WAGONER ENTERPRISES FEDERAL AI ASSURANCE
DATA & POLICY AUTOMATION

ClassifyAI

Classify with confidence. Keep the data inside your boundary.

ClassifyAI recommends classification and CUI markings from a document's content and the program's handling policy, then routes every recommendation to a cleared reviewer. Each decision is kept as signed, tamper-evident evidence, and no content leaves the boundary for inference.

Request Technical Walkthrough See the review screen
Self-hosted Air-gap capable No external inference Signed audit trail
Ideal buyer Classification authorities, security managers, ISSMs Programs handling CUI or classified-adjacent content
Primary output Reviewed decision with a signed record Decision records, release packets, evidence packs
Deployment Self-hosted on containers or Kubernetes Connected, air-gapped, or split network
Standards context CUI · CMMC · OSCAL 32 CFR 2002, SP 800-171, SP 800-53
FOR THE EXECUTIVE

Consistent markings you can defend, without moving the data.

Classification becomes a reviewed, recorded process, and sensitive content stays inside the accreditation boundary throughout.

The outcome

Faster, more consistent marking, with the reasoning attached.

Analysts start from a recommendation instead of a blank page. Each decision carries the rationale, the policy version that applied, and the person who approved it, so an inspector's question has a written answer.

The risk it retires

Spills, over-classification, and assessments without evidence.

Marking is usually manual, and programs rarely keep a record of which rule or reviewer produced a marking. A consistent process reduces spills and over-marking, and the signed trail is the evidence an assessor asks for.

Why now

CMMC assessments and CUI designation are arriving together.

Third-party CMMC assessments under Phase 2 start in November 2026, and agencies are being pressed to tell contractors explicitly what is CUI. Most commercial classifiers assume content can leave the boundary, which rules them out for these holdings.

HOW IT WORKS

The system proposes. A cleared person decides.

Every document follows the same path inside the deployment boundary, and each stage leaves a record you can inspect.

1

Ingest and screen

Documents are fingerprinted, checked for file-type spoofing, and scanned for malware with a check that fails closed. Text is extracted, with OCR for scanned pages and images.

2

Detect

Existing banners, portion marks, dissemination controls, and CUI markings are recognized and validated. Personal identifiers are flagged.

3

Recommend

A proposed marking comes with a readable rationale and the version of the classifier that produced it.

4

Evaluate against policy

Versioned policy as code checks the recommendation, and the signed decision records which policy version applied. With no policy loaded, evaluation fails closed.

5

Adjudicate

A cleared reviewer approves or rejects with a written rationale. Separation of duties is bound to the signed-in identity, and privileged actions need step-up authentication.

6

Release and prove

Approved decisions feed two-person controlled release. Every step lands in a signed audit trail that exports as an evidence pack, a SIEM feed, or an OSCAL security plan.

REFERENCE ARCHITECTURE

Content stays in. Only signed outputs leave.

Handling policy and software come in as signed inputs, and inference happens inside the boundary with no external calls. What leaves is limited to signed release packets, signed sync bundles, and audit exports.

ClassifyAI reference architecture Outside a self-hosted accreditation boundary sit the sources and authorities: users and the identity provider, imported documents and scans, handling policy and software delivered as signed offline bundles, and imported threat-intelligence data. Inside the boundary, an authenticated access edge and reviewer console fronts a classification pipeline. Intake and screening, extraction and OCR, and marking detection feed a recommendation step that produces a proposed marking with a readable rationale and classifier provenance. A policy decision point evaluates the recommendation with versioned rules and produces a signed decision record. Every decision goes to human review by a cleared reviewer under separation of duties. Approved decisions flow to label translation and to controlled release, and threat correlation runs alongside. A trust and evidence fabric spanning the pipeline keeps a signed, tamper-evident audit trail and produces evidence packs, OSCAL and SIEM exports. Only signed release packets, signed sync bundles exchanged with peer enclaves, and SIEM, OSCAL and evidence outputs leave the boundary. Raw content for inference, calls to hosted AI models, and private signing keys never cross. ACCREDITATION BOUNDARY: NO CONTENT LEAVES FOR INFERENCE SOURCES AND AUTHORITIES NEVER CROSSES LEAVES THE BOUNDARY CLASSIFICATION PIPELINE GOVERNANCE Users and identity SSO, MFA, smart card Documents imported files and scans Policy and software handling policy as code signed offline bundles Threat-intel feeds imported indicator data Access edge and reviewer console single authenticated entry: federation, MFA, step-up, rate limiting Intake and screening integrity, file type, malware Extraction and OCR text and scanned pages Marking detection banners, CUI, identifiers Policy decision point versioned rules as code fails closed, version recorded Recommend proposed marking, readable rationale classifier version and provenance Signed decision record policy and classifier provenance Human review, every decision cleared reviewer, separation of duties Label translation national, allied, and local taxonomies Threat correlation indicators, cases, signed case evidence Controlled release two-person, redaction, signed packet Trust and evidence fabric signed, tamper-evident audit: every step above emits an event evidence packs, OSCAL SSP, SIEM export, segregated signing keys each record verifies independently; unsigned evidence is rejected Raw content for inference Calls to hosted AI models Private signing keys Release packet signed, redacted Peer enclaves signed sync bundles SOC and assessors SIEM, OSCAL, evidence packs SINGLE GUARANTEE content stays in the enclave, every recommendation is policy evaluated and human reviewed, only signed outputs leave
Figure 1. Reference architecture, public-safe. Conceptual view only; classification logic, policy contents, thresholds, and interface design are not depicted.

