Cookie settings

Necessary, always on: they keep the site working and remember your cookie choice.

Skip to content
Skip to content

Glossary. Data classification.

What is data classification?

A match becomes a label when a check, its context and a named owner agree. Eight sample values, two of them rejected, show where each one decides.

By Dipal Panchal, co-founder and CEO · Updated November 7, 2026

Type, level and the person who confirmed it

Data classification is the practice of labeling data by what it is, such as a card number or a diagnosis, and by how sensitive your policy says it is. Unlike sensitive data discovery, which finds where data sits, classification decides what each found value is and which handling rules apply to it. A pattern match is only a candidate until a check, its context and a named owner make it a label.

A scheme names the levels and the handling each one gets. A label records which level a dataset, table, column or file belongs to, and who confirmed it. In ISO/IEC 27001:2022, the scheme is Annex A control 5.12 and the labeling is control 5.13.

How a value becomes a label

A pattern proposes a candidate, such as 16 digits that could be a payment card. A check then tests the format where the identifier has one, such as the Luhn check digit for cards, the mod-97 check for IBANs and Verhoeff for India’s Aadhaar numbers. Context settles what a check can’t: the column name, the neighboring columns and the system the table lives in.

Card numbers are issued with a Luhn check digit, and about one random 16-digit number in ten passes by chance, because that digit has ten possible values. In our test on November 7, 2026, 10,226 of 100,000 random 16-digit numbers passed. A column of real card numbers passes in nearly every row and a column of order numbers in about one row in ten. Judge a column by its pass share, and treat a single passing value as a candidate.

Names and addresses have no check digit. Name recognition and column names find them, which makes those labels the ones most worth a person’s review.

Eight sample values, kept or rejected

Sensitivity uses a four-level scheme (public, internal, confidential, restricted), and your policy’s names will differ. We computed every check with a Python script on November 7, 2026.

Records
SourceLocationDetected typeWhySensitivityDecision
CRM database, UScrm.customer.card_numberPayment card number4111 1111 1111 1111 passes Luhn, next to customer namesRestrictedKept; the data owner confirms the label
Sales databasesales.order.order_refNone (card candidate)2026 1107 0000 4181 fails Luhn: a date-prefixed order reference (not issued)NoneRejected: a look-alike, dropped from the card count
Billing databasebilling.payee.ibanBank account (IBAN)GB21 TEST 0000 0012 3456 78: length 22 for GB, mod-97 passes (not issued)ConfidentialKept; the data owner confirms the label
Billing databasebilling.payee.ibanBank account (IBAN)AE77 0000 0000 0012 3456 789: length 23 for AE, mod-97 passes (not issued)ConfidentialKept; the same column holds accounts from two countries
Newsletter platform exportmarketing.subscriber.emailEmail addressjordan.lee@example.com matches the format; email addresses have no check digitInternalKept: a subscriber list
Clinic scheduling databaseclinic.appointment.patient_emailEmail addressThe same address, in a table that shows the person is a patientRestrictedKept at the higher level the table sets
HR database, Indiahr.employee.aadhaar_noNone (Aadhaar candidate)2222 2222 2222 is a repeated-digit placeholder and fails Verhoeff (not issued)NoneRejected: a placeholder, not an identifier
CRM database, UScrm.person.full_namePerson name“Example Subject” found by name recognition and the column name; no check existsConfidentialKept for a person’s review

Why one email address gets two labels

The two email rows hold the same value at two levels. A subscriber list says little about a person, while a clinic’s appointment table says they are a patient, so the table, and not the address, sets the level.

The two rejected rows show the checks doing the volume work. The order reference fails Luhn, and the Aadhaar candidate is a run of one digit that also fails Verhoeff, so neither reaches an owner’s queue.

The card number is a published processor test number, and the email domain is reserved for documentation under RFC 2606. The IBANs, the order reference and the Aadhaar candidate were built to pass or fail their checks, and none was ever issued.

Our rule for turning a match into a label

Our rule has four conditions, applied in order.

  • A type with a check digit must pass it, value by value
  • A type without one takes its label from the table and its neighbors
  • A table the scan couldn’t read stays unclassified until it is read
  • A named owner confirms each label before it drives access

Where owner sign-off is worth the time

Confirming every label by hand doesn’t scale, and the rule doesn’t ask for it. Checks clear the look-alikes, as the two rejected rows show, and owners sign only the labels that decide who can open a table or how long it is kept.

A label that only feeds a report can wait for the next review. A label that grants a team access to a restricted table needs a named person who can say why it is right.

Where the law sets the category for you

Each law below defines sensitive categories of its own. We recommend mapping each category to a level of your scheme, in writing.

