AA Advanced RPA cheat sheet

RPA bots execute deterministic, rule-based tasks via UI or API integration; AI models add probabilistic judgment (classification, extraction, prediction) where rules are insufficient.

Rule of thumb: automate a step with RPA when decision logic can be expressed as explicit if/then conditions; use AI when the input is unstructured (free text, images, speech) and labeled examples exist.

Supervised learning requires labeled training data; unsupervised learning finds structure without labels; reinforcement learning optimizes actions via reward signals - choose based on whether ground truth exists.

Precision = true positives / (true positives + false positives); recall = true positives / (true positives + false negatives). High-recall models suit triage; high-precision models suit auto-approval.

Confidence thresholds route AI output: above upper threshold auto-process, between thresholds send to human review, below lower threshold reject or escalate.

OCR converts images to text; document understanding combines OCR with layout and classification to extract fields from invoices, forms, and contracts.

Human-in-the-loop (HITL) design is required when AI output feeds a regulated or high-value decision; the bot must pause, queue for review, and resume with the reviewer's decision.

Straight-through processing (STP) rate = transactions completed without human touch / total transactions; it is the primary throughput KPI for attended-plus-unattended hybrids.

Process selection scoring typically weights volume, standardization, exception rate, cycle time, and regulatory risk; high-volume, low-variation, rules-based processes score highest for RPA.

Task mining observes desktop events to find candidate steps; process mining reconstructs actual flows from system event logs - combine both to size automation opportunity.

Exception handling pattern: retry transient failures with exponential backoff, then route persistent failures to a queue with structured error context for a human or a remediation bot.

Queue-based decoupling lets producer bots enqueue work items and consumer bots process them; this enables horizontal scaling and isolates failures.

Attended automation is triggered by a user session; unattended runs on a server on a schedule or trigger; hybrid models hand off between the two via queues.

Selectors should prefer stable attributes (automation IDs, accessibility names) over screen coordinates or index positions, which break on UI layout changes.

Idempotency is required for any bot that can retry: use a transaction ID or dedupe key so re-processing the same item does not duplicate downstream effects.

Cognitive automation pipelines chain components: ingest → OCR/extract → classify → validate against business rules → post to system of record → log confidence and outcome.

Layered bot architecture: entry point (trigger) → orchestrator (queue/workflow) → reusable component library (sub-processes) → connectors (UI, API, DB) → logging/telemetry.

Every bot needs a defined state model (queued, in-progress, succeeded, business exception, system exception) persisted outside the bot process so restarts are safe.

Configuration values (credentials, endpoints, thresholds) belong in a secure config store, never hardcoded; secrets go in a vault referenced by key.

Logging must capture timestamp, transaction ID, step, outcome, and error code; logs feed both operational dashboards and audit evidence.

Deployment lifecycle: dev → test/UAT → production, with versioned packages and a rollback path; promotion requires passing regression test suites.

Environment parity matters: screen resolution, locale, OS version, and application versions must match between test and production or selectors and timings break.

Reusable components should be unit-tested independently and versioned; a change to a shared component triggers regression across all consuming bots.

Orchestrator capabilities to rely on: queue management, credential vaulting, scheduling, bot health monitoring, and centralized audit logs.

Center of Excellence (CoE) operating models: centralized (CoE builds all), federated (business units build, CoE sets standards), and hybrid (CoE owns platform and reusable assets, BUs own pipelines).

Automation governance gates: intake/assessment → design review → build standards compliance → UAT sign-off → production release → periodic health review.

A bot inventory/registry with owner, criticality tier, dependencies, and last-reviewed date is the minimum control for managing a growing estate.

Change management control: any change to a production bot requires a change request, impact assessment on dependent bots, and re-test evidence.

Criticality tiers drive monitoring cadence and incident SLAs; tier-1 bots (customer-facing, financial) get real-time alerting and on-call coverage.

Scaling metrics to track: pipeline conversion rate, bots in production, STP rate, exception rate, hours saved, and cost per transaction.

Role separation: developers build, business analysts define requirements and process maps, and a control/risk function reviews for compliance - avoid one person owning all three for critical bots.

License and infrastructure capacity planning must account for concurrent bot execution, peak queue volumes, and orchestrator licensing before scaling the pipeline.