Sensitive data

Emails about someone's health, private life or beliefs do not belong in an AI prompt, nor in an analysis report. Loop & Roll reads every email before it is sent to an AI model and sets these contents aside. They are counted, by category, and never quoted.

What is filtered

The special categories of Article 9 of the GDPR, criminal convictions, and manifestly private life, in English and in French:

Category	For example
Health	sick leave, a medical certificate, a doctor's appointment, an illness, a pregnancy, a disability, mental health
Sex life and orientation	sexual orientation, gender identity, sex life
Racial or ethnic origin	someone's origin, racist remarks
Political opinions	activism, party membership, a declared vote
Religious or philosophical beliefs	religious practice, leave for a religious holiday, freemasonry
Trade union membership	union representative, membership, union dues
Genetic and biometric data	a DNA test, genetic testing
Criminal convictions and proceedings	a criminal record, police custody, a conviction
Private and intimate life	love letters, intimate photos, domestic violence, divorce and child custody

Personal identifiers — social security numbers, IBANs, card numbers — do not set an email aside: an invoice stays an invoice. They are masked before anything is sent ([redacted-iban]).

The filter looks for phrasings, not words. "Health insurance", "the project is in good health", "SSL certificate", "a Bluetooth speaker" or a job ad that says "regardless of religion or sexual orientation" are not filtered.

How it works

1. Local rules, on your instance, read the subject and content of each email. No network call, nothing leaves the instance. Each detection carries one or more categories and a confidence level.
2. Optionally, an AI check: a small model reads a cleaned-up excerpt of 420 characters at most — never the whole email, without addresses, phone numbers or identifiers. It is always the instance's own AI provider or the OVHcloud connection, with an economical model (gpt-oss-20b at OVHcloud by default), never Anthropic. It has three levels, see "The settings" below. An email the rules judge sensitive is never sent to the model for confirmation.

In the mailbox analysis

A sensitive email is left out of the analysis: it is not grouped, not quoted in the report, not sent to the AI, not offered as an example. The report only says how many emails were left out, by category — for example "4 emails left out as sensitive (Health 3, Trade union membership 1)". A sender whose emails were mostly left out does not appear among the top senders.

When the rules hesitate and the AI check does not settle it ("rules only" level), the analysis leaves the email out to be safe: one email fewer in a sample costs nothing.

The same applies to the mailbox reading tools of the assistant, and to the examples suggested for a categorisation prompt.

In workflows

The five AI nodes — Categorize, Extract, Summarize, Compose (AI) and Free instruction (AI) — check the triggering email and the content they are about to send. If it is sensitive, the node does not call the model and the run stops, with the code ai.sensitive_content and the detected category. Nothing is sent, neither to the model nor further down the workflow. The email stays in the mailbox, to be handled by hand. The run shows up in Activity › To do.

To route these emails somewhere instead of stopping, set the node's If the failure persists to Follow the Error output (see Errors and replay): the Error output receives the ai.sensitive_content code, and can lead to an approval or to a folder for manual handling.

The instructions you write in a node (categories, descriptions, reply guidelines) are not filtered: a "Sick leave" category in an HR triage sorts invoices as usual.

When the rules hesitate and the AI check does not settle it, a workflow lets the email through: stopping a business process on a harmless email would cost more than it protects.

The settings

Administration › Sensitive data, for administrators:

• Filter sensitive data — on by default. Turning it off is recorded in the audit log.

• AI check — an option, per instance, with three levels:

  1. Rules only (default): no call to an AI. The rules catch about 65 to 70% of phrasings they did not know;
  2. Rules + AI on doubts: the small model only settles the cases where the rules hesitate;
  3. Rules + AI on every message: mailbox analysis and audits send every message to the model, after cleaning, even when the rules found nothing. This is what raises recall. Workflows, evaluation and the assistant's tools stay at the "doubts" level at most: a model call per step or per chat message would add latency and cost on a path meant to be fast.

  The screen shows the estimated cost (a few cents per 1,000 messages at level 3, and a few more minutes for a large analysis) and the model used. These calls are metered separately (usage "sensitive-data check") and deducted from your AI credits when the connection comes from the platform. With no eligible AI connection (OVHcloud, OpenRouter, Ollama or OpenAI), the option cannot be enabled and the screen says so. If the model does not answer, a doubt is treated as sensitive; an email the rules judged clear stays clear. Every level change is written to the audit log.

• Workflow exceptions — a workflow that deliberately handles sensitive data, such as an HR triage, can be exempted. The organisation takes responsibility for it: the exempted workflow sends its content to the AI without filtering or masking, and every sensitive email it lets through is recorded in the audit log.

The page also shows what was filtered over the last 30 days, by category and by place (analysis, assistant, workflows, evaluation). Only numbers: never a subject, a sender or any content.

What is recorded

Every exclusion is written to the audit log with its category, its count and what was being analysed — never a word of the email. Changing the setting or the AI check level, granting or removing an exception are recorded too. On a managed instance, the usage report counts the filtered emails by category and place.

Limits

An automatic filter is never perfect, and Loop & Roll does not claim otherwise:

• it may miss a phrasing it does not know, or an allusion that only a human would understand;
• it may set aside a harmless email in the analysis when a clue is ambiguous;
• it reads English and French only;
• it reads text: a scanned sick note without text, an image, is not analysed;
• it protects what is sent to the AI and what reports show, not the mailbox itself: the webmail shows emails as they are;
• an exempted workflow is not filtered at all.

The filter greatly reduces the risk of sending private information to an AI. It does not replace the scope — to keep a whole sender or domain out of workflows, exclude it — nor the vigilance of the people who build workflows.