In 30 seconds

Work can be handled at different levels, from fully manual to an AI that chooses and performs actions on its own. Each step up adds capability, and also adds risk and maintenance.

The goal is not to climb to the top. It is to stop at the level that solves the problem reliably.

The spectrum

MANUAL
Human performs the task
      ↓
DIGITAL
Software records and moves information
      ↓
DETERMINISTIC AUTOMATION
Fixed rules perform predictable actions
      ↓
AI ASSISTED
AI drafts, extracts, classifies or recommends
      ↓
AI WITH APPROVAL
AI acts only after a human decision
      ↓
AGENTIC
AI chooses and performs actions within defined authority

Each level is a valid place to stop. A workflow that is comfortably handled by a fixed rule does not become better by adding a model.

Why higher is not automatically better

Every level up the spectrum adds three costs: more ways to be wrong, more things to test, and more to maintain when the world changes.

A deterministic rule fails in predictable ways. An AI-assisted step can be wrong in surprising ways. An agent that acts on its own can be wrong and take action before anyone notices.

That does not mean stay low. It means the extra capability should be earning its keep.

The question that replaces "Can AI do this?"

The useful question is not whether AI is capable of a task. Modern models are capable of a great deal.

The useful question is:

That single question saves a surprising amount of time and money. See Use the least complex mechanism.

A small example

A company routes incoming email to the right team.

  • Manual: someone reads each message and forwards it.
  • Deterministic: messages to invoices@ always go to finance.
  • AI assisted: an AI suggests a team for ambiguous messages, a person confirms.
  • Agentic: the AI reads, decides and routes on its own within set limits.

Most of the value often comes from the first two levels. The AI earns its place only for the genuinely ambiguous messages that rules cannot sort.

What this does not solve

The spectrum tells you how much capability a step needs. It does not tell you whether your data is reliable, who owns the workflow, or how you will test it. Those are separate stages.

Good enough to move on when

  • You can place each step of your workflow on the spectrum.
  • You can justify any step that sits above deterministic automation.
  • You are not adding autonomy the task does not need.