Agents in the Enterprise: Who Owns the Outcome?

When agents take over tasks, people do not disappear from the process. They move to checking the outcomes, and sometimes more of them are needed. That was one of the clearest lessons from the keynote panel at the ALIGN AI Executive Summit Chicago 2026 on July 15, Agents in the Enterprise: Hype, Reality, and Risk, brought to you by HPE and moderated by Michael Emerick of HPE. Panelists were Enrique Quevedo of University of Chicago Medicine, Garrett Woody, Enterprise Architect at the Federal Home Loan Bank of Chicago, Russ Felker, CTO of MegaCorp Logistics, and Kelly O'Brien, VP of Product Strategy for Commercial Card at JPMorganChase.

Where are agents working today?

Felker's company names its agents. Its most-used one, Maggie, is an orchestrator that contacts carriers about loads in transit to check status and location, and carriers now interact more with AI agents than with people for those routine exchanges. After one rollout, a team asked to turn the agent off. Nothing had gone wrong, Felker said; they just didn't know what to have their intern do anymore.

Woody said his bank is further along with single agents than with fully agentic workflows. One agent reviews municipal bonds, which the bank accepts as collateral and which arrive in many formats, and recommends whether each is related to real estate. For a handful of bonds the time savings are small, he said, but for a hundred at once they are large, and a person still checks every recommendation.

Quevedo's analytics team bridges business intelligence and hospital administration. He compared today's moment to data democratization 15 years ago: non-technical colleagues now produce plausible-looking AI results, and his team spends time validating them because real decisions depend on them.

How do you decide what's reusable?

O'Brien said the key is the long view: building composable pieces that can be reused across use cases, such as card-present and card-not-present payment workflows, so teams don't start from zero. Felker took a different angle. His company has shifted from prediction to experimentation and now treats software as disposable. "The data is what becomes reusable," he said. Woody added that models should be swappable too, since regulators may ask where a specific provider's models are used and whether they could be replaced.

How do you govern while moving fast?

Quevedo embeds data scientists in pods aligned with clinical and research departments, so each model is built with the leaders who understand that line of business. Woody starts with lower-risk use cases, such as public data and agents that make "recommendations, not decisions," to get through enterprise risk reviews faster.

Felker said governance has to be decomposed to the task level. Each agent should be treated as a non-human identity with one task and the minimum data it needs, applying least privilege and zero trust. "It's the agent taking over tasks that used to make up a job," he said.

O'Brien said the line between recommendation and decision makes auditability essential, because the human who signs off must be able to explain the decision to a customer or regulator. She added that risk and compliance partners will need deep AI understanding.

What about memory and drift?

Woody described agents that hit a database error, see no data and carry on as if nothing were wrong, which is why his team is experimenting with orchestration-layer memory across task-specific agents. Felker said bad outcomes left in an agent's memory get treated as good, so his team measures drift in outcomes and removes problem cases. The work is done by people reviewing outputs. One 24/7 department grew about 10 percent because agents now handle every transaction rather than a sample, creating more outcomes to validate. He counts that as service lift rather than direct ROI.

Who owns the outcome?

Quevedo's team runs AI models in parallel with ones they build themselves and won't rely on them yet. "Trust is gained in drops and lost in buckets," he said. Woody wants to trace every decision in a workflow back to its source data. Felker's rule is simple: "A human owns every outcome." O'Brien separated two questions, whether an agent is working and why, and said answering both is what lets compliance teams accept the risk.

Where to hear more in 2027

Enterprise leaders will keep comparing notes on agents at the 2027 ALIGN AI Executive Summits, invitation-only gatherings in eight U.S. cities: Atlanta, New York City, Dallas, Chicago, San Francisco, Boston, DC and Los Angeles. Registration is coming soon. See the ALIGN AI Executive Summits page.

Frequently asked questions

How should enterprises govern AI agents?

Panelists recommended decomposing work into tasks, giving each agent minimal data access, and making decisions auditable.

Do AI agents reduce headcount?

Not always. One panelist said a department grew about 10 percent to validate the larger volume of agent outcomes.

What is agent drift?

Outcomes moving outside an acceptable range over time, often because bad results remain in an agent's memory.

This article is based on remarks by Michael Emerick, HPE (moderator); Enrique Quevedo, University of Chicago Medicine; Garrett Woody, Enterprise Architect, Federal Home Loan Bank of Chicago; Russ Felker, CTO, MegaCorp Logistics; Kelly O'Brien, VP, Product Strategy, Commercial Card, JPMorganChase at ALIGN AI Executive Summit Chicago 2026 on July 15, 2026. More coverage: ALIGN AI Executive Summits.

Filed under: Agentic AI, AI Governance & Trust, People, Culture & Leadership