Automation and AI Fundamentals
What you must know
Advanced RPA work sits on a layered stack: deterministic rule-based automation at the base, then orchestration, then intelligence (OCR/IDP, classification, ML scoring, LLM-based extraction or summarisation). Candidates should be able to distinguish deterministic tasks (fixed rules, structured input) from probabilistic ones (unstructured documents, judgement, ambiguity) and explain why the latter needs confidence thresholds, human-in-the-loop review, and exception routing rather than silent auto-approval.
Know the vocabulary precisely: attended vs unattended bots, process discovery and task mining, document understanding pipelines (capture → classify → extract → validate → export), and the difference between accuracy, precision/recall, and straight-through processing (STP) rate. STP is the business metric; model accuracy is only an input to it.
Traps
- Assuming AI makes a process automatable end-to-end. Most failures are upstream data quality, not model quality.
- Confusing RPA (UI/API-level task execution) with ML (pattern inference). A question may describe a rules-based invoice match and dress it in AI language.
- Treating a confidence score as a probability of correctness without calibration.
- Forgetting that LLM output is non-deterministic and must be validated before it writes to a system of record.
Rule of thumb
If a competent human needs judgement on more than ~20% of cases, design for human-in-the-loop from day one - the bot handles the easy majority, queues the rest with context attached. Build the exception queue before you build the happy path.
Worked example
An invoice process: 70% arrive as structured EDI (deterministic, 99%+ STP), 25% as clean PDFs (IDP with a 0.90 confidence threshold, ~85% STP), 5% as scans/handwriting (route to review). Overall STP ≈ 0.70 + 0.25×0.85 = 91%. Quote 91%, not the model's 97% field accuracy.