Gen8

What data does an AI workflow process?

Knowledge

What data does an AI workflow process?

The question “what happens to our data?” always comes, usually at the worst moment. Eight questions answered beforehand on one page turn that security review into a formality.

Why beforehand, and why on one page

An AI workflow almost always touches systems with different owners: the user's email, sales' CRM, IT's documents. Until someone writes down the whole picture, everyone answers for their own part and nobody can approve the whole. That is why projects get stuck in review — not because the answer is bad, but because it is not written anywhere.

The eight questions

1. What data enters the workflow?

Name it concretely for each step: message text and attachments, a contact's name and role, amounts, a transcript. "Client data" is not an answer. Also note what you deliberately exclude — that is the sentence that shortens a review fastest.

2. Does it include special categories of personal data?

Health, sickness absence, origin, memberships, criminal-offense data. If the answer is yes, that is not a detail but a redesign: that data usually belongs outside the workflow, or the workflow needs a separate, narrower setup.

3. Which systems are read, and with what permissions?

For each system: which part, on whose behalf, and read-only or write access too. The distinction between "reads the employee's mailbox" and "reads one shared mailbox" determines the entire discussion.

4. What is written, and can it be reversed?

Every write action, with its target system, control level (read-only, human approval, or automatic with a log) and whether the previous value is retained. A write action without a way back should not be automatic.

5. Where does processing take place?

For each component: where the workflow runs, where the model is hosted, where the logs reside. For many organizations, processing in the EU is a firm requirement, and that requirement applies to the chain, not one link.

6. What is retained, and for how long?

Distinguish three things often lumped together: input (the message, the document), intermediate results (extracted fields, transcriptions) and the log of what the workflow did. They deserve different retention periods, and usually the middle category should be kept for the shortest time.

7. Who can access it?

Who sees proposals, who sees logs, who can change the workflow. And: what does the supplier see? That last question deserves a concrete answer, not a reference to a policy document.

8. What happens on departure?

Where the workflow, credentials and logs reside when the collaboration ends, and what is deleted. Beforehand this is a paragraph; afterward it is a negotiation.

What this page does not fill in for you

The answers depend on the systems and suppliers you choose. Request them and write them down literally; an answer you cannot explain back is not an answer.

The gain

These eight questions take a morning and move the conversation with IT and management to the beginning, when changes are still cheap. They also regularly lead to a simpler design: if the answer to question 1 gets too long, the workflow sees too much — and seeing less often also means working better.

Next step

What to do with this

Answer the eight questions for the workflow currently on the table on one page, and send it to the person who will eventually conduct the security review anyway. The question you cannot resolve is exactly the part that needs to change in your design. What we document on our side about processing, permissions and exit is covered under security and data.

Frequently asked questions

Is this the same as a DPIA?

No, but it is the information you need to perform one. Whether a formal assessment is required depends on the data and the purpose; this checklist is the practical groundwork.

Who should answer these questions?

The person designing the workflow, together with the owner of each affected system. Leaving the answers to IT alone produces a technically correct account that does not match the process.

What if an answer is unknown?

Write it down explicitly as unknown, with a name attached. A visible gap is manageable; a gap covered with vague wording returns when something goes wrong.

Want to apply this to your own process?

Bring one process to a conversation. We assess suitability, risk and the cost of the first step.

  • Reply within one business day
  • No obligation