Open the full-resolution diagram

THE PRODUCT

Every recommendation waits for a person.

A reviewer cannot approve or reject until a written rationale is entered, and that rationale becomes part of the signed record.

ClassifyAI · Review Center
ClassifyAI Review Center in a dark navy theme with a purple CUI banner at the top and bottom. A sidebar lists Operations, Governance, Audit and Compliance, and Administration. Tiles count the decisions and workflows awaiting review. A table lists open decisions with their type, the proposed change shown as two classification badges such as CUI to CUI//SP, and a confidence value. The selected decision opens a review panel with the system's rationale and a reviewer's typed concurrence, which enables the Approve and Reject buttons.
Figure 2. Review Center. Open decisions with the proposed change, confidence, and rationale. Sanitized screen with representative, synthetic data.
ClassifyAI · Document detail
ClassifyAI document detail page for a synthetic logistics plan annex marked CUI and in Reviewing status. A status timeline shows Ingested, Parsed, Analyzed and Reviewing. A recommendation card proposes CUI//SP with its rationale. An extracted entities list shows classification markers, a CUI category, a masked identifier, an example email address, and a placeholder name. A decision chain lists an earlier reviewer decision and the open reclassification.
Figure 3. Document detail. The evidence behind a recommendation and the document's decision history. Sanitized screen with representative, synthetic data.
WHAT SETS IT APART

What makes the classification defensible.

No external inference

Extraction, detection, recommendation, and policy evaluation all run inside the deployment boundary. Content never has to cross the accreditation perimeter to be classified.

Policy as code with provenance

Handling rules are versioned and tested as code, and a built-in editor validates and activates new versions. Each decision records the policy version and the classifier that produced it.

Accountable human review

Every decision opens a review, and separation of duties is tied to the signed-in user, with CAC and PIV supported. Clearance is enforced on reads, so an uncleared viewer cannot open a document recommended at a higher level.

Controlled release

Release packets need two-person approval and recipient checks. Their classification comes from the documents themselves, never from the requester.

Signed evidence, ready for assessors

The audit trail, decision records, and evidence packs are signed and verify independently. The same record feeds a SIEM and an OSCAL System Security Plan for ATO and CMMC preparation.

Built for disconnected enclaves

The platform installs from an offline bundle. Separated enclaves stay aligned through signed sync bundles that are checked against a per-peer trust store and quarantined on failure.

HOW IT COMPARES

A different job from data security or marking add-ins.

ClassifyAI recommends like an assistant, governs like a policy engine, and records like an evidence system. Established categories are ahead on personal-data detection and native label integration.

Category-level comparison based on typical capabilities in each category; individual products vary.
Capability ClassifyAI Cloud data-security platforms User-driven marking tools Cross-domain guards Declassification AI platforms
Fully self-hosted or air-gappedYesMostly SaaSYesYesYes
Recognizes classification and CUI markingsYesPartialOwn labelsYesYes
Policy as code with decision provenanceYesRule or ML settingsNoRule setsVaries
Human review with separation of dutiesEvery decisionLimitedUser decidesNoYes
Two-person controlled releaseYes, signed packetNoNoTransfer rulesVaries
Signed audit and evidence packsYesLogsLogsLogs and attestationsVaries
WHERE IT LANDS FIRST

Start read-only. Add governance as it proves out.

The first deployment changes nothing about how the program works today.

Entry use case

CUI marking assist for a CMMC enclave

Recognize CUI markings and identifiers, recommend markings, and keep the signed trail an assessor asks for.

Entry use case

Consistency review of existing holdings

Run recommendations read-only against a sanitized corpus. Measure analyst time per document and how often reviewers agree.

Entry use case

Release governance

Two-person release with signed packets for documents crossing organizational boundaries.

Buyers are classification authorities, security managers, ISSMs, and ISSOs in DoD and IC-adjacent programs, and defense industrial base suppliers preparing for CMMC Level 2.

STANDARDS AND MATURITY

Aligned with the rules your assessor already uses.

Alignment means the design follows the direction each reference sets. It is not an assessment, a certification, or an authorization.

Standards and maturity
CUI Program · 32 CFR Part 2002Recognizes and validates CUI banners and markings
NIST SP 800-171 · CMMCDocumentation maps which Level 2 practices the platform supports; the rest are organizational controls
NIST SP 800-53 Rev. 5 · OSCALA maintained control mapping drives the generated System Security Plan
Section 508 · WCAG 2.1 AAA VPAT 2.5 is published
Technology readinessTRL 5. Components validated end to end in a representative laboratory environment, with the core pipeline running as a containerized deployment.
Intellectual propertyPatent pending. Wagoner Enterprises, LLC has filed United States patent applications covering technologies in this product. No patent has yet issued.
Responsible use
  • Output is a recommendation. The classification authority makes the determination
  • Every decision goes to a human reviewer
  • The platform supplies evidence toward an authorization package; it is not itself an accreditation
  • Decision quality depends on extraction fidelity and on the policies the program maintains
This is a public-safe overview. Classification methods, policy contents, and thresholds are available under appropriate agreements.
GET STARTED

See it classify content you recognize.

We run a sanitized walkthrough on representative content and hand you the outputs to inspect: the recommendation and its rationale, the signed decision record, the release packet, and the audit trail.

Request Technical Walkthrough Back to all solutions Sanitized demonstrations available for classified environments.