Your AI Agent Can Finish the Work. It Can't Sign Off on It.

When an owner asks whether AI agents can work inside his company, I don't answer that question first. I ask a different one: when this task is finished, whose name goes on it?

The question is easy to answer. What matters is the silence that follows it.

The owner asks what the agent can do. The blocker is who signs.

Most stalled adoption projects are not stalled by model capability.

In my experience, a project moves forward once the conversation reaches the question of who owns the output. Projects that never reach that question usually die after the second meeting.

At an engineering consultancy in Taichung, the owner spent the meeting asking whether an agent could generate drawings and connect to their project management system. A project manager in his fifties asked one thing: whose name goes at the bottom of the drawing.

His question was more accurate than the entire presentation.

In his role, whether a task can be handed off has never been about speed. It is about who gets called when something fails.

An AI agent can complete an entire body of work. It cannot be the person who gets called when that work fails.

Approval chains are accountability design, not efficiency design

In Taiwan, a decision moves through a chain of approvals. Each person stamps a personal seal — a physical chop carrying their name — onto the document before it moves up.

People call this bureaucracy. I used to call it that too.

But the seals are not decoration. Each one means a specific person absorbed a portion of the responsibility at that step. He reviewed it, he approved it, and if it goes wrong he is part of it. The process is slow because responsibility is being transferred one step at a time, and transfer takes time.

This has nothing to do with efficiency. It solves a different problem: if something goes wrong, the trail leads back to a person.

So when we propose handing a step to an agent, the people who designed that process hear something specific. They do not hear “this step can be automated.” They hear “this step will be empty.”

Someone used to be standing in that step.

Slow approval chains are usually not a sign of poor efficiency. They are a sign that a person is carrying risk at every stage.

An agent can finish the job. It cannot be sued.

Engineering makes this obvious.

Suppose I use an AI agent to design a dam structure, then run it through NVIDIA Omniverse — in plain terms, building the entire dam inside a computer first, calculating water pressure, stress, and flow before anything is poured. The output may well be finer and faster than human work. We tender it. We build it.

Then one day it fails.

Regulators arrive. Investigators arrive. Families arrive. In that room, nobody can shrug and say the AI did it. That sentence does not survive a single hearing, because the engineering certification carries a human name. The law needs someone it can fine, delicense, or prosecute.

A model is none of those things. A model cannot even be named as a defendant.

This is not an argument against using AI in engineering. It is an argument that the accountability layer currently has nothing to catch it. The technology arrived years ago. The institutional design has not, and institutional design does not appear on its own just because the tools improved.

I now start by asking who absorbs the cost of an error

I changed my approach. In the first meeting I no longer ask whether an agent can do the task. I ask who pays when the task is wrong.

The answer sorts the work into two piles.

The scope that survives this exercise is much smaller. It also moves.

And the quiet project manager usually starts talking at this point. You are not taking his seal away. You are removing the three hours that come before he uses it.

The real bottleneck in AI adoption is not model capability. It is that no one can yet sign their name on behalf of a model.

Once he stamps it, it's his

Before we left, I asked the project manager why he only ever asked about the name.

He said, “Once I stamp it, it's mine.”

Until that sentence has an answer, no tool gets into his step of the process.