Building an AI agent is easy; running thousands of them is not. At COLLIDE 2026, IBM's Ann Leach explained how agent sprawl creates security and cost risk, the five questions an agentic control plane must answer, and how Starlink coordinated procurement agents.
Key takeaways
- Sprawl is the new problem. Leach cited a study in which 94% of enterprises said agent sprawl raises security risk, but only 12% have a centralized platform to manage it.
- Building is the easy part. Most effort goes into integration, testing, identity, observability, policy and cost.
- A control plane answers five questions: what an agent can do, who runs it, how it behaves, what it costs and where it can operate.
- Control must span vendors. If it only works in one ecosystem, Leach said, it's a console, not a control plane.
- Governance belongs in the life cycle, from design through monitoring, not as a final gate.
The first phase of enterprise AI asked whether a team could build an agent that does something useful. The harder question now is how to run hundreds or thousands of them "safely, reliably, and economically," Ann Leach, Executive Director, Data and AI Product Marketing at IBM, told attendees at COLLIDE Data + AI Conference 2026 in Atlanta.
In her session, "From AI Agents to AI Operations: The Missing Layer for Enterprise Scale," Leach argued that the missing piece is an agentic control plane, a software layer that governs every agent a company runs.
What is agent sprawl?
Leach described a pattern she said is happening now. One business unit adopts an assistant. Another team builds an agent on a different platform. A third connects an agent to a SaaS application. Each project makes sense on its own, she said, "but at the enterprise level, the result is sprawl."
Agents built by different teams end up with different access models, policies, tools and monitoring. Often no one can answer basic questions, according to Leach: what agents exist, who owns them, what they can access, and what happens when one behaves unexpectedly.
She cited an OutSystems study in which 94% of enterprises said agent sprawl is increasing security risk and complexity, while only 12% have implemented a centralized platform to manage it. "Enterprise AI is moving from experimentation challenge to an operations challenge," Leach said.
Why is building an AI agent only part of the work?
Building an agent is easy, Leach said. She has built several herself. Most of the effort lies below the surface: connecting to back-end systems, testing and evaluation, identity and access, observability, runtime policy, cost tracking and coordination with other agents. All of that has to keep improving after deployment.
Early market attention focused on how fast teams could build agents, she said. The enterprise question is how to manage the full life cycle and operate the agent as part of the business.
What is an agentic control plane?
Leach defined it as the layer that answers five questions:
- What can the agent do? Policy guardrails on which models, tools, data and actions are allowed.
- Who can run it, and on whose behalf? Identity, access and delegation.
- How is it behaving? Observability and tracing that create an audit trail.
- What does it cost? Consumption by agent, model, use case or line of business. Leach said cost came up in many of her conversations at the conference.
- Where can it operate? Across clouds, vendors, frameworks and agent platforms.
On the last point, she drew a line between real control and vendor lock-in. "If the control works only inside one ecosystem, it's not really a control plane," Leach said. "It's a console."
She added that governance belongs throughout the life cycle, from design through build, testing, deployment and monitoring. "Governance is not a gate that is added at the end," she said.
How does IBM approach agent operations?
Leach presented IBM's watsonx Orchestrate as its control plane, calling this part of the talk her "IBM product pitch." She said agents from different platforms can be registered in it so they become discoverable and auditable, without forcing every agent onto the same runtime. The design principle, she said, is to "meet the agent estate where it is."
She walked through three requirements using product screenshots:
- Visibility: a common view of the agent estate. Business owners can ask questions conversationally instead of navigating dashboards.
- Economic visibility: showing which agents and models drive token consumption, so teams can ask whether a cost supports a high-value use case or whether a different model should be used.
- Consistent control: enterprise-wide rules, including controls for sensitive data such as PII, applied across agents rather than one at a time.
How did Starlink use AI agents in procurement?
Leach described Starlink's use of IBM Bob, the company's coding assistant, to build and test procurement agents, and Orchestrate to coordinate them. The agents connect to systems Starlink already uses, including Oracle NetSuite, DocuSign and Automation Anywhere, for ERP validation, document processing and quote handling.
"Starlink didn't create a different island for AI," she said. Leach said purchase order and quote comparison is projected to drop from about 30 minutes to 3 minutes, saving an estimated 11,500 hours of procurement effort a year. She stressed that both figures are projections.
Her advice was to start with one meaningful use case and check it early against operating questions: how the agent will connect, how it will be governed and observed, and who owns it. "Building agents is becoming easier," she said. "Operating agents across a complex enterprise is not."
Agentic AI at COLLIDE 2027
COLLIDE, the largest AI conference in the Southern U.S., brings more than 1,000 enterprise data and AI leaders to Atlanta for sessions like this one. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Learn more about COLLIDE, or catch Data Science Connect's Agentic AI Showcases among its free 2027 webinars.
Frequently asked questions
What is AI agent sprawl?
The result of teams building agents on different platforms with different access models, policies and monitoring, so no one has a consistent view of what agents exist or who owns them.
What is an agentic control plane?
A software layer that governs an enterprise's agents by setting policy guardrails, managing identity and access, tracing behavior, tracking cost, and working across clouds and vendors.
Why does AI agent cost management matter?
As agents multiply, model and token consumption spread across teams. Without a common view, Leach said, cost is hard to attribute and harder to optimize.
Watch: From AI Agents to AI Operations: The Missing Layer for Enterprise Scale on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
