Skip to article

What Does an AI Agent Cost After It Launches?

A business case that stops at the build leaves out what comes next. What three leaders said about model changes, maintenance, token use, evaluations and runtime policy, and the line BNEDai adds for each.

By Jacqueline V. TwillieNovember 20, 20264 min readGuide
Quick answer

After launch, an AI agent costs money in four places: the next model move, the people who keep it running, token use per workflow, and evaluations. Cardinal Health's Alyssa Cox and Bloomberg's Shawn Edwards described them at an online enterprise AI summit on November 12, 2026.

A business case that stops at the build leaves out what follows it. Three leaders at an online enterprise AI summit on November 12, 2026 spoke about what comes after launch. What each one said is reported below, and BNEDai adds its own recommendation after each part.

The next model move belongs in the business case

Alyssa Cox, VP of Transformation at Cardinal Health, said that when the platform underneath changes its infrastructure or moves from one model to the next, the vendor does not fund Cardinal's side of the transition. Cardinal has to do it. Her team estimates what that ongoing support will cost and puts it in the business case.

BNEDai's recommendation: put a line for the next model move in the business case before anyone approves the build, with a name next to it.

The people who keep the AI agent running need a line of their own

Cox said Cardinal needs people on the ground maintaining and managing these solutions, and making enhancements as processes change. Naming what it costs to maintain them, she said, lets Cardinal show that those people deliver value. That moves the IT staffing conversation from keeping the lights on to value delivery. She added that returns on implementation projects fund discovery work, and discovery never funds itself.

BNEDai would give the people who will maintain the AI agent once it is live a line in the business case, with a cost beside each.

Token use needs tracking for each workflow

Cox said business cases for AI rest on the same fundamentals as any other. What differs is understanding the ongoing cost. Old business cases tracked headcount reductions, she said, and bodies are easy to understand. The ongoing cost was whatever the software fees were. Tracking token use for a particular solution takes observability across your systems, down to which query a workflow is making. That means a different kind of collaboration with IT, she said, to collect and understand the data, than the one you have with finance.

BNEDai advises getting IT and finance in one room before launch to agree on how each workflow's token use is tracked, who reads the report, and what happens when a number looks off.

Bloomberg spends heavily on evaluations

Shawn Edwards, Chief Technology Officer at Bloomberg, called evaluations the cornerstone of building a trustworthy system, or building AI, period. Bloomberg has put a huge amount of time, effort and money into them, he said. Bloomberg takes a model agnostic approach, constantly updates its models, and uses different models for different AI agents. The evaluation framework, he said, is incredibly necessary to sustain that.

BNEDai's view: budget evaluations as a recurring cost and rerun them whenever someone swaps a model.

Policy has to be enforced while the AI agent runs

Jenn Kosar, PwC's US AI Assurance Leader, said most large organizations already have principles and policies, and the challenge executives describe is turning them into controls that work at scale. The controls she typically sees are concentrated before deployment. There is work to do on enforcing policy at runtime, she said, and there is still a lot of reliance on humans in the governance process and in the execution. Her aim is to move past relying mainly on humans to get it right, and past policies that sit on the shelf instead of being enforced day to day.

Her answer on modernizing governance names which data platform the AI agents have authority to reach, and how identity and access management is controlled. It also names defined decision rights, how you test and monitor, and how you escalate and have humans step in.

BNEDai suggests picking one live AI agent and checking it against that answer: what it can reach, who can approve what it does, and when a person steps in. Then repeat with the next one. The page on autonomy against consequence covers how to set the approval step for each action.

Frequently asked questions

Who said the vendor does not fund the transition to a new model?

Alyssa Cox, VP of Transformation at Cardinal Health. Her team estimates the ongoing support cost and puts it in the business case.

What does tracking token use require?

According to Cox, observability across your systems, down to which query a workflow is making, and a different kind of collaboration with IT than the one you have with finance.

Why does Bloomberg invest in evaluations?

Shawn Edwards called evaluations the cornerstone of building a trustworthy system. Bloomberg uses different models for different AI agents and constantly updates them, and the evaluation framework is necessary to sustain that, he said.

Source: remarks by Alyssa Cox, VP of Transformation, Cardinal Health; Shawn Edwards, Chief Technology Officer, Bloomberg; and Jenn Kosar, PwC's US AI Assurance Leader, at an online enterprise AI summit, November 12, 2026.

Reviewed on November 20, 2026.

More on Leadership for Corporate Executives