Start with the process, not AI
Document ten recent executions. Count active minutes, waiting, rework, exceptions and system changes. If nobody can define a correct result, the process is not ready for automation.
Volume
Occurrences by day, week and season.
Variability
Distinct paths and the percentage of exceptions.
Friction
Active time, waiting, application switching and corrections.
Risk
What happens when the system is wrong or unavailable.
Classify the work
Not everything needs an agent. The cheapest solution that meets the objective is usually the easiest to maintain.
Deterministic rule
Use conventional code when conditions are stable.
Extraction or classification
Use models for ambiguous input while validating critical fields.
Generation
Define criteria, examples, voice and risk-based review.
Agent
Reserve it for work involving decisions, tools and error recovery.
Calculate full value
Count quality, speed and released capacity as well as integration, inference, maintenance and supervision.
- ✓Baseline before building.
- ✓Estimated cost per run and per month.
- ✓Maximum acceptable human-review rate.
- ✓Value of fewer errors or faster responses.
- ✓A clear threshold to stop or redesign the pilot.
Design control and exit
The company should understand the flow, switch providers and recover its data. Separate business logic, model, connectors and observability.
- ✓Client-controlled data and credentials.
- ✓Fallback when a provider or model fails.
- ✓Traceable decisions without logging secrets.
- ✓Human approval before irreversible actions.
- ✓Delivered code, documentation and operating procedure.