An AI data flow register: trace a document from input to deletion
Map how data moves through an AI task, from input and integrations to copies, access and closure.
Article language: English

In this article
A document goes into an AI tool. Where does it go next: an output file, a search index, logs or an employee's copy? A data flow register traces that path for a specific task. Record the source, each place the material reaches, who can access it and what happens to the copies afterwards. Without a map, saying "the data is in the system" is about as precise as saying lost keys are somewhere in the flat.
This is a working map of the technical flow. It should distinguish known facts from questions that still need answers. A list of purchased tools cannot replace it. One service may be used in several processes, each with different data and permissions.
What should one register entry describe?
Choose a specific task with a clear beginning and end. "We use chat at work" is too broad. "An employee summarises an approved document and saves the result in a team folder" lets you identify the material, the action and the result.
Record the product, plan and sign-in method. Establish whether the account belongs to the organisation, who administers it and which integrations are enabled. If the flow differs between teams, describe each variation separately. Do not treat access as uniform across the company. If a solution is described as local, check what actually works without an internet connection and which connections remain necessary for installation and updates. The guide to testing offline operation and remaining connections (in Polish) can help structure that check.
Give the entry a simple identifier and an owner. Instructions, tool inventories and change plans can then refer to it. Each time you update the entry, record the date and reason for the review.
Where does the data come from, and how much is read?
Identify the source: pasted text, a file, email, a drive, a form or a connection to another service. Record the categories of information and the scope needed for the task. Check file names, comments and metadata as well. They can travel with the content.
For an integration, distinguish the access available from the access actually used in a test. A connection may be allowed to read a whole folder even when someone intends to use one file. Record both, along with the evidence for each.
Trace the flow using test material or material approved for that purpose. Keep the steps so another person can reproduce the observation. Do not add confidential documents just to see what happens.
What is created besides the final result?
Look for conversation history, attachments, transcripts, logs and exports. The final document may be only one of several records. Note where a person later copies the result too, such as a task system or a message to another team. For images, trace the origin of AI material from source to publication (in Polish), including exported versions and their roles.
Microsoft's documentation for Copilot in Microsoft 365 describes interaction history and its administration separately from model training policies. This is an example of why a register should keep those questions separate. Check the data and privacy documentation for the product against the plan and configuration in use.
For each storage location, record who can read it and the applicable retention rule. If the answer is unknown, keep the question open and name the person responsible for finding out. Do not fill the gap with an assumption about a default setting. For diagnostic locations, use guidance on AI logs without excess data (in Polish) to establish a minimum activity record, so the log does not become another complete copy of the material.
What might a flow record look like?
This exercise concerns an internal meeting note. The input is an approved recording, the output is a summary, and the reference for review is a list of decisions confirmed by the meeting lead. The example is a template for the description. It makes no claim about a particular tool or real client.
- Purpose and owner: what the note is for and who is responsible for the process.
- Input: the recording, its storage location and permissions.
- Transfer: how the file is attached or which integration is used.
- Intermediate records: transcript, history and other confirmed copies.
- Output: the document, recipients and the person reviewing it.
- Closure: the rule for keeping or deleting each record.
- Evidence: the date checked, the source of the settings and open questions.
Separate automatic actions from manual ones. This can reveal where to narrow the scope or stop a transfer before it reaches another system.
How do you check that the flow has ended?
Define what closing the task means: approving the result, disconnecting an integration, removing a role or taking an agreed action on the files. Each operation can concern a different location. A document disappearing from a user's view should not be the only evidence used to describe the whole flow.
Before changing services, check whether you can export the material and end access. Record limits imposed by settings and supplier terms. A precise account helps plan the next steps without promising control over locations that have not been assessed.
Do not put passwords or key values in the register. Record the credential's role, its owner and where it is managed securely. If people or integrations change, the team can see what needs review.

How can the register guide a small improvement?
Choose one place where more material is shared than the task needs. Check whether you can limit the folder, disable an unnecessary connection or prepare the right input file. Before making a change, record the current state and agree how you will judge whether the task still works correctly.
After the change, repeat the test with the same approved material. Check whether the result remains useful and whether the scope of access has actually changed. Record evidence from the relevant service. Only then update the entry's status. A proposed restriction, a completed change and a confirmed effect are separate stages.
How do you distinguish a setting, a supplier's statement and a test result?
Add the type of evidence to each finding. Documentation describes product behaviour under specified conditions. A setting shows the state of a particular account or organisation. A controlled test shows the effect of one action. These sources complement one another, but they do not automatically verify the entire data lifecycle.
For example, an administrator sees that access to a folder is restricted. An operator opens one file and saves the result elsewhere. Checking only the input folder says nothing about who can receive the export. Each stage needs its own location, role and responsible person. If you cannot access a setting, record that as a limit of the assessment.
A record can use three simple labels: confirmed, needs checking and not applicable to this flow. Explain the last label too. A blank field does not show whether someone forgot it or actually ruled the stage out. The article on Shadow AI in a law firm (in Polish) discusses how to review unapproved accounts.
How can you test sharing without exposing a client's material?
Prepare a fictional document labelled as test material. Give it an easily recognised identifier, but no real case data. Reproduce the file attachment, creation of the result and a sharing attempt within the approved scope. Record who could open the result and whether sign-in was required.
The test should reflect different roles, not just administrator accounts. A user with broader access may miss an error that an ordinary recipient encounters. If another person needs to check a link, agree their participation and use only the fictional material. Do not create a public link to a real document as a shortcut for testing.
Afterwards, record which test copies were created and who will clean them up. The marker helps reconstruct the path observed. It does not prove that no other records exist elsewhere. Make the limits of the test clear to the person making the decision.
How do you describe deletion and the end of access without overpromising?
Distinguish disconnecting an integration from deleting a visible file, deleting user history and the rules for other records described by the supplier. Do not replace those fields with a single "deleted" status. For each action taken, record the date, the person who performed it and confirmation from the relevant place. For a question that depends on the supplier, identify the current document or response and its scope.
If the company requires a particular condition to be met at closure, check it before working with the material. A condition that cannot be confirmed should go to the process owner for a decision. Do not add a retention period from memory or from documentation for a different plan. Before handing over a solution, use the code and access handover checklist (in Polish).
How can the map support a decision to approve the process?
Compare the scope of access needed with the actual scope. If summarising one file requires access to an entire directory, record the difference and the alternatives considered. You might narrow the task, prepare a separate input or look for a different configuration. The fact that a tool works does not settle the choice.
Put the reasoning in the project decision log (in Polish). Then test whether the restricted option is useful using the guide to evaluation criteria for an AI pilot (in Polish). Approval should name the task, data, account and owner, as well as events that require another review. The map can then help beyond the first audit meeting.
For each storage location, record the retention period and the action needed to delete the information separately. The guide to retention and deletion in AI tools (in Polish) helps distinguish a product setting, a copy in an integration and material kept by an employee. That description is more useful than a single general claim that the data has been deleted.
When should you update the map?
Return to it when the plan, supplier, role, integration or type of material changes. Add that step to purchasing and pilot processes. A register stops helping when it describes a configuration from before several important changes.
Connect the map to the service inventory (in Polish), rules for AI agents (in Polish) and the materials prepared for an AI audit. The guide to preparing materials for an AI audit (in Polish) can help organise those materials. The Aeneia Audit service covers the review of sources and responsibilities.
Do not overlook ordinary spreadsheets between systems. A spreadsheet in an AI process (in Polish) can change the format of an identifier, retain a hidden column or contain a formula rather than a value. Record the actual input and output in the map, along with the person responsible for that stage.












