What "agentic" really means
This module sets the ground the rest of the course stands on: what an agent actually does, what today's AI still cannot do, and why — for now — the experienced person in your business knows the things the machine does not. Not chatbots, not hype. A plain account, with one real example, of why the human stays in command.
1.1 Answering is not acting
A chatbot answers a question. An agent takes an action — it sends, schedules, files, provisions, updates. That is the whole subject of this course: an action carries consequences a sentence does not, and an action-taker must be held to a different standard than an answer-giver. The moment a tool can do rather than only say, the question stops being "is it clever?" and becomes "what will it do, and to whom?"
Discussion topics
- Where in your business would an AI acting on its own carry real consequence — and where would it merely be convenient?
- What is the most consequential action you would never want taken without a person in the loop?
1.2 Efficiency is not the point
The first thing most organisations do with AI is the wrong thing: they use it to run their existing processes faster. But speed applied to a flawed process only produces the flaw sooner. An agent that automates a broken step repeats the same mistake at scale and at pace — you have not fixed it, you have industrialised it. And automating a process that is no longer fit for purpose entrenches it: once a machine depends on it, it becomes harder to question. Before "can an agent do this faster?", ask the prior question — "should this work be done at all, and in this shape?" The agentic shift is an invitation to rethink the work, not a tool for cementing it.
And efficiency is the smaller prize. AI is not mainly a way to do the same things more cheaply. Its real value is in effectiveness — doing the right work, and doing it better — and in the room it opens to innovate, to offer what you could not offer before. Efficiency is worth having, but on its own it is close to irrelevant, and more often than not a fast road to failure. The reason is competitive, not moral: if you only use AI to run your existing model faster, you are polishing a position your competitors are using AI to move past. Efficiency tunes the engine; effectiveness and innovation decide whether you are still driving the right vehicle.
Three kinds of gain
- Efficiency — the same work, faster or cheaper. Worth having; rarely decisive on its own.
- Effectiveness — the right work, done better. Where the real advantage lies.
- Innovation — value you could not offer before. AI as a springboard, not just a lever.
Discussion topics
- Which of your processes would you not design this way if you were starting today?
- Where would speeding a process up make it harder to change later — and would that matter to you, or not?
1.3 The thing today's AI cannot do
Yann LeCun, the field's most cited researcher, left the largest AI lab to say it plainly: today's models cannot truly reason or plan, because they have no model of the world — they cannot predict the consequences of their actions. They have read almost everything and lived nowhere. They can sound right and have no idea what their next move will do. Fluency is not understanding.
Further reading
- Taonga in the Latent Space — what a "world model" is, and why an agent must foresee what its action will do, brought down to a community and a business.
- AI that stays in its lane — bounded competence and built-in deference: the shape of an agent you can trust.
- Yann LeCun — A Path Towards Autonomous Machine Intelligence — the world-model architecture this module draws on.
- See the full Further reading list.
1.4 Two ways to act
There are two ways to act. The first is reactive — straight from impression to action, with nothing weighed. The second imagines the result, and rejects what breaks the rules, before it moves. Today's agents mostly do the first. The second is what makes an action-taker safe — and, for now, the second is supplied by a person.
Key teaching points
- Reactive action is fast and blind to consequence; considered action weighs the result first.
- An agent without a model of your world cannot do the second on its own.
- The person who weighs the consequence is doing the work the machine cannot.
1.5 You are the world model it does not have
The experienced person in your business already carries the picture the machine does not — what this client does if the invoice is wrong, what the service breaks if a step is skipped, which promise must never be missed. Until a governed model of your world exists, that person is the world model. This is the technical reason people come first in this course, not a sentiment. Remove the person who holds the picture, and you have removed the only thing that can foresee the harm.
Discussion topics
- Whose departure would cost you knowledge you could not quickly rebuild?
- Where do you already rely on a person's foresight to catch what a checklist would miss?
- Where in your work is a person actually the unreliable part — and does that change where you'd want a machine to act?
1.6 A worked example
One agentic tool built on this principle watches a platform's operations — its health, its alerts, its failures — and acts within a narrow lane more steadily than a person could. It exceeds a human at the watching. And it stops: anything that touches a member, a payment, or an irreversible change is handed to a person to decide. Superhuman in its lane; deferential at the edge. That pairing — capable where it is bounded, and answerable where it is not — is the shape worth looking for.
1.7 What to ask of any agentic tool
When you weigh an agentic tool for your business, the question is not how clever it is. It is two plainer ones: does it exceed a person at one well-bounded job, and does it hand back, by rule, everything that should stay human? Cleverness without that second property is not yet worth your trust. Hold both questions in mind through the rest of the course; Module 2 turns them into a method you run across your whole business.
Key teaching points
- Bounded competence beats broad cleverness when an agent can act.
- Built-in deference — not goodwill — is what makes an agent safe to trust with real work.
- Cleverness without deference is a hazard, not an asset.
Questions by role — not by rank
Grouped by role, not by rank — and the roles are flattening: in a small business one person often holds several, and in a one-person-plus-agents business, one person holds them all. What an action is, and who it lands on, only comes clear if every role can speak to it — the newest included.
(a board, the owners — or you.) Where would an agent acting on its own carry real consequence in our work — and are we reaching for efficiency when the prize is effectiveness, and the room to do what we could not before?
(whoever runs the change — or you.) Hold any tool to the two plain tests before it acts: does it exceed a person at one bounded job, and does it hand back, by rule, everything that should stay human?
(team leads — or you.) Which of our processes would you not design this way if you were starting today — and where would speeding one up make it harder to change later?
(the people closest to the task — including the newest.) You can contribute: where you already hold the picture no software has — what this client does, the step that must never be skipped. You're entitled to see: that the question being asked is "what will it do, and to whom?", not only "is it clever?"
Self-check
1. What is the difference this module says "matters most"?
An action carries consequences a sentence does not — that leap is the whole subject of the course.
2. Why, for now, do people come first — as a technical matter, not only an ethical one?
Until a governed model of your world exists, the person who can foresee the consequence is doing the work the AI cannot.
3. What two things should you ask of any agentic tool?
Bounded competence plus built-in deference. Cleverness without the second is not yet worth your trust.
4. Why is "just use AI to run our existing processes faster" the wrong starting point?
Going faster in the wrong direction is not progress — question the process before you accelerate it.
5. Why is using AI purely for efficiency gains a risky strategy?
Efficiency tunes the engine; effectiveness and innovation decide whether you're still driving the right vehicle.