An AI pilot is ready when its task, data, permissions, human checks and stop condition can be described and tested. Before connecting a model to email, documents or business software, choose a narrow workflow: what goes in, what comes out, which rules are deterministic and which cases need a person.
Map inputs, decisions and systems
Describe the current process from a trigger to an observable result. For example, an email could be classified, enriched with authorised CRM information and turned into a draft for an employee to approve. Record input formats, manual steps, exceptions, source of truth and destination. Use code for stable rules such as required fields; use a model only where language interpretation adds value.
- Trigger and scope: the event that starts the flow and records that are excluded.
- Data contract: expected fields, types, size limits and incomplete-input behaviour.
- Roles: process owner, technical administrator, reviewer and person who can stop the system.
- Allowed output: draft, classification or update; specify actions that are not permitted.
- Exceptions: abstention, escalation and a return to the manual process.
Keep the model between controlled steps
A reviewable flow validates schema and permissions, minimises data, calls the model, validates its response and then pauses for review before an authorised write. Structured output can reduce formatting ambiguity, but the application must still validate types, allowed fields and business rules. Valid JSON does not prove that an answer is correct.
Test errors, uncertainty and stop conditions
Prepare representative examples with an expected result or evaluation rule. Include normal, ambiguous, incomplete, duplicate, out-of-scope and deliberately hostile input. Test prompt injection in user text and documents. Unexpected output, timeouts and provider outages should stop safely rather than cause a partial write.
Isolate the pilot and measure the whole process
Start in a separate environment or review queue with dedicated credentials and minimised, authorised data. Version the prompt, schema, model and integration code. Set request limits, timeouts, retries, a spending cap and a way to stop new runs while completing open cases manually. Record a baseline for handling time, corrections, exceptions and review effort; include model, service, maintenance and human-review costs before deciding whether to continue.
Make human review and ownership visible
Show the reviewer the approved input, generated draft, available sources, checks and any reason the model abstained. They must be able to correct, reject or reassign the case. Record who can disable the connector, how interrupted work is recovered and what manual procedure takes over. The NIST AI RMF organises ongoing work around Govern, Map, Measure and Manage.
Set continuation, change and stop criteria before the pilot. A narrow workflow with clear controls is easier to evaluate than a broad promise to automate a whole department.