Core•8–12 hours

AI Process Automation Map

Turn an as-is process into an evidence-backed automation map covering bottlenecks, controls, AI candidates, deterministic alternatives, human checkpoints, and measurable acceptance criteria.

process mappingautomation discoverycontrol designrequirements traceabilityAI capability fit

Scenario

Task

An operational process spans email, spreadsheets, and CRM, with manual checks, duplicate data entry, and several exception paths. The team proposes adding AI but lacks evidence for where a model creates value versus where deterministic workflow automation is enough. Build an as-is/to-be map, separate deterministic automation from probabilistic AI, preserve control points, and define how the result will be verified before rollout.

Step-by-step execution

1. Reconstruct the real as-is process

Outcome: The process map is evidence-backed and includes exceptions rather than only the happy path.

Tasks

→Collect 10–20 real process instances
→Mark actors, systems, queues, approvals, and rework
→Record cycle time, wait time, and error or rework signals
→Link critical steps to source documents or system events

Checks

✓Every high-impact step has an evidence source
✓Exceptions are not hidden in notes
✓Every manual check has an explicit purpose

2. Separate deterministic, AI-assisted, and agentic automation

Outcome: AI is used only where probabilistic capability is needed rather than as an expensive if/else.

Tasks

→Define at least one non-AI alternative for each bottleneck
→Assess input variability, ambiguity, reasoning needs, tool access, and failure impact
→Assign an authority class: read, recommend, draft, bounded action, or prohibited action
→Identify dependencies on data quality, retrieval, identity, permissions, and human review

Checks

✓Every AI candidate has a non-AI baseline
✓Model output is not treated as permission or an authoritative result
✓High-impact actions do not receive autonomy without control evidence

3. Build the to-be flow with controls and failure paths

Outcome: The future process describes validation, escalation, retry, and reconciliation, not just an AI stage.

Tasks

→Draw the normal path and at least three failure paths
→Add schema and field validation before domain logic
→Define preconditions, idempotency, and authoritative postconditions for write actions
→Define fallback for unavailable models, stale data, low confidence, or policy conflict

Checks

✓A timeout after a side effect enters reconciliation rather than blind retry
✓Uncertain or unsupported results have a safe exit
✓Human checkpoints are tied to risk and authority

4. Turn the map into a decision-ready experiment

Outcome: The team can make a go, iterate, or kill decision from measurable signals.

Tasks

→Create a representative evaluation set with happy paths, edge cases, and costly failures
→Record baselines and targets for quality, time, cost, and rework
→Design shadow, draft, and canary rollout stages
→Define rollback, owner, and a production-feedback-to-regression loop

Checks

✓Each success metric has a baseline and ground-truth source
✓A critical failure blocks rollout regardless of average score
✓Cost is measured per successful task including review and retry overhead

Acceptance criteria

The as-is map has evidence for critical steps, handoffs, exceptions, and controls
Every AI candidate is compared with a deterministic or non-AI baseline
The to-be flow includes authority, validation, escalation, retry or reconciliation, and fallback
A representative evaluation set and predefined go, iterate, or kill criteria exist
High-impact automation requires an explicit human or permission boundary and authoritative verification

Materials before execution

Human-in-the-loop for AI

A practical production architecture for human oversight: involve a person at a specific risk point with enough context for real, not ceremonial, control. Covers contracts, authority boundaries, failure modes, evaluation, and controlled rollout.