Automation Decision Matrix

Summary

A practical tool for turning the BAU vs Pilot intuition into a decision: should this task be automated, and at what autonomy level? Built around two axes — how repeatable the task is, and how high the stakes are.

The 2×2

                 HIGH STAKES (Pilot)
                        │
   L1 Assisted /        │      L2 Approval-gated
   keep human           │      (agent does it,
   (judgment-heavy,     │       human approves)
   one-off)             │
   ─────────────────────┼─────────────────────  →  HIGH REPEATABILITY
   L0 Manual /          │      L3 Autonomous
   not worth automating │      (full agent,
   (rare, low value)    │       audit log)
                        │
                 LOW STAKES (BAU)

• Top-left (high stakes, low repeatability): creative/strategic one-offs → human or L1 assisted.
• Top-right (high stakes, high repeatability): the sweet spot for L2 approval-gated automation — speed and control.
• Bottom-right (low stakes, high repeatability): automate fully, L3 autonomous. The clearest wins.
• Bottom-left (low stakes, rare): leave manual; automating costs more than it saves.

Scoring a task (quick rubric)

Score each 1–5; higher total → higher autonomy.

Factor	1 (keep human)	5 (automate)
Repeatability	one-off	daily, identical shape
Rule-clarity	fuzzy judgment	explicit rules exist
Reversibility	hard to undo	trivially reversible
Stakes if wrong	brand/legal/money	cosmetic
Measurability	no clear metric	clear proxy metric

• 20–25 → L3 autonomous
• 14–19 → L2 approval-gated
• 8–13 → L1 assisted
• <8 → L0 manual

Add the static/dynamic lens first

Before scoring autonomy, classify the task using Static and Dynamic Functions:

• Static tasks usually score higher on repeatability and rule-clarity. They are better candidates for scripts, SOPs, templates, and L3 automation once stakes are low.
• Dynamic tasks usually score lower on rule-clarity and reversibility. They may still be automated, but usually as L1/L2 assisted work with human judgment around the final decision.
• Hybrid tasks should be split before scoring. Example: “digital marketing” is too broad; website tracking setup is static, while content strategy is dynamic.

This prevents broad departments from being misclassified. Score the specific task, not the umbrella.

The approval-gate rule

Some things never reach full autonomy regardless of score

Any task that moves money, files with authorities, or speaks publicly as the brand in a non-reversible way caps at L2 (approval-gated) — never L3. Examples: ad budget changes, payments, tax filings, PR responses, partnership commitments. The surrounding production can be autonomous; the commit needs a human click.

Don’t over-classify — the script test

Many "automation candidates" aren't agent tasks at all

Before assigning an LLM agent, ask: does this need judgment, or just execution? News scraping, tracker audits, scheduling, inventory sync are deterministic — a plain script/cron does them more reliably and cheaply than an LLM. Reserve agents for genuinely fuzzy work (content angles, copywriting, analysis). See 5. Don’t reach for an LLM when a script will do.

Output of this exercise

For every task you classify, record:

• Autonomy level (L0–L4)
• Whether it’s script or agent
• The approval gate (if any)
• Where it logs its work

…in the agent spec and the decision log.

Related

• Part B2 - BAU vs Pilot Framework — the conceptual layer
• Part A2 - Multi-Agent Orchestration — building the automatable ones