Skip to content
Back to the context layer
Match the level to the job

The automation level-matcher

Most “AI projects” are really automation problems one or two levels simpler than the team assumes. Describe the work and see the level that actually fits, and the trap on either side.

Describe the workLive · adjust any answer

Does the work need reading, writing, or judgment?

No language or judgment is the bright line. Without it, AI is the costlier, riskier tool.

How predictable is the input?

Structured data vs. free-form language is the gap rules could never cross.

Do the steps change from case to case?

Who decides the path, you or the AI, is the single biggest architectural choice.

How much runs at once?

Coordinating several agents is the only thing that earns multi-agent overhead.

If it goes wrong, how bad is it?

Stakes don’t change the level. They decide how much guardrail it needs.

The automation spectrum0 → 3
0Rules / RPA
1AI workflow
2Autonomous agent
3Multi-agent system
Deterministic · predictableProbabilistic · adaptive
Recommended starting point

Rules / RPA

Level 0Deterministic
Why this level

There’s no language or judgment in the work. A rule or a bot is faster, cheaper, and auditable.

A fixed “if-this-then-that” path, or a bot that clicks through your apps. Same input, same output, every time. Predictable, cheap, auditable.

Email filters · a thermostat · copying invoice fields between two systems

What this needs to be safe
  • Light guardrails, then ship it and watch
  • A couple of sanity checks on the output
The mismatch trap

Don't over-build ↑

There is no “over-building” below this. But forcing rules onto work that needs judgment is the mirror mistake.

Don't under-build ↓

Under-building isn’t possible here. The risk is the opposite: reaching for an agent when a rule would never fail.

Start at the lowest level that solves the job. Move up only when real variability forces you to. A workflow that runs flawlessly beats an over-engineered agent every time.

A teaching tool, not a verdict. The level is a starting point for a real scoping conversation. A lot of work labeled “AI” turns out to be a plain automation problem. The discipline is matching the rung to the need.