McKinsey recently published a conversation with Kate Smaje, global leader for technology and AI at the firm and co-author of Rewired, alongside senior partners Brooke Weddle and Bryan Hancock. In it, Smaje asks a question most organisations have never seriously addressed: who actually evaluates the performance of your AI agents? Companies know exactly how to hire, develop, assess and offboard human staff. About the performance of what Smaje calls "nonhuman labor", there has rarely been an internal conversation.
Agents need a people strategy too
The conversation starts from a simple observation. AI agents, like people, need proper management. Engineers to correct them when they drift, compute as their energy source to keep functioning, and clear lifecycle management, because an agent cannot exist indefinitely without maintenance or oversight. Smaje has a name for what happens without that management: "abandonware". Forgotten, outdated agents that keep wandering somewhere in the corporate network, with nobody quite sure what they do or why they still exist.
That problem, according to Smaje, is not a technology issue but an organisational one. Many leaders measure their talent transformation by how much new external talent they bring in, when the real gain lies in how they grow their existing experts, the people who already know the business inside out. Smaje quotes a client who put it sharply: "It is far easier for me to teach my metallurgists AI than for me to teach my AI specialists metallurgy."
The question that actually matters: who owns it?
The most concrete statement in the interview concerns ownership. Smaje is explicit: it is not the CTO or the IT department that should manage an organisation's entire population of agents. It is the business owners who deploy the agents within their own workflows. Her example: the head of finance who today manages the people running reconciliation and the month end close should tomorrow also manage the agents working within that same process. "I don't think this is a technology problem," she says. "I don't think the CTO suddenly needs to manage half the organization's workforce."
That is not a licence for unchecked sprawl. There remains a need for shared guardrails, standards and lifecycle management across the whole organisation, otherwise every process ends up with its own abandonware. But day to day responsibility belongs with whoever already knows the process, not centrally with IT.
For organisations, that tension is a familiar one. As soon as AI agents move from experiment to production, the instinct is to place them centrally under IT. That makes sense from a control perspective, but it creates an ownership gap. IT manages systems it does not understand at a domain level. The process owner does sense when an agent gets something wrong, but has no lever to correct it.
What does this mean in practice for how you structure and follow up your agents?
-
Ownership belongs with the process, not with IT. Whoever manages the people in a workflow today, whether in finance, customer service or procurement, should manage the agents in that workflow tomorrow. IT facilitates the infrastructure and the guardrails, but day to day follow up belongs with the process owner. Whoever manages the people in a process also manages the agents in that process.
-
Teach your metallurgists AI, not the other way round. The domain knowledge needed to properly judge an agent already sits in house, with the people who have run the process for years. It is more efficient to make them AI capable than to teach outside AI specialists the nuances of your sector. The expertise is already there, the AI skill gets added to it.
-
Shared guardrails prevent abandonware. Distributed ownership does not mean every department invents its own rules. Standards for safety, data handling and lifecycle management need to be agreed centrally, so a forgotten agent never keeps wandering the network unnoticed. Distributed ownership only works with a shared framework underneath it.
-
Human judgement becomes the scarce resource. Smaje is clear on this: "The rise of AI doesn't mean we all suddenly delegate our responsibilities for quality control and critical thinking. That's not the world we live in. There's a premium on those capabilities now." Whoever hands quality control entirely to the agent delegates exactly the work that requires the most judgement. The more an agent takes over, the more valuable the judgement of whoever still stands above it becomes.
The real lesson
At Rescope we have long called this ownership with the client: Rescope guides, the client leads. What Smaje describes is that same logic, carried through to agents. Not Rescope or an IT department manages your entire agent population, but your own Champions and their team leads: the people who already know the process. They learn to follow up their agents, correct them and phase them out where needed. Rescope's role is to help set up the guardrails and transfer the knowledge, not to hold on to the controls.
The question, then, is no longer whether your organisation uses AI agents. Most already do, in more processes than IT suspects. The question is who is responsible today for the performance of those agents, and whether that person actually knows the process they work in.
Who manages the agents in your organisation?