IBM switched its HR system from Workday to SAP and most employees never noticed, because their HR interface was an AI agent. At COLLIDE 2025, IBM's Melissa Hadley explained systems of intelligence, the problem of agent sprawl and how orchestrator agents route work across systems.
Key takeaways
- Agents are a new layer. Hadley said systems of intelligence sit on top of systems of record, engagement and insight, and act rather than report.
- Three obstacles: agents built in silos, weak integration with core systems, and agent sprawl.
- Orchestrators route the work. An orchestrator agent decides which agent or combination of agents handles each request.
- No-code and pro-code both matter. IBM pairs a catalog and no-code builder with a developer kit and MCP-based tool builder.
- The interface of work is changing. IBM's HR agent hid a Workday-to-SAP switch from most employees, Hadley said.
When IBM switched its HR system from Workday to SAP, most employees didn't notice, according to Melissa Hadley, AI Productivity Sales Leader at IBM. The reason, she said, is that the interface they used for HR was an AI agent. Hadley told the story at COLLIDE Data + AI Conference 2025 in Atlanta, in a session sponsored by IBM, to make a broader point: AI agents are changing the user interface of work.
What are systems of intelligence?
Hadley described a progression in enterprise software. Systems of record, such as CRM, ERP and HR information systems, hold the data. Systems of engagement, such as applications and websites, are how people interact with them. Systems of insight, the analytics and dashboards, tell people what is happening.
AI agents add a fourth category she called systems of intelligence. "These aren't dashboards, they're doers," she said. Agents execute complex workflows across the other systems, so knowledge workers no longer have to become experts in each application's interface. "The complexity of our systems is almost hidden by AI agents," Hadley said.
She cited a forecast that 85% of enterprises would use agentic AI in some way by the end of 2025, and said that inside IBM, agents now handle 94% of HR inquiries, sales team members get a quarter of their work week back, and procurement sourcing and contracting are about 70% faster.
What is holding enterprise AI agents back?
Hadley named three problems. First, agents are often built in silos, so HR, finance and the IT help desk each have their own, while employees work across all of them. Second, agents often aren't well integrated with the core systems where data and daily work live, which limits their return.
Third is what she called agent sprawl. Nearly every application now has an AI assistant in the corner, she said, and moving from experiments to enterprise-grade deployment takes specialized knowledge, time and budget.
What is an orchestrator agent?
Hadley's answer to those problems is an orchestrator agent, which she described as a master supervisor for AI work across the enterprise. It acts as a router and planner: it interprets what the user wants and decides which agent, or which combination of agents, should do the work, regardless of vendor or coding language. Connected to systems of record and insight, she said, it unlocks a new way of working.
IBM's version is watsonx Orchestrate. Hadley said IBM uses it internally to orchestrate AI for more than 200,000 employees in 170 countries, with hundreds of agents working together, connected to core applications including Salesforce and SAP. She said IBM's internal HR agent alone handles more than 11 million interactions a year.
How do non-technical employees build AI agents?
IBM's goal, Hadley said, is for anyone "regardless of technical skill" to build an agent in minutes. She described several pieces of the platform:
- Agent catalog: prebuilt, domain-specific agents for areas like sales, HR and procurement, with governance built in, plus agents from partners
- No-code agent builder: for employees such as HR managers or sales reps who don't write code
- Agent developer kit: a pro-code option for engineers who want to build and test custom agents
- Tool builder: uses the Model Context Protocol to bring existing workflows from systems of record into agents, such as transferring an employee in an HR system or resetting a password
- AgentOps: traceability, transparency and failure alerts so teams can review what an agent did
In a recorded demo, Hadley copied IBM's sales prospecting agent from the catalog, added a workflow for scoring and prioritizing leads, and attached a prebuilt CRM agent from the Salesforce catalog. The model and knowledge sources could be swapped, she said, including bringing your own documents for retrieval-augmented generation.
The demo then switched topics mid-conversation. The seller asked about retail accounts, then about booking vacation time, then about a broken laptop, and the orchestrator routed each request to the right sales, HR or IT agent. The same conversation then moved into Slack. On tracing, Hadley argued that the black-box problem many teams hit with generative AI "didn't quite get solved for agentic AI until now."
Why does multi-agent task execution matter?
Hadley distinguished multi-agent task execution from traditional task automation such as robotic process automation. Orchestrated agents can carry out processes of "seven, eight, nine and more" steps across multiple systems on a user's behalf, she said. That is why integration with core systems matters: she called it a recipe for return on investment.
IBM's view, Hadley said, is that agents, orchestrator agents and the systems of record and insight that "aren't going away" together change how people work. The Workday-to-SAP switch was her example. She said that even she, as an AI strategist at the company, wasn't aware of the change. "This is what I mean by changing the user interface of work," she said.
See enterprise AI in practice 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 keynotes, panels and demos like this one. The 2027 conference takes place Wednesday, October 13, 2027, at the Sandy Springs Performing Arts Center. Details are on the COLLIDE event page, and Data Science Connect's free webinars, including its Agentic AI Showcases, cover agent platforms throughout the year.
Frequently asked questions
What are systems of intelligence?
Hadley's term for AI agents that act on insights and execute workflows across systems of record, engagement and insight, rather than only reporting.
What is an orchestrator agent?
A supervising agent that interprets a user's request and routes it to the right agent or combination of agents, regardless of vendor or coding language, according to Hadley.
What is agent sprawl?
The proliferation of separate AI assistants and agents across applications and departments, which Hadley said makes cross-functional work and enterprise deployment harder.
How is multi-agent task execution different from RPA?
Hadley said orchestrated agents can carry out processes of seven or more steps across multiple systems, rather than automating single tasks.
Watch: Usher in the Future of Work With AI Agents Built for the Enterprise, brought to you by IBM on YouTube
Event: COLLIDE AI Conference · More coverage: COLLIDE AI Conference insights
