team running diligence across a mid-market acquisition used to spend three weeks assembling a data room picture. A Superagent, with human oversight, can now pull financials, flag covenant risks, cross-reference prior deal patterns and surface the three issues that actually need a partner's attention, before the first internal call.
Many disconnected agents increase complexity. A small number of governed Superagents create leverage through reusable capabilities, consistent governance and scale without rebuilding logic for every use case. They do not expose dozens of agents. They expose one interface to get work done.
Small and governed beats many and disconnected, every time.
The question in production isn't "what can this agent do?" It's "can this system be trusted at scale?" Most organizations stall there. They don’t have an agent control system — a shared model for visibility, identity, permissions, access and traceability that spans every agent in the enterprise, across platforms, built in from day one.
Without it, the cost surfaces in unexpected places. A legal team's unconstrained agent ran overnight summarizing its own prior outputs, saturating the shared inference tier. By morning, the deal team's diligence harness was timing out at quarter close — a brownout triggered by an agent nobody had bounded.
Governance is what makes scale possible. When agents share a control model with clear ownership, consistent access rules and traceable decisions, organizations deploy with confidence. Without that discipline, every new agent adds risk. With it, every new agent compounds value.
Once governance and scale become the goal, most AI environments show their limits. Many platforms make it easy to build agents. Few are designed to operate them as an enterprise system.
At enterprise scale, three requirements become non-negotiable:
Google Cloud addresses this as a platform. Its agent capabilities across the breadth of GCP are built to support how agents are operated, not just how they’re created. A platform alone doesn’t deliver outcomes. Execution does.
Superagents only work if the platform supporting them enforces control at scale. Google Cloud’s powerful, optimized AI stack provides solutions across all three requirements:
Agents operate within the same identity, data and security frameworks as core systems. Superagents sit on that foundation, translating intent into coordinated action.
Between concept and production is where most agent strategies fail. Enterprises don’t struggle with ideas. They struggle to deploy them safely and quickly. Handoffs, long cycles and isolated proof of concepts create distance between design and execution and that distance is where risk accumulates.
Forward Deployed Engineering (FDE) removes it. Teams build and deploy directly in production environments, alongside clients. Architecture, built-in trusted practices, governance and performance are addressed in real time. These teams are agile, responsible and deliver results quickly.
Before FDE, a client's contract review cycle might be generating a proof of concept every six months and restarting, for example. Once FDE embedded a team in their legal operations, they could deploy a governed Superagent in six weeks and have measurable cycle-time reduction before the next steering committee met.
These systems touch data, decisions and workflows. They need clear ownership, defined guardrails and measurable outcomes from day one. Otherwise, it’s a demo factory.
Getting one Superagent into production is a milestone. Scaling them is the test. A pod-based structure brings engineering, data, security and domain expertise together with end-to-end ownership. Each pod builds, deploys and evolves Superagents as a system. The organization stops starting from zero and every deployment makes the next one faster and the governance stronger.
KPMG defines the operating system for enterprise agents – superagents, skills and governance through trusted AI. Google Cloud provides the execution substrate where that system can run securely, observably and at scale.
Start with one workflow. Deploy a governed Superagent. Prove it in production with human oversight. Scale through reuse, not sprawl.
The real measure isn't how many agents are running. It's whether the work itself has changed. Agent sprawl adds lanes to a bottleneck. Superagents redesign the route.
| Enterprise AI Challenge | Superagent Approach |
| Many disconnected agents | Small number of governed Superagents |
| Siloed workflows | End-to-end orchestration |
| Limited visibility | Centralized governance and traceability |
| Duplicate implementation effort | Reusable skills and capabilities |
Complex human handoffs | Shared work surface for humans and agents |
| Scaling operational risk | Scaling through governance and control |
The blog was prepared by KPMG. Authors: Mark Shank, Google Cloud Platform Leader and Swami Chandrasekaran, Global Head of AI & Data Labs.
Definitions
