10.1 Classification Rules

Policy → Rules → Classification Rules holds the content rules. A rule looks at the text of a document or email and, when its conditions are met, takes one action:

Every enabled rule in scope is evaluated every time; there is no rule order and no “first match wins” for classification rules (see How the engine decides).

The rule list

Badges next to the name

Delete asks: “Are you sure you want to delete the rule”…“? To disable it instead, turn off the Enabled switch rather than using Delete.” Disabling keeps the rule (and its history in reports) and is the safer way to switch a rule off temporarily.

Adding a rule: blank or from a template

+ Add Default rule (or + Add rule to Group) opens the Add a rule drawer (“Start blank or from a template”).

• Blank rule — “Start from scratch with one empty condition.” The new rule is called New Rule, action Recommend, match All conditions (AND), one empty pattern condition, and the first label in your Labels list pre-selected.
• Template cards — each shows a category, name, description and a summary line such as 2 conditions · all (AND) · enforce. A template fills in the conditions, the match mode and a suggested action, and deliberately leaves the Label empty: you must pick one before you can save.

In both cases a rule created while a group is selected in Scope is assigned to that group.

The 25 built-in templates are written for organisations that process Turkish personal data under KVKK, plus international sets. Their names and keywords appear in Turkish in this release. By category:

Creating a rule step by step

1. Go to Policy → Rules → Classification Rules. Set Scope to Default for a rule everyone receives, or to a group for a group-only rule.
2. Click + Add Default rule (or + Add rule to Group) and choose Blank rule or a template.
3. In the drawer, enter a unique Policy name.
4. Choose the Action. For Recommend or Enforce, choose the Label and, if the label has sub-sensitivities, optionally a Sub-sensitivity. For Warn, write the Warning message.
5. Set Match (All / Any) and build the conditions: + Pattern condition, + Keyword condition, + AI document type. For each, set Min. count and Counting.
6. Optional: add Exceptions, Recipient exceptions (Outlook), restrict Modules, and set Group targeting (groups and exclude groups).
7. Click Save in the drawer footer. If Save is disabled, the footer says why (e.g. “Select a label (data class) to enable Save.”).
8. Paste sample text into Test a rule at the bottom of the page and click Run Test.
9. Click Publish & Sign in the top bar.

The rule drawer

The drawer title is the rule name (or New rule), subtitle “Edit classification rule”. Fields are edited in a local copy; nothing changes in the draft until you press Save. Closing the drawer discards the edits.

General settings

Actions in detail

Recommend produces a suggestion. Where it surfaces:

• the classifier window (save, send, ribbon Classify, right-click Classify) opens with the highest recommended label pre-selected, with the rule’s sub-sensitivity if set;
• the background file monitor adds the file to the user’s files waiting list and shows a notification (on live changes) — it never writes a recommended label to a file;
• the Ada web panel pre-selects it.

Enforce means “this content requires this label”. The exact effect depends on where the content is (see Where and when rules run): the file monitor labels silently or raises the label; classic Outlook labels silently, raises, or blocks; Office save blocks a save below the required label.

Mask (***) is an independent, irreversible content change:

• Every occurrence of every matched value is replaced by the same number of * characters (100 occurrences → 100 replacements). Counting affects only whether the rule fires (Min. count), never what is masked.
• With Match: Any, only the conditions that are satisfied are masked; with All, all conditions’ matches are masked.
• Exceptions apply. In Outlook, recipient exceptions apply (no masking when every recipient is excepted).
• No label is applied and mask rules never take part in label selection.
• Mask rules cannot contain an AI condition (“A masking rule cannot use an AI condition.”) — AI finds document types, not text spans.
• Desktop Office (Windows) and classic Outlook show a confirmation, Masking Will Be Applied, with a count per rule and a preview: “If you continue, the items below will be PERMANENTLY replaced with ***. This action cannot be undone.” Buttons Mask and Continue / cancel. Cancel stops the save or send.
• Outlook on the web, new Outlook and Outlook for Mac mask automatically at Send (no confirmation is possible there) and the send continues.
• The background file monitor never edits files for masking. It records Masking pending (user warned), lists the file in the user’s masking list and notifies on live changes; the user opens the file in Office and saves to mask it with confirmation.

Warn (user continues) shows a message you write. It never applies a label and combines freely with Recommend, Enforce and Mask rules on the same content.

When several Warn rules match at once, the user sees one window: the messages one after another, the strictest mode (Confirm beats Notify) and the strictest justification (Required > Optional > Off).

Where Warn appears:

Conditions

Pattern (regex) condition

Keyword condition

Comma-separated list fields. Keywords, supporting keywords, recipient exceptions and mail rule domains all use the same input: type entries separated by commas; while the field has focus you see exactly what you typed, and when you leave it the list is shown normalised (a, b, c). Empty entries are dropped. Domain and address lists are also trimmed, lower-cased, and a leading @ is removed.

