What it does
Process Automation is the pattern for replacing manual, rules-plus-judgement work with a governed automation: IntraLLM workflows orchestrate the steps, models handle the judgement calls, tools perform the actions, and humans approve the consequential ones.
Typical shape:
Trigger/Upload → Classify → Extract fields → Validate → Approve → Act → Record
Use-cases
| Process | Before | With automation |
|---|---|---|
| Invoice capture | Keying totals and codes by hand | Classify → Extract → post to finance; exceptions only to a person |
| Employee onboarding | Checklist emails to five teams | One submission fans out tasks and drafts every document |
| Report production | Analyst copy-pastes into templates | Scheduled run reads data, writes the report, files it |
| Support triage | Manual ticket routing | Classify by topic and urgency, route, draft first response |
| Data cleansing | Spot checks in spreadsheets | Batch extract → Format Check → correct or flag |
Building it
- Map the process — list steps, decisions and systems involved. Mark every point where a human currently exercises judgement.
- Choose the engine — a scripted workflow for known steps; an agentic step where the path can't be pre-scripted.
- Connect the systems — expose each external system as a tool so the workflow can act. See Tools (Admin).
- Add quality gates — Format Check on model output; Data Extraction with field definitions.
- Insert approvals — a User Approve node before anything is sent, posted, or changed.
- Pilot — run on copies of real cases; compare accuracy and time against the manual process.
- Monitor — Workflow History for failures, Usage for cost.
Governance checklist
- Approvals before side effects
- Error path defined (who is told when a run fails?)
- Data retention and export policy for outputs
- Group permissions limit who can run and edit
- Guardrails configured for user-facing stages. See Guardrails