Data control

# What may your AI coding tool send?

“The tool only sees the code” is not a complete description of an AI workflow. Before using it with client material, map what may become part of a request and where that information goes.

28 October 2026 · 5 min read

[← All articles](/blog)

[Start with the information route](#inputs)[Name the information that needs different treatment](#categories)[Separate the provider questions](#provider)[Check where a rule is applied](#boundary)[Treat the records as information too](#records)

## Start with the information route

Depending on the tool and its configuration, a request may contain the person’s prompt, selected files, retrieved context, conversation history or output from another tool. Identify the actual behaviour of the version and settings you plan to use.

Do not assume every file in a project is sent, or that only the text visible in a chat box is sent. Both assumptions can be wrong. Inspect the documented behaviour and test representative requests through the selected route.

## Name the information that needs different treatment

Client source code, credentials, personal data and business-confidential material may require different handling. Ownership and a client’s contractual conditions also matter. A generic “sensitive data” label is a starting point, not a complete operating rule.

Information | Question to resolve |

Project code | Which repository and files may the tool use? |

Credentials | How are secrets kept out of the permitted request path? |

Personal data | What purpose and data-protection arrangements apply? |

Client material | Which client conditions govern its use and recipients? |

Tool output | Could a log, command or retrieved document introduce excluded content? |

## Separate the provider questions

Ask who receives the information, why it is processed, what is retained and whether it may be used for model training. Those are separate questions. An answer about training alone does not describe the full processing arrangement.

When personal data is involved, GDPR includes requirements about purpose, lawful processing, data minimisation, processor arrangements and security. Have the relevant owners review the actual workflow rather than treating a model setting as the whole assessment.

[GDPR: official text](https://eur-lex.europa.eu/eli/reg/2016/679/oj)

## Check where a rule is applied

A written instruction tells a person or model what should happen. A configured technical check makes a decision at a particular connection. Establish which requests pass through that check and which routes remain outside it.

Test an allowed request, a request that should be refused and an interruption. Review both missed detections and unnecessary blocks. A control that prevents useful work without a clear reason is not a finished design.

## Treat the records as information too

Prompts, files and tool outputs can appear in retained records. Decide what the reviewer needs, who can read it and when it is deleted. A useful account may need a rule and decision without a full copy of every sensitive input.

Write the data route and record policy together. That gives the team a concrete basis for choosing an integration, adjusting the workflow and explaining its limits.

[Build an AI policy](/policy)

[Questions about deployment and data](/trust)

Keep exploring

[Use the workflow checklist](/guides)
