Companies that really make money on AI didn't take the human out of the process. They moved the human to a different place.
I say this as someone who built a company on AI agents and who has, in every process, one place where a human clicks "approve".
McKinsey calls them high performers: companies that attribute at least 5 percent of profit to AI and speak of significant value. That is 6 in 100 surveyed, the same as a year ago. They differ from the rest not by the model but by the way they work. Three quarters of them rebuilt processes around AI, against a quarter of the rest. They are 3.3 times more likely to intend to transform their business with AI within three years. Twice as often they have a board that is engaged and a defined way of measuring the effect. A year earlier the same survey showed that 65 in 100 of these companies use a human in the loop. BCG reached the same 6 percent and the same conclusion: value comes not from the number of pilots but from a few processes rebuilt end to end.
Anthropic, in its economic index, measured how people really use models: in 52 percent of cases AI supports a human, in 45 it does the task itself. Support wins even where automation is technically possible.
What it looks like in practice. In a process with AI there are two kinds of steps: those that require judgment and those around it. Collecting data, reading, classifying, preparing a proposal, executing after the decision. Companies that win don't automate judgment. They automate everything around it.
At the start the process looks like this: a human decides, AI collects and reads, AI proposes an action, a human approves, AI executes. Only when months show that the proposals are good does the process move one step: AI decides in typical cases, executes, and a human handles the exceptions. Not the other way round.
At our company it works like this. An agent answers customers on the website. A second agent reads all the conversations every day and proposes fixes to the knowledge base, each already tested on a control set. A human sees the test result and clicks "approve" or not. It takes fifteen minutes a day. Without that quarter of an hour, the agent answers worse after six months, because nobody knows something has changed.
The same in sales. The agent recognizes that the customer wants to buy and asks whether to connect them with a human. It doesn't sell on its own. A human takes over the conversation with the whole history and closes.
A human in the loop has its cost: someone has to be in it. If there is no person in the company who looks at the report once a day, a process without approval will run longer than a process with an empty approval slot. Decide who does it before you deploy.
Today's task: take one process you want to automate and mark in it the one place where someone makes a decision. Everything before and after it can be handed to AI. That place cannot.
Want to see how the agent handles your customers' questions? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your website's offer, ready to be quizzed.
Sources
- McKinsey, The State of AI in 2026: On the road to ROI: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- McKinsey, The State of AI in 2025: Agents, innovation, and transformation: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
- BCG, Look Past Productivity to Get Real Value from AI (AI Radar 2026): https://www.bcg.com/publications/2026/why-ai-pilots-rarely-deliver-value
- Anthropic Economic Index, report IV (November 2025 data)