Chapter 22: Identifying Automation Candidates

Not every process should be automated. Not every task suits agentic AI. Choosing the right candidates is critical.

The Automation Sweet Spot

The best candidates share common characteristics:

• High volume: Enough activity to justify the investment
• Repeatable: Consistent patterns that can be learned
• Rules-based (mostly): Clear logic, even if complex
• Bounded scope: Well-defined inputs and outputs
• Low consequence of errors: Mistakes are recoverable
• Data is available: Information needed is accessible

The Human-Best Zone

Some tasks should remain human-led:

• High stakes/irreversible: Decisions that can't be undone
• Novel/unprecedented: Situations without clear precedent
• Deeply relational: Work that depends on human connection
• Ethically complex: Decisions requiring moral judgment
• Politically sensitive: Situations requiring organisational navigation

A Simple Assessment Framework

For each candidate process, score these dimensions (1-5):

Higher total scores = better automation candidates.

Start Assistive, Graduate to Autonomous

Even for good candidates, consider a phased approach:

1. Assistive: Agent drafts, human approves
2. Active: Agent executes routine cases, human handles exceptions
3. Autonomous: Agent handles end-to-end, human governs

The PIMS asset register agent described in Chapter 4 illustrates this graduation in practice. It started as a Level 2 (Assistive) capability — the agent suggests, the human approves. Adding a confidence threshold graduates it to Level 3 (Active), where high-confidence extractions proceed automatically and only uncertain ones are escalated. No architectural change is required — just an update to the agent's DNA and a confidence threshold in the code.