Workflow Decomposer

Overview

Turn human process descriptions into machine-readable task graphs that expose what's actually happening - including the judgment calls, timing bottlenecks, and undocumented expertise that determine where agents can safely operate.

Core principle: Expose what's unknown, not just what's stated. A good decomposition reveals gaps in understanding as clearly as it documents the happy path.

When to Use

• •SME interview transcript needs structuring
• •SOP document needs automation analysis
• •Shadowing notes need systematization
• •Evaluating where AI agents can assist regulated workflows
• •Justifying automation boundaries to compliance

The Decomposition Process

digraph decomposition {
    rankdir=TB;
    node [shape=box];

    extract [label="1. Extract Raw Steps\n(what was said/written)"];
    judgment [label="2. Mark Judgment Points\n(where human decides)"];
    gaps [label="3. Flag Knowledge Gaps\n(what's NOT explained)"];
    timing [label="4. Add Timing Layer\n(latency per step)"];
    volume [label="5. Add Volume Layer\n(repetition markers)"];
    compliance [label="6. Tag Compliance Needs\n(audit/approval/segregation)"];
    output [label="7. Generate Task Graph\n+ Required Follow-ups"];

    extract -> judgment -> gaps -> timing -> volume -> compliance -> output;
}

Do NOT skip steps. The value is in steps 2-6, not just step 1.

Step Details

1. Extract Raw Steps

List every action mentioned, in sequence. Don't interpret yet.

2. Mark Judgment Points (CRITICAL)

Look for these signals in the source material:

For each judgment point, document:

• •What decision is being made
• •What information the human uses
• •What the decision criteria ARE (or flag as UNKNOWN)
• •Whether criteria could be codified

3. Flag Knowledge Gaps

Create explicit GAPS section listing:

• •Processes mentioned but not explained
• •Decision criteria not specified
• •Exception handling not described
• •"Shortcuts" or optimizations mentioned but not detailed

Each gap becomes a required follow-up question.

4. Add Timing Layer

For EACH step, estimate or flag:

• •Work time: Actual human effort
• •Wait time: System response, approvals, external parties
• •Variability: Normal vs exception case timing

The SME saying "could be an hour" is not enough - WHERE does that hour go?

5. Add Volume/Repetition Markers

Flag steps that are:

• •High-frequency (many times per day/week)
• •Batch-able (similar items grouped)
• •Seasonal (month-end, quarter-end spikes)

Repetitive + deterministic = automation candidate.

6. Tag Compliance Requirements

For regulated environments, mark:

• •Audit trail: Must log who/what/when
• •Approval: Requires sign-off (by whom? threshold?)
• •Four eyes: Requires independent verification
• •Segregation: Cannot be done by same person as step X

Output Format

workflow:
  name: [Process Name]
  source: [interview/sop/shadowing]
  source_date: [when captured]

  steps:
    - id: step_1
      name: [Action Name]
      type: [action/decision/validation/system_action]
      inputs: [what's needed]
      outputs: [what's produced]

      # REQUIRED for decision/validation types
      judgment:
        decision: [what's being decided]
        criteria: [how it's decided, or "UNKNOWN - requires follow-up"]
        codifiable: [yes/no/partial]

      timing:
        work_time: [estimate or "UNKNOWN"]
        wait_time: [estimate or "UNKNOWN"]
        exception_multiplier: [Nx normal, or "UNKNOWN"]

      volume:
        frequency: [per day/week/month]
        batch_potential: [yes/no]
        peak_periods: [when volume spikes]

      compliance:
        audit_required: [yes/no]
        approval_required: [yes/no, by whom]
        segregation: [cannot follow step X]

      automation_assessment:
        level: [full/assisted/human_required]
        blockers: [what prevents automation]

  gaps:
    - id: gap_1
      description: [what's not explained]
      source_quote: [verbatim from input]
      follow_up_question: [specific question to ask]
      blocking: [yes/no - can we proceed without this?]

Automation Assessment Rules

Do NOT assess automation potential for a step until:

1. •The process for that step is documented (not just mentioned)
2. •Decision criteria are defined (not UNKNOWN)
3. •Inputs and outputs are specified

If any blocking gap exists for a step:

automation_assessment:
  level: BLOCKED
  blocker: "GAP-X: [description]"
  prerequisite: "Resolve gap before assessment"

Do not speculate about automation potential for undefined processes. "UNKNOWN" and "BLOCKED" are valid, honest outputs.

Common Mistakes

Red Flags in Your Output

If your decomposition has these, you're not done:

• •Steps without timing estimates (even rough ones)
• •Decision steps without documented criteria
• •"To be determined" or "varies" without follow-up flagged
• •Gaps section is empty (there are ALWAYS gaps)
• •No compliance tags in regulated workflow
• •Automation assessment without blockers listed

Rationalizations to Avoid

Financial Services Context

This skill justifies where agents can safely take over without destabilizing regulated workflows:

1. •If judgment criteria can't be codified → Human stays in loop
2. •If compliance requires audit trail → Agent must log everything
3. •If timing is dominated by wait time → Agent can't help much
4. •If volume is high AND logic is deterministic → Prime automation target

The output should make these tradeoffs explicit and defensible to compliance.