Now consider the domains where agents should, in theory, be simplest to deploy. Marketing. Content generation. Internal operations. The regulatory exposure is minimal, the liability diffuse, and nobody faces consequences if a draft email is slightly off.
And yet the value stays thin.
McKinsey's 2025 survey found 62% of organizations experimenting with AI agents. But no more than 10% were scaling agents in any individual function. Only 39% attributed any EBIT impact to AI at all, and most of those said less than 5%. The tools are everywhere. The organizational return is hard to find.
The technology explanation is tempting. The structural one is more convincing: these domains never had to build delegation infrastructure, so they don't have it. There's no equivalent of a payment token, no chargeback process, no regulatory mandate to log what happened and why. Agents operate in a structural vacuum. They can produce things, but nobody has formalized what they're authorized to produce, how their work gets verified, or what happens when the output is wrong.
What fills that vacuum is a dynamic HBR researchers labeled "workslop." In a survey of 1,150 desk workers, 40% said they'd encountered low-substance AI-generated work in the prior month. Each instance consumed an average of nearly two hours to deal with. More than half of recipients felt annoyed. About half viewed the sender as less capable. AI makes it cheap to produce output, so people produce more of it, and the cognitive labor of evaluating quality shifts downstream to whoever receives it.
The agent, or the person wielding one, has no clear authority boundary, no quality contract, no accountability mechanism. Work gets done fast and evaluated slowly, if at all.
McKinsey's data on high performers sharpens the point. The roughly 6% of organizations seeing meaningful EBIT impact from AI were roughly three times as likely to have redesigned how individuals actually work. They were also more likely to define processes for when model outputs need human validation. They built delegation infrastructure voluntarily. Most organizations haven't, because nothing forced them to.
In payments, confused delegation creates a disputed charge, a fraud investigation, a regulatory finding. The cost is immediate and legible. In marketing or internal ops, confused delegation creates a mediocre deck, a slightly wrong summary, a meeting that could have been an email. The cost is real but diffuse, showing up as organizational drag rather than an incident with a case number.
So agents in these domains stay in demo-mode. Useful for first drafts, risky to trust with anything consequential, and perpetually hovering somewhere between fast intern and accountable participant without quite becoming either.