Mauricio Tapia.
← AI course

Chapter 02 · Foundations

Autonomy is not an on/off switch

An agent needs an objective, a working cycle and boundaries. How much it may do alone depends on the risk.

“Let AI take care of it” leaves important questions unanswered: take care of what, using which information, how far may it go and who checks?

An agent combines an objective, state, tools, a working cycle, limits and evaluation.

The cycle that gives it purpose

Observe → decide → act → verify. Verification matters just as much as action. If a system makes recommendations but never checks the outcome, we do not know whether it helped or merely generated activity.

In a service team, the objective might be to reduce overdue requests. The state includes open cases and due dates. Tools let the agent read authorized records and prepare drafts. A boundary may be clear: it proposes replies, while a person approves anything sent to customers. Evaluation asks whether cases were resolved without sacrificing quality.

Autonomy in proportion to risk

  • It can do independently: organize information, detect duplicates, flag exceptions and draft summaries.
  • It can propose: priorities, drafts and alternatives supported by evidence.
  • It must escalate: external messages, changes to commitments or prices, and decisions that are hard to reverse.

Less autonomy does not make a system less useful. The design decision is where human judgment changes the risk and how to measure whether the whole process works better.

Your turn

A system identifies urgent requests and drafts replies. What is needed before it can send those replies automatically: A) a larger model, B) more speed, C) boundaries and risk evaluation, or D) a name?

If you came from LinkedIn, share your answer there before opening the explanation.

Show the explanation

C) Boundaries and risk evaluation. Before expanding autonomy, define sources, exceptions, quality criteria, the consequences of an error and where a person needs to intervene.