NEC has a department with AI bosses, AI managers and AI staff. Here's what really gets faster, what the reported ¥350,000–¥410,000 figures mean, and why humans still sign off.
AI can take over a large share of status tracking, task assignment, reporting and analysis. Humans still have to settle conflicts, make consequential decisions and own the results. NEC's experiment is fascinating precisely because it draws that line at the level of an entire department—not because 35 human employees have magically been replaced.[1]
An AI department head, AI managers and AI workers. Who books the team dinner?
On August 1, 2026, Japanese technology company NEC established its Corporate AI Workforce Department. Its August 10 announcement described four levels: an AI department head, an executive-style AI board, AI managers and AI employees.[1]
The managers assign and monitor work; AI workers perform tasks; a separate layer evaluates the operation. The system even gathers improvement suggestions from its AI staff and can address some issues internally.
An AI department dinner sounds straightforward until somebody asks the AI boss to reserve a table. "Request denied: restaurant-booking permission not granted." Corporate culture meets access control.
One detail matters: NEC explicitly leaves final evaluation, decision-making and quality oversight to people. The organization has AI roles; it has not transferred legal or business accountability to a chatbot.[1]
The biggest gain may be skipping the wait for your boss
NEC says its trial cut the time required for specified management analysis, scenario planning and early risk detection to about one-seventh of the previous time.[1] That is a result for particular tasks, not proof that the entire company became seven times as productive.
Think about an ordinary reporting chain. Someone asks for a presentation. Another person arranges a meeting to review it. A third person prepares an update on the progress of the presentation. Suddenly the job has grown a family tree.
AI systems can collect information, draft reports, compare options and revise them in parallel. They do not need a diary slot or a carefully worded message to protect a superior's pride. That can remove waiting and coordination overhead from computer-based work.
But relationships are not pointless overhead. Someone still has to negotiate with a supplier, support a stressed colleague or explain a hard choice to an affected person.
Does 35 AI agents for ¥350,000 mean 35 salaries disappeared?
News reports described roughly 35 AI agents and an AI processing charge of about ¥350,000 a month.[2] A widely shared October 14 post gave figures of roughly ¥350,000 for July and ¥410,000 for August.[3]
The tempting calculation is ¥350,000 divided by 35: around ¥10,000 per agent. Then comes the dramatic headline, "Thirty-five employees for the cost of one!" There is a catch. An AI role is not equivalent to a full-time human worker with the same workload, authority, reliability or accountability.
Usage fees also differ from the total cost of designing a system, connecting company data, securing it, supervising results and repairing mistakes. NEC's official release does not specify the agent count or those monthly amounts.[1]
A meaningful comparison asks: how much did the same verified output cost before and after, including human checking time and rework?
Which management jobs are most exposed?
Collecting updates, distributing repeatable tasks and summarizing a dozen reports are good candidates for automation. Resolving a disagreement, accepting responsibility for a budget, protecting a customer or making a safety call require an accountable human decision-maker.
There is another trap: AI agents can agree with one another while sharing the same false assumption. A polished report is no guarantee that the underlying numbers are right. Research published in September 2026 points to two opposite failures: too many human approvals create bottlenecks; too little supervision lets errors build up quietly.[4]
A skilled individual contributor may soon coordinate a small fleet of AI researchers, writers and analysts. In some settings, that person could deliver more than a traditional manager whose main output is passing reports up and down a hierarchy. That is a plausible direction, not a universal prediction.
How would you test an AI manager fairly?
Measure end-to-end completion time, corrections, missed errors, total cost and decision quality on comparable tasks. Count the human hours spent verifying the AI's work, too. Begin with reversible internal tasks; keep meaningful human sign-off for contracts, payments, employment and safety.
The question is not whether the AI deserves a corner office. It is whether a piece of work gets completed more reliably with less wasted effort.
The last boss may still be human
NEC offers a glimpse of flatter organizations: fewer people serving as report relays, more people setting priorities and taking responsibility for decisions.
Ask an AI manager for an update and it may respond in seconds. Ask a human organization to schedule a meeting to prepare the update for the meeting about the update, and the AI may discover the strangest business problem of all.
Why does a company sometimes create three new tasks to confirm that one task is finished?
References (4)
- NEC press release, August 10, 2026 jpn.nec.com
- Nikkei report on NEC's AI-run department (reported figures) nikkei.com
- October 14, 2026 social post quoting July/August costs x.com
- Yan et al., “Managerial altitude,” September 8, 2026 sciencedirect.com
