The team named it Lisa. It was meant as a joke, a way to make the new AI agent feel like part of the team. But then something shifted. “Lisa thinks we should go with the higher bid,” someone said in a meeting. No one questioned it. The decision was made.
Language is not a nice extra. Language is a steering decision. When we call a tool a colleague, people begin to behave as if responsibility has moved. That is usually the beginning of confusion, not the beginning of progress.
I research how AI changes responsibility, trust, and decision habits in teams, and I have learned to be careful with the stories we tell. If you’ve been following me, you know that I gain a lot of insights from old sources. In this case, I think Aristotle is useful because he would not start with what the tool can do. He would start with what the tool is for.
Purpose comes first, always
An Aristotle style test is surprisingly simple. You ask, “What is this for?” You do not ask, “What is possible?” You ask, “What is good?” and “Good for whom?”
So before you deploy any agents, you need to be able to say, in one clear sentence, what it is meant to improve for humans. You also need to be able to say what it must not do, even if it looks efficient. If you cannot say both, you are not ready to scale. If you do not choose the purpose, your organisation will choose it for you. The default purpose tends to become speed, because speed is measurable and rewarded. This is how tools quietly reshape cultures.
“Surely that for the sake of which all else is done.” Aristotle
A tool is not a person, even if it speaks like one
A colleague can be accountable. A tool cannot be accountable. A colleague can refuse unethical work. A tool cannot, a colleague can carry consequences. A tool cannot. A colleague can care, a tool cannot.
This is why the “AI colleague” metaphor is not harmless. It blurs who makes choices and who owns outcomes. It encourages people to say “the AI decided” when the truth is that humans decided to let a system act, and humans set the boundaries, or failed to set them. It also tempts teams to outsource judgement. That outsourcing rarely happens in one dramatic moment. It happens quietly. The agent sounds confident. The team is busy. The human brain wants relief. Over time, “check it” becomes “accept it.”
“It’s not going to be about human vs. machine" ...... "We will still expect a human to be ultimately accountable for the outcomes.”
Microsoft CEO Satya Nadella
The lifebuoy, not the swimmer
I prefer the lifebuoy metaphor. A lifebuoy supports you when the water gets rough. It can stabilise you. It can prevent panic. It can buy time. But it does not swim for you. It does not decide where you should go. And if you hold it wrong, it can still pull you into trouble.
That is how I see AI agents at work. They are extremely helpful when they reduce coordination friction and prepare options that humans can evaluate. They become dangerous when they start replacing human judgement in situations where values, fairness, and consequences matter.
The Hidden Danger: Cultural Amplification
This brings us to the deepest risk, one that hides behind the glossy outputs. Agents do not create a new organisational story. They amplify the one you already live in. If your culture rewards speed over care, the agent will strengthen that. If your culture avoids conflict, the agent will become a convenient way to avoid tough human conversations. If your culture is unclear about accountability, the agent will make that fog thicker. This is why human factors are not a soft add-on. Psychological safety, clear roles, and ongoing learning are the non-negotiable operating system for human-agent teams.
This is the moment in the conversation where the conversation needs a pause:
If this agent makes a bad call, who will feel it first, and who will be blamed? Are we building capability in the team, or are we building dependence on the system? What part of our culture will this agent amplify, and are we proud of that?
These questions are not meant to slow innovation. They are meant to stop fantasy.
The 5 Non-Negotiables Before You Scale
The 5 Non-Negotiables Before You Scale
Wise limits beat maximum autonomy. Before you scale an agent beyond a small pilot, I would insist on clear answers to these five points. Each one should be answerable in plain language, and each one must have a named owner.
01: One-Sentence Purpose. We can explain what the agent is for and name the specific human benefit. We have also written down what it must never do.
02: Named Accountability. One person or role is clearly accountable for the outcomes. Not “the system,” not the vendor, but a human.
03: Defined Autonomy Level. We can state whether the agent only suggests, acts with explicit approval, or acts on its own within a defined safe zone.
04: Mandatory Human Handoff. We have clear stop rules for when the agent must hand off to a human (e.g., high uncertainty, impact on people, legal or reputational risk).
05: An Instant “Off” Switch. We can pause the agent instantly, and we can undo its most recent actions.
If you do not have these answers, you do not yet have a human-agent team. You have an experiment without a safety rail. And that is a story that rarely ends well.
„For the things we have to learn before we can do them, we learn by doing them.“ Aristotle