Where the law sets the category for you
WhereProvisionCategoryWhat it lists (abridged)
EU and UKGDPR Art. 9(1); UK GDPR Art. 9(1)Special categories of personal dataRace, ethnic origin, political opinions, beliefs, union membership, genetic, biometric, health, sex life, orientation
CaliforniaCal. Civ. Code § 1798.140(ae)Sensitive personal informationGovernment ID numbers, account log-ins with passwords, precise geolocation, citizenship or immigration status, health
US health45 CFR 160.103 (HIPAA)Protected health informationIdentifiable health information in any form; excludes FERPA records and records a covered entity holds as employer
UAEFederal Decree-Law 45 of 2021, Art. 1Sensitive personal dataFamily, ethnic origin, opinions, religious beliefs, criminal record, biometric, health and genetic data
Saudi ArabiaPDPL, Art. 1 (official English translation)Sensitive dataEthnic origin, beliefs, criminal record, biometric or genetic data, health, data showing a parent is unknown
IndiaIT Act s.43A; SPDI Rules 2011, rule 3Sensitive personal data or informationPasswords, financial and card details, health conditions and medical records, sexual orientation, biometrics
Payment cards (contract, worldwide)PCI DSS v4.0.1, requirement 3Account dataCardholder data (card number, name, expiry, service code) and sensitive authentication data (track, CVV, PIN)

India’s list, and the Gulf free zones

India’s list still comes from section 43A of the IT Act and rule 3 of the SPDI Rules 2011. The Digital Personal Data Protection Act 2023 has no separate sensitive category, and its section 44(2), which removes section 43A, commences 18 months after the notification of November 13, 2025. Until then the SPDI list applies. When those sections commence, the Act lists the volume and sensitivity of the data among the factors the government weighs before it notifies a Significant Data Fiduciary (s.10(1)(a)).

The UAE law has been in force since January 2, 2022, but its Executive Regulations had not been issued when this page was checked on November 7, 2026. Its Article 2 also leaves out health and banking data that have their own legislation. Firms in the DIFC and ADGM free zones answer to those zones’ own laws, the DIFC Data Protection Law No. 5 of 2020 and the ADGM Data Protection Regulations 2021. Where a column falls under two laws, we label it at the stricter level, and recommend you do the same.

How many levels a scheme needs

The UK government runs on three tiers, OFFICIAL, SECRET and TOP SECRET, under its Government Security Classifications Policy (version 2.0, August 5, 2024). US federal systems rate information by potential impact, low, moderate or high, under FIPS 199. NIST SP 800-53 control RA-2 then has agencies categorize each system and the information it processes, stores and transmits.

ISO/IEC 27001:2022 lists classification and labeling as Annex A controls 5.12 and 5.13, and doesn’t fix the number of levels. Use as few as you enforce differently: if two levels get identical handling, merge them.

What classification can’t tell you

A checksum proves a value is well formed. It doesn’t prove the value belongs to a real person, and a format-preserving token can pass the same check as the value it replaced.

A label also says nothing about who can reach the data or whether anyone uses it; access mapping answers that. And a scan covers what it read: a sample supports a decision about the tables it tested, and can’t show that the rest of the estate is clear.

With IRIS

IRIS classifies what it reads in the sources you connect with pattern rules, checksums and local name recognition. Each database finding stays pending until a reviewer decides.

Check digits tested value by value: IRIS produces part of this, for nine identifier types, such as cards (Luhn) and Aadhaar (Verhoeff).

Labels taken from the table and its neighbors: IRIS produces part of this, from column names and context; your policy sets the level.

Unread tables kept unclassified: IRIS produces this.

An owner confirms each label: IRIS produces part of this, a reviewer approves, reclassifies or declines each database finding; the owner’s sign-off stays yours.

Explore discovery and classification

Test your rules on a table you already know

Before you trust a classification run, point it at a column you know holds card numbers and one you know holds order references, using an account that can only read. Check at least 1,000 values from each and record how many were checked. Expect close to all 1,000 to pass Luhn in the card column and about 100 in the order column, and record the run date and rule version beside the findings.

Classification depends on the discovery step finding the tables first. Once columns are labeled, the next question is who can reach them: How to run a database access review.

Official sources

  1. ISO/IEC 27001:2022, Annex A controls 5.12 (classification) and 5.13 (labeling), ISO. Opened November 7, 2026.
  2. NIST SP 800-53 Rev. 5, control RA-2 Security categorization, NIST. Opened November 7, 2026.
  3. FIPS 199, Standards for security categorization, NIST. Opened November 7, 2026.
FAQ

Data classification questions

Protect your brand. Start with your data.

Bring your classification scheme. We’ll show you how IRIS findings line up with your levels.