Safe AI Launch

# Test the work. Test the controls.

An AI trial can produce a polished result without answering the questions that prevent people from using it on real work. Test the useful task and its controls as two separate outcomes.

28 October 2026 · 5 min read

[← All articles](/blog)

[Choose a result someone can judge](#result)[Write the rules before the demonstration](#rules)[Use more than the happy path](#cases)[Read the records as a reviewer would](#records)[Make two decisions at the end](#decision)

## Choose a result someone can judge

Start with work that already has an owner and a recognisable result. A small code change with a clear review process is more useful to assess than an instruction to transform the engineering function. Name the person who can decide whether the output is fit for its intended use.

Write down the current workflow, the intended improvement and the limits. If a person must still approve a release, that approval remains part of the task. The trial should not quietly turn a review step into an automatic action.

## Write the rules before the demonstration

List the information the AI tool may use, the routes it takes and the actions it may propose or execute. Some decisions belong to the business or client owner; others need a technical implementation. Keep those responsibilities visible.

For example, a team may allow an agent to prepare a patch using an approved repository while excluding credentials and a separate client dataset. Publishing the release may require a named maintainer. This is an illustrative scope, not a customer result.

## Use more than the happy path

A permitted task checks whether the workflow remains useful. A refused request checks the intended boundary. A protected action checks whether the correct person must intervene. An interrupted connection checks whether the team can explain what is known and what is not.

Case | Useful result |

Permitted task | Work can proceed and a reviewer can assess it |

Restricted information | The selected rule produces the expected decision |

Protected operation | The required approval is checked for that proposal |

Interruption | Missing or uncertain results are visible rather than silently treated as success |

## Read the records as a reviewer would

Ask someone who did not run the task to explain the recorded activity. Can they identify the tool, the rule, the decision and the reported outcome? Do they know which parts of the task were outside the connection?

Keeping every prompt in full is not the same as keeping the information needed for a review. Decide the purpose and necessary context, then set access and retention. A missing event may indicate missing collection rather than an event that never occurred.

## Make two decisions at the end

First decide whether the work was useful. Then decide whether the controls and records met the agreed requirements. Either can fail while the other succeeds. Record the gaps and the changes needed before expanding the task.

NIST’s AI risk-management resources support a wider process of defining responsibilities, evaluating risks and deciding what to manage. Our suggested test sequence is one practical way to make an initial workflow discussion concrete.

[NIST AI Risk Management Framework](https://www.nist.gov/itl/ai-risk-management-framework)

[NIST generative AI risk profile](https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence)

Keep exploring

[Use the planning worksheet](/guides)
