Label sensitive data once, at the door, and enforce it everywhere.
Classification in Data.co Control happens at ingest, before content is indexed or reaches any model. Labels travel with every field, record, and chunk, so the policy engine can decide what an identity may retrieve and where it may be sent without guessing from free text.
Labels and what policies typically do with them.
Label names are defaults. Your organization decides what each one means in policy.
| Label | Examples | Common policy use |
|---|---|---|
| PHI | Diagnoses, clinical notes, medical record numbers | Private models only; treating-relationship conditions |
| PII | Names with identifiers, SSNs, addresses, contact details | Restrict exports; private models only |
| Financial | Account numbers, balances, payment data, deal terms | Approval for bulk export; team-scoped access |
| Privileged | Attorney-client communications, legal work product | Matter-based access; never external endpoints |
| Custom | Trade secrets, deal codes, controlled technical data | Defined by your policies |
Detection proposes. People confirm.
No classifier is perfect, so labels have a status. Policies can treat proposed and confirmed labels differently.
- 1
Hint
Source context provides strong signals: a column named ssn, a SharePoint site for Legal, an EHR notes table.
- 2
Detect
Pattern and dictionary detectors scan content for identifiers and known formats as it is ingested.
- 3
Propose
Labels are attached with a proposed status and surfaced to the asset's owner in the catalog.
- 4
Confirm
Owners confirm, correct, or add labels. Confirmed labels become the basis for policy and audit.
Define what is sensitive in your business.
Custom labels and detection rules are code, versioned with your policies. Rules can be scoped to specific sources to reduce false positives.
label: deal_confidential
description: Material non-public deal information
detect:
- pattern: "\\bPRJ-[A-Z]{3}-\\d{4}\\b" # internal deal codes
- dictionary: classification/dicts/codenames.txt
- source_hint: { site: "/sites/CorpDev" }
scope:
sources: [sharepoint-corpdev, email-archive]
status_on_detect: proposed
reviewer: group:corpdev-data-ownersIllustrative example · final API surface documented in /docs
Fields, records, documents, and chunks.
Structured data is labeled at the field level, so a policy can allow an agent to query a claims table while excluding the diagnosis column. Documents are labeled at the document level and the label is inherited by every chunk in the retrieval index, with chunk-level detections added where they occur.
Because labels are stored with the indexed content, a policy change takes effect at the next request. You do not need to reclassify or re-index to tighten what a label permits.
Each signal has a different failure mode.
Classification combines several signals because no single one is reliable across every source. Knowing how each tends to fail tells you where owner review matters most.
| Method | Strong at | Tends to miss or over-flag |
|---|---|---|
| Source hints | Whole sources with a known purpose, such as an EHR notes table or a Legal site | Sensitive content stored somewhere unexpected, like PHI in a general file share |
| Pattern detectors | Identifiers with a fixed format, such as account or record numbers | Free-text descriptions; numbers that only look like identifiers |
| Dictionaries | Known code names, product names, or matter names | New terms not yet on the list |
| Custom rules | Business-specific categories scoped to the sources where they apply | Anything outside the sources a rule is scoped to |
| Owner confirmation | Context only the business has, such as a deal being non-public | Volume; owners need a focused queue, not every field |
Decide how to treat a label nobody has confirmed yet.
Every classification program trades false positives against false negatives. Over-labeling blocks legitimate work and trains users to route around controls. Under-labeling lets regulated content reach destinations it should not. Because labels in Control carry a status, policy can make that trade-off explicit instead of leaving it to the classifier.
A conservative posture treats any proposed sensitive label as if it were confirmed, so content is restricted from the moment it is detected and owners widen access by correcting labels. A permissive posture enforces only confirmed labels and relies on owners to keep up. Many teams use both: conservative for PHI, PII, and privileged content, permissive for custom business labels while their rules are being tuned. Either way the choice is written in policy, tested, and visible in review.
Questions buyers ask
Do you use an LLM to classify data?
Classification relies on source hints, pattern and dictionary detectors, custom rules, and owner confirmation. Whatever produces a proposed label, access decisions are made by the deterministic policy engine on confirmed metadata, never by a model at request time.
How accurate is classification?
It depends heavily on your data, which is why we do not quote a figure. Labels carry a proposed or confirmed status, owners review them, and conservative policies can treat proposed sensitive labels as if they were confirmed.
Can we import labels from an existing classification tool?
Where existing labels are available from the source, such as document sensitivity labels, they can be used as source hints. Specific integrations are confirmed during scoping.
What happens to content added after the initial scan?
New and changed content is classified during each sync before it is indexed, so it is never available for retrieval unlabeled.
Is classification output sent anywhere?
No. Detection runs inside your environment, and labels are stored in the local catalog and indexes.
Can labels be set for a whole dataset rather than field by field?
Yes. Dataset-level labels are often the right starting point for sources whose purpose is clear, with field-level labels added where a policy needs finer control, such as excluding one column from an otherwise permitted table.
See it against your own use case.
A 30-minute technical walkthrough, policy simulation on your scenarios, and a deployment recommendation for on-prem, private cloud, or hybrid.