.svg){kind=icon}
Learn
Automated data classification is the use of AI to detect sensitive content and personally identifiable information in defense data and propose classification and control markings - such as CUI or SECRET - for authorized human review.
Last updated: October 18, 2026
Why it matters to your mission. Hand-marking every document for sensitivity is slow and inconsistent, and a single missed marking is a spill. Letting AI detect sensitive content and PII and propose markings for authorized human review changes the economics. The outcome is faster, more consistent classification at scale, so data reaches the people who need it without loosening control.
Automated data classification is the use of AI to detect sensitive content and personally identifiable information in defense data and propose classification and control markings - such as CUI or SECRET - for authorized human review. It does not replace the classification decision. A person with original or derivative classification authority still decides what a document is and how it must be marked; the AI does the reading, flagging, and proposing that used to consume an analyst's day, and it logs every proposal so the result is auditable.
Classification and marking are policy-governed and human-authoritative. What automation changes is the cost and consistency of the work, not who holds the authority. Read "automated classification" as "AI-assisted classification with a human in authority," and the rest of this page follows.
Key takeaways
Mission data arrives faster than people can hand-mark it. A single ingest can bring in thousands of documents, messages, images, and records, each of which may contain PII, a sensitive selector, or content that is classified by reason of its subject. Marking all of it by hand is slow, and consistency erodes as volume climbs, because two reviewers rarely apply the same rule the same way under time pressure.
Inconsistent marking creates decision risk in both directions. Over-restrict, and analysts cannot reach information they are cleared for, so a cleared picture sits unused. Under-restrict, and sensitive content moves where it should not, which is a spill. Automated classification is decision-support aimed squarely at that gap: it keeps marking in pace with ingest while keeping a human in authority over every decision.
The government has made the stakes explicit. The U.S. Government Accountability Office found that across roughly 2,900 DoD systems handling Controlled Unclassified Information, none of the components reviewed were fully compliant with the department's own protection requirements, against a standard of 100 percent compliance (U.S. Government Accountability Office, Defense Cybersecurity: Protecting Controlled Unclassified Information Systems, GAO-22-105259, 2022). Getting the marking right at scale, consistently, is a precondition for protecting the data correctly downstream.
The two words are often used together, but they name different acts, and automation touches them differently.
Classification is the determination that information requires protection and at what level: whether a document is Controlled Unclassified Information, CONFIDENTIAL, SECRET, or higher. In the national security context this is a human authority's decision, governed by executive order and by DoD Manual 5200.01 Volume 2 for classified information.
Marking is applying the standardized labels, banners, portion markings, and dissemination controls that carry the classification decision with the document. For CUI, the marking rules are set in DoD Manual 5200.48 and in the CUI Registry maintained by the National Archives' Information Security Oversight Office (see the CUI Registry).
| Dimension | Classification | Marking |
|---|---|---|
| What it decides | Whether and at what level information is protected | How that decision is written onto the document |
| Who holds authority | Original or derivative classification authority (a person) | Applied per policy once the level is decided |
| Governing policy | Executive order; DoD Manual 5200.01 Vol 2; CUI authorities | DoD Manual 5200.01 Vol 2; DoD Manual 5200.48; CUI Registry |
| Where AI helps | Proposes a candidate level for human decision | Applies marking rules uniformly at scale |
| Failure mode | Wrong level: over- or under-classification | Right level, wrong or missing banner or portion marking |
AI can propose a candidate classification and pre-apply the corresponding markings, but the authority still owns the classification decision, and a reviewer confirms the markings before the document is treated as marked.
The design principle is that AI is decision-support, not decision authority. In practice the flow is:
The authority gate in step four is what keeps the AI from classifying on its own. The AI narrows the work from a blank page to a reviewed proposal; the human still decides. Only the designating agency may apply certain controls at all: the CUI Registry states that "only the designating agency may apply limited dissemination controls to CUI" and that authorized holders may apply those markings "only with the approval of the designating agency" (National Archives, CUI Registry: Limited Dissemination Controls). An automated proposal respects that line by routing to the right authority rather than asserting the control itself.