Built-in validators

Validators are what make a pattern a reliable data type: the regex finds candidates, the validator rejects numbers that merely look right.

Patterns supplied by the templates. These are the expressions the templates use; you can copy them into your own rules.

National ID (use with TCKN checksum)   \b[1-9][0-9]{10}\b
Payment card (use with Luhn)           \b(?:[0-9]{13,19}|[0-9]{4}[ -][0-9]{4}[ -][0-9]{4}[ -][0-9]{1,7}|[0-9]{4}[ -][0-9]{6}[ -][0-9]{4,5})\b
IBAN, no spaces (use with IBAN mod-97) \b[A-Z]{2}[0-9]{2}[A-Z0-9]{11,30}\b
E-mail address                         [\w.+-]+@[\w-]+\.[\w.-]+
IPv4 address                           \b(?:[0-9]{1,3}\.){3}[0-9]{1,3}\b
Turkish mobile number                  \b(?:\+90|0)?5[0-9]{9}\b
Passport number (letter + 8 digits)    \b[A-Z][0-9]{8}\b
AWS access key                         \bAKIA[0-9A-Z]{16}\b
Private key block                      -----BEGIN (?:RSA |EC |OPENSSH |PGP )?PRIVATE KEY-----
10-digit tax number (no checksum)      \b[0-9]{10}\b

The payment-card expression accepts only real card groupings (contiguous 13–19 digits, 4-4-4-4, 4-6-5, 4-6-4) with a space or dash as separator. Looser patterns that allow a separator after every digit join unrelated numbers in spreadsheets and PDFs into “card numbers” that pass Luhn by chance.

Writing your own patterns

• The same expression runs in the console and web add-ins (JavaScript) and in the Windows/macOS agents (.NET). Use syntax both understand. Inline flags such as (?i) are rejected by the console’s check; use character classes ([Cc]onfidential) or a keyword condition instead.
• Prefer \b word boundaries and explicit lengths; avoid unbounded .*.
• Agents stop any single regex evaluation after 2 seconds and analyse at most the first 2,000,000 characters of a document’s text.
• A pattern that fails to compile on a client is skipped; it never stops the other rules.
• Example for IBANs written in blocks of four (verify with Test a rule before use): \b[A-Z]{2}[0-9]{2}(?: ?[A-Z0-9]{4}){2,7}(?: ?[A-Z0-9]{1,3})?\b with IBAN mod-97.

AI document type conditions

An AI document type condition is satisfied when the on-premises model is at least as sure as the threshold that the document is of a given type (payroll, CV, medical record, …). It involves no keywords. Document types, training and measured accuracy are covered in Chapter 11 (Policy → AI Classification); this section only lists the rule fields.

• AI conditions are evaluated by the Windows and macOS agents (file monitoring; on Windows also Office save/send and right-click Classify), and only when the AI module is on (Policy → Rules → Agent Rules → Modules, off by default). Web add-ins and the console’s Test a rule never satisfy an AI condition.
• If you set Enforce on an AI rule whose measured precision at the chosen threshold is below 95 %, never measured, or uses Document or a part, the drawer shows Enforcing with low measured accuracy. with the reason and “Run it in test mode first, or raise the threshold.” It is a warning; Save is still allowed.
• The quickest way to create AI rules is from Policy → AI Classification, which opens this drawer pre-filled.

Supporting keywords

Exceptions (exclude)

Exceptions are conditions that suppress the rule. If any exception is satisfied, the rule produces no finding, no warning and no masking — like an exception in a DLP policy. They use the same fields as conditions and are added with + Exception pattern, + Exception keyword and + Exception AI type. Exceptions are optional.

Typical uses: a test-data marker (TEST DATA – NOT REAL), a public template, or the organisation’s own published IBAN.

Recipient exceptions (Outlook)

Matching rules:

• Every recipient must be covered. A single outside recipient makes the rule run, so adding one trusted address to Bcc cannot switch enforcement off.
• Matching is exact: contoso.com does not cover sub.contoso.com or contoso.com.tr; a full address covers only that person.
• If any recipient address cannot be resolved to an SMTP address, no exception applies.
• Applies to every action: Recommend, Enforce, Mask and Warn. In classic Outlook it applies to email messages; meeting requests are evaluated without it.

Modules

Modules restricts where the rule is evaluated. Nothing ticked = the rule runs in all modules (“Runs in ALL modules. Tick one or more to restrict where this rule is evaluated.”). With ticks: “Runs only in the ticked modules (n of 11).”

Group targeting

Group rules are added to the Default rules a group receives; they do not replace them. To stop a Default rule in one group, exclude the group.

