The Human Measure of AI

In the first part, What Your Friend Knows and Your Colleague Shouldn’t, I suggested handling confidentiality with AI assistants the same way we handle it with people. The same pattern shows up in other places, too. We build trust with agents in the same steps as with a new colleague, and we split complex tasks for agents the same way a company splits its work into departments.

This article looks at both, including the limitations we import with them, and explains why I think it is good news. As we build AI systems the way we understand people, we keep the concepts we need to control them, and that’s a good reason to worry less about AI and to make more use of its opportunities.

The Human as the Measure

“Man is the measure of all things” goes back to the Greek philosopher Protagoras, who lived almost 2,500 years ago. He meant it as a statement about knowledge: whatever we say about the world, we say from our own perspective. For new technology there is a very practical reading of it. We understand something new by comparing it with something we already know, and there’s nothing we know as well as ourselves and the way we deal with each other.

In the first part this was the basis of the argument: if you know how to talk about confidential matters with friends, colleagues and therapists, you have almost everything you need for an assistant. Let’s look at two more places where we do the same.

Trust Is Built Step by Step

Let’s say a company hires a new colleague. In Germany, a probation period of up to six months is common, and during this time the colleague gets small tasks first, and somebody looks at the results. Payments above a certain amount need a second signature, and the authority to sign for the company comes later, if at all. Trust grows with the track record, and with it the scope of what the colleague may decide alone.

We give agents trust in the same steps. At first, a coding assistant asks before every change and every command, and we approve each one. Once we’ve seen enough, we allow certain commands for good, let it edit files without asking, or let it work in a sandbox which limits it to one folder. What it does is logged, so we can review it afterwards. And as with the colleague, the scope grows with the track record, and it shrinks again after a mistake.

The trust zones from the first part belong here, too. An agent which may only see one project and one folder is in the same position as a colleague who is briefed for their own work package and nothing else.

Cutting Work So It Fits into a Head

Let’s say a family plans a wedding with 120 guests. One person takes care of the guest list, another one of the venue and the caterer. The caterer learns the number of guests, the allergies and the time of the dinner, but not who isn’t speaking to whom. Nobody has to keep the whole wedding in mind, and each person gets what they need for their part.

A company works the same way at a larger scale. If it has a large number of tasks and a lot of information to process, it is split into departments and teams, so that each person only deals with an area of information which still fits into their head. Herbert Simon’s work on bounded rationality, which started with his book Administrative Behavior in 1947, is a classic reference for the idea that organizations have to deal with the limits of human attention and knowledge. More recently, Matthew Skelton and Manuel Pais made cognitive load the guiding principle in Team Topologies (2019) for deciding how much a team should be responsible for.

Agents have the same limit in a different form. A language model can only process a limited context, and it gets less reliable well before this limit is reached. In 2025 the vector database company Chroma tested 18 models and found that their results became less reliable as the input grew longer, even for simple tasks.

So we cut tasks for agents the way we cut them for people. In the research system which Anthropic described in 2025, a lead agent splits a question into subtasks and gives each subagent an objective, an output format, the tools and sources to use, and the boundaries of its task. That is a briefing as a project lead would write it. Coding assistants start subagents with a context window of their own, which hand back a summary of their results. Anthropic also writes that this approach does not fit tasks in which all agents need the same context or depend on each other a lot. Every company knows this cost: what’s split across departments has to be put together again at the interfaces.

Melvin Conway observed in 1968 that organizations which design systems produce designs which copy their own communication structures. Agent systems are a very clear case of this, as we give them a lead, roles, briefings and handovers, as we would for a department. Side by side, the two look like this:

The Limitations Come Along

There’s a counterargument, and Marshall McLuhan made it in 1967: we look at the present through a rear-view mirror. New media imitate the old ones at first. The first cars looked like carriages without horses, and it took a while until people asked what the new thing could do which the old one couldn’t.

If we model AI on people, we also import limitations which exist because of people. Two examples:

• Hierarchies, spans of control and coordination meetings exist because a person can only talk to a few others at the same time. An agent can be started as often as needed, so an agent system may need far less hierarchy than the organization it copies.
• A person cannot unlearn what they once heard, which is why we need trust and sanctions in the first place. A context can be discarded, and for a system which can forget, there may be better answers than the ones we know from people.

There will be more of these places, and some of them will change how we build agent systems. Yet, this does not speak against the human pattern, as the pattern is the starting point from which we notice the differences.

Why This Keeps Us in Control

Above all, building AI in our own image keeps the systems understandable. A role called “reviewer” is clear to everyone, a briefing with an objective and boundaries can be read by anyone, and a permission which grows with the track record is something most managers have given before. These are the concepts with which we control the work of people, and we can apply them to agents without learning new ones.

In some places we even get more control than with people. We can log every input an agent had, which we could never do for a colleague, and we can end a context when the task is done. McLuhan described media as extensions of man in 1964. An assistant fits this picture, as it does research, writes and programs much like we would, only faster, and so our rules for this work can be extended to it.

Fewer Worries, More Opportunities

A large part of the public discussion about AI is about control: do we still understand what a system does, and can we stop it if it does something we don’t want? These are fair questions, and I don’t want to wave them away. I think the way we’re building AI systems gives a good part of the answer, as we build them with roles, briefings, approvals and logs, which are the instruments we’ve used to control work for a long time.

So I’d worry less and make more use of the opportunities. Start with small tasks, check the results and extend the trust as the track record grows, as you would with a new colleague. Cut larger tasks so that each part fits into one context, as you would for a team. And keep confidential things in their own circle, as described in the first part.

What I assume here is that we keep building AI systems this way. Whether we’ll leave the human pattern behind at some stage, as the car left the carriage behind, I don’t know. As long as we build agents like colleagues and departments, I’d expect us to keep enough control over them to feel comfortable working with them.

Conclusion

We build trust with AI agents the way we build it with people, step by step and with a growing scope, and we cut complex tasks for agents the way companies cut their work into departments, so that each part fits into one head or one context window. The human pattern brings along some limitations, e.g. hierarchies which agents may not need, and we’ll probably find more of them. Above all, though, it keeps the concepts with which we control work, so that we can keep enough control to feel comfortable with AI. That’s a good reason to fear it less and to make more use of its opportunities.