Personally identifiable information - names, identifiers, contact details, and the combinations that single out an individual - is one of the most common reasons unclassified data has to be controlled. The CUI program exists in part to standardize how the executive branch handles exactly this kind of sensitive unclassified information: the National Archives describes CUI as information that "requires safeguarding or dissemination controls pursuant to and consistent with law, Federal regulations, and Government-wide policies" (National Archives, Controlled Unclassified Information).
Detecting PII reliably across free text, scanned documents, and records is a reading problem before it is a policy problem, and it is where automation earns its place. A model can flag PII and privacy-category content at ingest, map it to the correct CUI category, and propose the handling control, so the reviewer confirms a finding rather than hunts for it line by line. Privacy is one of the eighteen organizational groupings in the CUI Registry's category list, which is why PII detection and CUI marking are two halves of the same workflow rather than separate tasks.
Data-centric security attaches protection to the data itself - its markings and handling rules - rather than relying only on the boundary of a network. When a marking travels with the data, each environment that receives it can apply the right handling, instead of trusting that the perimeter already filtered everything correctly. Automated marking at ingest is what makes this practical at volume: the protection is written onto the data at the first opportunity, not bolted on later.
Build-low-enrich-up means ingesting and enriching data at a lower classification level first, then moving validated, correctly marked data upward. Classification markings proposed at ingest travel with the data, so each level can apply the handling its policy requires. Movement between security domains, though, is not a classification function and not something an AI performs: it requires an accredited cross domain solution governed by DoD Instruction 8540.01. Keeping that boundary clear - the AI marks, an accredited solution moves - is part of what makes the overall design accreditable.
Torch.AI builds government-owned reasoning infrastructure that treats classification as decision-support with a human in authority. The relevant capabilities map cleanly onto the flow above:
Because the reasoning infrastructure is government-owned, the customer keeps control of the detection logic, the proposals, and the audit record, rather than renting them inside a vendor's proprietary environment. Markings and enrichment produced by NEXUS stay attached to the data as ORCUS makes it available, so secure sharing respects the classification each item carries, and every proposal remains inspectable. You can see how this is packaged on the Torch.AI software page.
For a defense or intelligence evaluation, require that the capability:
Evaluating a capability? Bring these seven requirements to a scoping call and we will walk each one against your environment and a sample of your data: request a technical walkthrough.
| Function | AI decision-support (NEXUS) | Human authority |
|---|---|---|
| PII detection | Detects and flags PII and sensitive indicators at ingest | Confirms findings and handling |
| Classification proposal | Proposes candidate markings (for example CUI, SECRET) against policy | Classification authority decides and marks |
| Marking consistency | Applies marking rules uniformly across large volumes | Reviews and corrects as needed |
| Audit trail | Logs every proposal, rule, and reviewer action | Owns the record of decision |
| Cross-domain movement | Does not move data between domains | Uses an accredited cross domain solution |
Classification is the determination that information requires protection and at what level; marking is applying the standardized labels, banners, and dissemination controls that carry that decision with the document. Classification is a human authority's decision; marking applies the policy rules once the level is set.
It reads data at ingest, detects PII and sensitive indicators, and proposes candidate markings against policy, then routes each proposal to a classification authority who accepts, edits, or rejects it. Nothing is treated as marked until a human confirms, and every proposal and action is logged.
PII detection finds personally identifiable information across text, documents, and records. Because PII is one of the most common reasons unclassified data must be controlled as CUI, detecting it at ingest and mapping it to the right CUI category is part of the same workflow as proposing a marking.
Data-centric security attaches protection to the data itself, its markings and handling rules, rather than relying only on a network boundary. Build-low-enrich-up means enriching and marking data at a lower level first, then moving validated, correctly marked data upward, with markings traveling with the data.
No. Classification and marking are separate from cross-domain movement. Moving data between security domains requires an accredited cross domain solution governed by DoD Instruction 8540.01, not a classification model.
Describe the problem.
Define the constraints. We will structure the session.
No demo.
No slide deck.
Real work.
Describe the problem.
Define the constraints. We will structure the session.
No demo.
No slide deck.
Real work.