Saving: what blocks the drawer and what blocks publishing

How the engine decides

The same engine runs in the console tester, the web add-ins and the Windows/macOS agents. For each piece of content:

1. Skip rules that are disabled, rules whose label is disabled, Mask rules (handled separately), and rules whose Modules do not include the calling module.
2. Exceptions — if any exception is satisfied, the rule is skipped.
3. Conditions — each condition counts its matches (after validators and the counting mode) and is satisfied when the count reaches Min. count; AI conditions compare the model probability with the threshold. Match combines them (All / Any).
4. A matching rule produces a finding (Recommend/Enforce), a warning (Warn) or, in test mode, a test-mode match (audit only).
5. Outlook only: findings and warnings of rules whose recipient exceptions cover every recipient are dropped.
6. Recommended label = the highest label among all Recommend and Enforce findings. “Highest” is the label’s position on the Labels ladder (the bottom of the Labels list is the highest). List order of the rules plays no part.
7. Recommended sub-sensitivity = the sub-sensitivity of the most confident finding that recommends that same label (if it names one).
8. Required label = the highest label among Enforce findings only. Recommend findings never restrict the user.
9. Mask rules are evaluated separately with the same exception/condition/module logic and produce the list of values to replace.

Confidence. Each finding carries a confidence from 0 to 100, shown in Test a rule, the classifier and the audit. It starts at 50, or 75 when a satisfied pattern condition has a validator; +15 when a keyword condition is satisfied; +5 when the rule has 3 or more matches; capped at 100. For AI conditions the model probability (×100) is used when higher. Confidence decides between sub-sensitivities (step 7); it does not change which label wins.

Enforce is applied in two strengths. On send in Outlook (classic and web) the required label is pinned exactly: a lower label is raised (classic) or blocked (web), and a different label of equal or higher rank is blocked with “Sensitive data detected in the content requires exactly the”…” label.” Everywhere else (Office save, right-click Classify, the web panel, the file monitor) Enforce is a floor: the user may choose that label or any higher one, and the file monitor only ever raises, never lowers.

Where and when rules run

AI conditions are satisfied only on the agents (file monitoring; on Windows also Office save/send and right-click Classify) and only with the AI module on.

Test a rule

The Test a rule card at the bottom of Classification Rules runs the clients’ engine against the draft, including rules you have not published. The text stays in your browser; nothing is sent to the server.

1. Paste or type sample text in the box (placeholder: “E.g. Employee national ID (TCKN) 12345678950, IBAN TR33 0006 1005 1978 6457 8413 26…”).
2. Click Run Test (enabled once there is text).
3. Read the result:
   • green box: “n rules matched:”, one line per rule — IBAN × 3 — confidence 80%, (enforced) for Enforce rules; Warn rules as … — warning (confirm, justification optional): “message”; and Recommended label: …;
   • red box: “No findings — no rule matched this text.”

What the tester does not do:

Worked examples

Example 1 — Payment card numbers → Confidential, enforce

1. + Add Default rule → Blank rule. Name: Payment card numbers.
2. Action Enforce, Label Confidential.
3. Match All conditions (AND). Pattern condition: the payment-card expression above, Validator Luhn, Min. count 1, Counting Unique matches.
4. Optional second condition to reduce false positives: Keyword card, visa, mastercard, cvv, expiry — with Match All the number must appear together with card context.
5. Save, test with 4111 1111 1111 1111, publish.

Effect: a file saved with a valid card number is labeled Confidential by the file monitor (or raised to it); Word refuses to save it as Internal or Public when label-on-save is required; an Outlook email containing one is labeled Confidential on send.

Example 2 — IBAN count ≥ 5 → Strictly Confidential

A single IBAN on an invoice is normal; a list of IBANs is a customer or payroll file.

1. New rule IBAN list, Action Enforce, Label Strictly Confidential.
2. Pattern condition with the IBAN expression, Validator IBAN mod-97, Min. count 5, Counting Unique matches (the same IBAN repeated five times counts once).
3. Optional: a second rule IBAN with Min. count 1, Action Recommend, Label Confidential.

With both rules a document with three IBANs gets Confidential recommended; one with five different IBANs requires Strictly Confidential — the higher label wins.

Example 3 — HR file: keyword AND valid national ID

Match All: Keyword payroll, employment contract, personnel file + Pattern national ID with TCKN checksum. Action Enforce, Label Confidential, Sub-sensitivity Personal data (if defined). Add an exception keyword SAMPLE – NOT REAL for training material.

Example 4 — Warn before personal data leaves

Action Warn, message “This content looks like it contains personal data. Make sure the recipient is authorised before you continue.”, Mode Confirm, Justification Required. Conditions: Any of national ID (TCKN checksum), IBAN (IBAN mod-97). Recipient exceptions: contoso.com, contoso.com.tr. Internal mail is not interrupted; any external recipient triggers the warning and the justification is audited.

Example 5 — Mask card numbers in email only

Action Mask (***), Pattern payment card with Luhn, Modules: Outlook and Web add-in · Outlook. Classic Outlook asks Mask and Continue; Outlook on the web masks automatically. Documents are not touched.

Audit events written by classification rules