Automation Triage
Takes a captured process and routes each step to the right execution model — keeping humans where judgment matters, handing rule-based drudgery to RPA, and giving genuinely ambiguous work to an AI agent. The output is a per-step verdict that becomes the execution model of the SOP:
process-standardization → automation-triage → sop-authoring
(discover + capture) (decide the how) (write the SOP)
Language rule
The skill's questions and working notes are in English. Generated deliverables follow the engagement's working language — English by default, Spanish for Costa Rica engagements. The per-step rationale must be legible to a non-technical executive: cost, error rate, brittleness, maintenance — not RPA jargon.
Output directory
Resolve the pipeline directory {sops-dir} first, in this order:
- If a business brain exists in the working directory (a
business-brain/folder, or a vault root with_procedimientos/— see/business-brain-setup), then{sops-dir}=<brain-root>/_procedimientos/. SOPs that belong to a business LIVE in its business brain; that folder is their canonical home. - Otherwise (no brain yet, e.g. a client engagement without one),
{sops-dir}=./fractional-cto/sops/.
Flujo del skill
Paso 1 — Obtener el proceso
Two paths, mirroring /project-takeover:
Path A (preferred): capture available.
Check for {sops-dir}/{slug}/{slug}-capture.md from /process-standardization. If it
exists, load the ordered step list and confirm with the user:
"I found your captured process for {slug} — {N} steps. Triaging each one. Correct?"
Path B (standalone): no capture. Run a condensed interview. Ask the user to list the process steps, and for each: what triggers it, what tool it touches, and how often it runs.
Paso 2 — Triage each step
Score every step on four axes, then read the verdict off the matrix.
| Axis | Question | Low ↔ High |
|---|---|---|
| Volume / frequency | How often does this step run? | rarely ↔ constantly |
| Rule-determinism | Is the logic fixed rules, or does it need interpretation? | judgment ↔ deterministic |
| Judgment / ambiguity | Does it need creativity, strategy, or reading a messy situation? | none ↔ high |
| System / UI stability | Do the tools/screens it touches change often? | volatile ↔ stable |
Verdict rules:
| Condition | Verdict |
|---|---|
| High volume · deterministic · low judgment · stable systems | RPA — unattended (bot runs alone) |
| High volume · deterministic · low judgment · needs a human trigger/checkpoint | RPA — attended (bot + human in real time) |
| Deterministic core but with ambiguous inputs or exceptions | Hybrid (RPA for the rote part, human/AI for exceptions) |
| Ambiguous, unstructured, or interpretation-heavy — but not high-stakes judgment | AI agent (plans, decides, adapts) |
| AI does the draft, a human approves/decides | Human + AI (AI proposes, human disposes) |
| Strategy, creativity, high-stakes judgment, or relationship work | Human (keep it) |
DECISION POINT: where a step sits between two verdicts, default to the one with the human closer to the loop. It is cheaper to remove a human gate later than to recover from an unattended bot that ran wrong 10,000 times.
Produce a per-step verdict table.
Paso 3 — Executive rationale
For each non-human verdict, write a one-line rationale a non-technical decision-maker can act on. Use business terms:
- Why this and not a person: [cost / speed / error-rate gain].
- Why RPA and not AI (or vice versa): RPA mimics fixed clicks and keystrokes cheaply but is brittle — it breaks when a screen changes. AI handles ambiguity and adapts but costs more per run and needs guardrails. Reserve humans for judgment, creativity, and relationships.
Frame the whole thing as an incremental transformation, not a big-bang: which one step to automate first (highest volume × most deterministic), then the next. This staged framing is how a manual process becomes AI-assisted one step at a time.
Paso 4 — Risk callouts
Flag the traps before they bite:
SCALE NOTE: RPA bots are brittle. A vendor UI change silently breaks an unattended bot; budget for monitoring and IT maintenance from day one, or the "automation" becomes a hidden outage.
LESSON LEARNED: never make a step unattended until it has run attended long enough to trust it. Attended → hybrid → unattended is the safe promotion path.
Note any step where automating it would remove a human check that catches expensive mistakes — those
stay Human + AI even if technically automatable.
Paso 5 — Handoff
Write the verdict to {sops-dir}/{slug}/{slug}-triage.md: the per-step verdict table, the
executive rationale, the risk callouts, and a recommended automation sequence (what to do first).
Suggested next step:
"Run
/sop-authoringto write the SOP. Steps you triaged as RPA or AI become the executable version of the SOP ({slug}-sop.skill.md) with approval gates; human steps stay in the written procedure."
Notas para el modelo
- This skill does NOT write the SOP — it produces the verdict that parameterizes it.
/sop-authoringturns RPA/AI verdicts into executable steps with gates and human verdicts into written procedure. - The four-axis matrix reuses the methodological pattern of
/takeover-assessment's risk matrix: a structured, defensible verdict from scored inputs. - Bias toward keeping the human close to the loop when uncertain. Automation that runs wrong at scale is more expensive than a human bottleneck.
- Map verdicts to the gate markers used by the executable SOP template:
Human/Human + AI→[APPROVAL REQUIRED]or[PRE-APPROVAL REQUIRED];RPA — unattended/AI agent→ no gate. - Keep the rationale executive-legible. If a sentence needs an RPA glossary to parse, rewrite it.
- Use the Acme universe for examples (e.g. "RapidEats refund step: high volume, deterministic → RPA unattended; dispute-escalation step: ambiguous → Human + AI").