The planner releases it
On 16 October Anthropic published a report on how work has changed inside the company. By August, the agents its employees direct were working six hours for every hour the employees worked. Most of that work is code.
The part that caught my attention is further in. People became the bottleneck. In interactive sessions, an agent spends about 80% of its time waiting for a person. Reading everything the agents produce would take non-technical employees more than half their week. One engineer describes facing "an army of Claudes, with every one asking for my attention."
In industry there is a person in the middle too, for a different reason. At Anthropic, the person can't keep up with what the agents produce. In a plant, the person decides when an agent is allowed to act at all.
The usual worry is a wrong move that can't be undone. Sometimes that is right. A plant is not a code repository, and some mistakes in a plant cannot be taken back. But a lot of what an agent would do can be undone with one click. A work order can be cancelled. So I think much of the hesitation is cultural. Some people prefer old sound systems because they are "better". It is the same resistance to something new.
That does not mean the answer is to push. The first thing an agent is usually allowed to do on its own is send a report or an email. What moves people from there to letting it open a work order is confidence, and confidence is not lost one alert at a time. A few false positives, one hallucination, and people stop trusting the whole thing, not just the alert that was wrong.
I saw it several times as a consultant. People tried to rush, lost confidence and killed the program.
So with one client, when we looked at letting our agents open work orders on their own, we decided together to start one step back. The agent drafts the work order and a planner releases it. Every draft released as it was written is the agent being right in front of the person who will one day decide to let it act.
Anthropic can show what being right looks like in its own work. Code written by its agents was reverted no more often than code written by its people. In a plant, the same number is how many drafts the planner released without changing anything.
What I'd bet on
Of the industrial AI programs cancelled between now and the end of 2028, more will be cancelled for going live too early than for costing too much.
Diego Mercadal started as a commissioning engineer on high-voltage motors and generators, joined an offshore drilling contractor as a rig hand, worked several positions in the drilling crew, and ended up running its AI/ML function. He is now co-founder and CEO of Wonder DataLabs.