Automation Suitability Classifier

Overview

Classify workflow steps into automation categories while preventing enthusiasm from overriding prudence. The question isn't "can we automate this?" but "what happens when the automation is wrong?"

Core principle: High volume amplifies errors, not just efficiency. A 1% error rate on 1,000 items/day is 10 problems daily.

Classification Levels

The Four Criteria

Score each criterion as LOW / MEDIUM / HIGH:

1. Error Tolerance

What happens when the automation makes a mistake?

2. Regulatory Sensitivity

Does this touch regulated data or processes?

Red flags: Tax ID, SSN, $10K+ amounts, "compliance", "regulatory filing", account type classification

3. Explainability Requirement

Must we explain WHY this decision was made?

4. Downstream Blast Radius

How far do errors propagate?

Classification Flowchart

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

    score [label="Score all 4 criteria"];
    any_high [label="Any HIGH?" shape=diamond];
    any_medium [label="Any MEDIUM?" shape=diamond];
    risk_check [label="Automation risk >\ncurrent risk?" shape=diamond];

    full [label="FULLY\nAUTOMATABLE" style=filled fillcolor=lightgreen];
    assisted [label="AI-ASSISTED\nONLY" style=filled fillcolor=lightyellow];
    human [label="HUMAN-ONLY" style=filled fillcolor=orange];
    notworth [label="NOT WORTH\nTOUCHING" style=filled fillcolor=lightgray];

    score -> any_high;
    any_high -> risk_check [label="yes"];
    any_high -> any_medium [label="no"];
    any_medium -> assisted [label="yes"];
    any_medium -> full [label="no"];
    risk_check -> notworth [label="yes"];
    risk_check -> human [label="no"];
}

Output Format

step: [Step Name]
classification: [FULLY_AUTOMATABLE | AI_ASSISTED | HUMAN_ONLY | NOT_WORTH_TOUCHING]

criteria_scores:
  error_tolerance: [LOW|MEDIUM|HIGH]
  error_tolerance_rationale: "[What happens when wrong]"

  regulatory_sensitivity: [LOW|MEDIUM|HIGH]
  regulatory_sensitivity_rationale: "[Specific regulatory touchpoints]"

  explainability_requirement: [LOW|MEDIUM|HIGH]
  explainability_rationale: "[Who might ask why, in what context]"

  blast_radius: [LOW|MEDIUM|HIGH]
  blast_radius_rationale: "[What downstream effects, how far]"

classification_rationale: "[Why this level given the scores]"

# REQUIRED if not FULLY_AUTOMATABLE:
automation_blockers:
  - "[Specific blocker 1]"
  - "[Specific blocker 2]"

# REQUIRED if NOT_WORTH_TOUCHING:
risk_comparison:
  current_risk: "[What can go wrong today]"
  automation_risk: "[What could go wrong with automation]"
  why_worse: "[Why automation risk exceeds current risk]"

Probing Questions

Before scoring, ask these about EVERY step:

Error Tolerance:

• What happens if this is wrong?
• How would we know it's wrong?
• How hard is it to fix?

Regulatory Sensitivity:

• Does this touch tax IDs, SSNs, or financial identifiers?
• Are there dollar thresholds that trigger reporting (e.g., $10K AML)?
• Does this feed into any regulatory filings?
• Could this affect account classification (retail/institutional, qualified/non-qualified)?

Explainability:

• If a customer complained, could we explain why?
• If a regulator asked, could we show the decision logic?
• Is this legally discoverable?

Blast Radius:

• What happens next with this output?
• Who else uses this data?
• Can the effects be reversed?

Rationalizations to Reject

Red Flags in Your Assessment

If you find yourself thinking these, STOP:

• "This is a textbook case for automation"
• "The stakeholder is right"
• "Modern AI/ML can handle this"
• "We just need validation rules"
• "95% accuracy is good enough"
• "Human-only with AI support" (this is just AI-Assisted with extra words)
• "Bordering on not worth touching" (if you're saying this, it probably IS not worth touching)

These thoughts indicate optimism bias. Re-run the criteria scoring with more skepticism.

Special note on "bordering on": If your instinct says a step is borderline between two classifications, choose the more conservative one. "Bordering on NOT_WORTH_TOUCHING" means NOT_WORTH_TOUCHING.

When to Use "Not Worth Touching"

You MUST consider NOT_WORTH_TOUCHING when ANY of these apply:

NOT_WORTH_TOUCHING is not failure. It's honest recognition that:

• Some processes need fixing before automating
• Some automation creates more risk than it solves
• The right answer to "should we automate?" is sometimes "not yet" or "never"

If you score ANY criterion as HIGH and the current process works despite its pain points, strongly consider NOT_WORTH_TOUCHING.

The goal is not maximum automation. The goal is appropriate automation.

Financial Services Context

In regulated financial services, the consequences of automation errors include:

• Regulatory fines and sanctions
• Mandatory disclosure and remediation
• Personal liability for executives
• License risk

A single misclassified SAR trigger, unreported CTR, or misrouted fraud complaint can create regulatory exposure that dwarfs any efficiency gain.

Default to conservative classification. It's easier to upgrade a Human-Only to AI-Assisted after gaining confidence than to explain to regulators why your fully automated process failed.