Skip to content

Company Intelligence Profile ​

The Company Intelligence Profile lets organizations define what counts as sensitive in their environment. Instead of relying solely on generic PII patterns, admins register company-specific terms — project codenames, internal product names, customer identifiers, proprietary jargon — so the detection pipeline catches data that off-the-shelf detectors would miss.

Why It Matters ​

Every organization has sensitive information that doesn't fit standard PII categories. A pharmaceutical company's drug candidate name, a defense contractor's program codename, or a startup's unreleased product name won't be caught by Presidio or NER models. The Company Intelligence Profile bridges this gap by feeding organization-specific knowledge directly into the detection pipeline.

How It Works ​

User Input → Standard Detection (Presidio + NER + AI) → Company Profile Terms → Personal Profile Terms → Tokenization

The Company Intelligence Profile injects its terms into the detection pipeline as a dedicated detection layer. When a message contains a registered term, it is flagged and tokenized just like any other PII entity — using the same confidence scoring, compliance tagging, and redaction style configured for the organization.

Detection Pipeline Integration ​

StepDescription
1. Standard detectorsPresidio, NER, Code & Secrets, AI Enhancement run as usual
2. Company termsRegistered sensitive terms are matched against the input
3. Personal termsUser-level terms from Personal Intelligence Profiles are applied
4. Merge & deduplicateAll detections are merged, overlapping spans resolved by highest confidence
5. TokenizationMatched entities are tokenized according to the configured redaction style

Company terms always run before personal terms, establishing an organizational baseline that individual users build upon.

Setup Guide ​

Admin Required

Only users with the Owner or Admin role can create and manage the Company Intelligence Profile. Regular users inherit the profile automatically.

  1. Open Admin Dashboard from the sidebar
  2. Navigate to the Intelligence Profile tab
  3. Configure the three core components described below

Sensitive Terms Registry ​

The Sensitive Terms Registry is the heart of the Company Intelligence Profile. Each term represents a word, phrase, or pattern that should be detected and protected.

Adding Terms ​

FieldTypeRequiredDescription
termstringYesThe word or phrase to detect (case-insensitive matching)
categorystringYesClassification category (e.g., project_codename, customer_name, product)
sensitivitystringNoOverride sensitivity: low, medium, high, critical. Defaults to the category default.
metadataobjectNoFree-form metadata for audit and reporting

Example Terms ​

TermCategorySensitivity
Project Falconproject_codenamecritical
Acme Corpcustomer_namehigh
NeuralForgeproductmedium
Q4-2026-Restructuringinternal_initiativecritical

Bulk Import ​

Upload a CSV or JSON file to register many terms at once:

bash
curl -X POST https://api.ghostai.com/v1/intelligence-profiles/company/{orgId}/terms/bulk \
  -H "Authorization: Bearer ADMIN_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "terms": [
      {"term": "Project Falcon", "category": "project_codename", "sensitivity": "critical"},
      {"term": "Acme Corp", "category": "customer_name", "sensitivity": "high"},
      {"term": "NeuralForge", "category": "product", "sensitivity": "medium"}
    ]
  }'

See the Company Profile API Reference for the full endpoint specification.

Safe Terms Allowlist ​

The Safe Terms Allowlist prevents specific words or phrases from being flagged as sensitive, even if they match a pattern in the Sensitive Terms Registry or a standard PII detector.

Use this for:

  • Public-facing product names that share words with internal codenames
  • Industry-standard terms that resemble PII patterns
  • Common abbreviations that trigger false positives
FieldTypeDescription
termstringThe word or phrase to allowlist
reasonstringWhy this term is safe (for audit trail)

WARNING

Allowlisted terms bypass detection entirely. Use sparingly and review the allowlist periodically.

Data Classification Policy ​

The Data Classification Policy defines the categories available for organizing sensitive terms. Each category carries a default sensitivity level and optional compliance tags.

FieldTypeDescription
namestringCategory identifier (e.g., project_codename)
display_namestringHuman-readable label
default_sensitivitystringlow, medium, high, or critical
compliance_tagsstring[]Compliance frameworks (e.g., ["SOC2", "ITAR"])

Built-in Categories ​

CategoryDefault SensitivityDescription
project_codenamecriticalInternal project or program names
customer_namehighCustomer and partner identifiers
productmediumUnreleased or internal product names
internal_initiativehighStrategic initiatives, M&A targets
financial_metriccriticalNon-public financial figures

Organizations can create custom categories to match their classification taxonomy.

Sensitivity Floor ​

The Sensitivity Floor sets the minimum detection sensitivity for the entire organization. Individual users can increase their sensitivity (via their Personal Intelligence Profile) but cannot reduce it below the floor.

Floor SettingEffect
RelaxedMinimal baseline — only critical-sensitivity terms enforced
BalancedStandard protection — medium and above enforced (default)
StrictMaximum protection — all registered terms enforced regardless of individual sensitivity
Organization Floor (Balanced) → User Sensitivity (Paranoid) → Effective: Paranoid ✓
Organization Floor (Balanced) → User Sensitivity (Relaxed)  → Effective: Balanced (floor enforced)

TIP

The Sensitivity Floor follows the same ceiling/floor pattern used elsewhere in the 5-tier configuration hierarchy (Cloud → Tenant → User → Conversation → Subject). Admins set the floor; users can only go higher.

Auto-Research from Public Information ​

Instead of manually filling every field, admins can use Auto-Research to pre-populate the Company Intelligence Profile from publicly available information about their organization. Auto-Research v2 uses multi-query parallel search, two-pass LLM synthesis, and computed confidence scoring to produce higher-quality suggestions with full source attribution.

How It Works ​

  1. Enter your company name (or let Ghost AI use the organization name and admin email domain)
  2. Click Auto-Research
  3. Ghost AI runs 3 targeted search queries in parallel, scrapes discovered pages with BeautifulSoup-based HTML parsing, and uses a two-pass LLM pipeline to synthesize structured profile data
  4. Review the suggestions — each field and term includes a confidence score and source attribution, and can be individually accepted or rejected
  5. Apply the selections to save them to your profile
Company Name → Domain Auto-Discovery → Multi-Query Parallel Search (3 queries) → HTML Scraping (BeautifulSoup + meta/OG/JSON-LD) → Pass 1: Fact Extraction → Pass 2: Profile Synthesis → Confidence Scoring → Suggestions → Admin Review → Profile

Research Pipeline ​

StageDescription
Domain auto-discoveryWhen only a company name is provided, Ghost AI discovers the company domain from search results
Multi-query parallel search3 targeted queries run simultaneously (company overview, products/services, industry/compliance) via Tavily/Brave/Serper
HTML scrapingBeautifulSoup-based parsing extracts page content along with meta tags, Open Graph data, and JSON-LD structured data
Industry-aware term seedingDetected industry vertical seeds additional search terms for better coverage of domain-specific terminology
Two-pass LLM synthesisPass 1 extracts raw facts from each source; Pass 2 synthesizes facts into a structured profile with cross-source reconciliation
Computed confidence scoringPer-field confidence scores (0.0–1.0) are calculated from source agreement, data completeness, and extraction quality — not self-reported by the LLM
Source attributionEach suggested term and field traces back to the specific page(s) it was derived from

Data Sources ​

SourceWhat It Provides
Web search (3 parallel queries)Industry context, news, public descriptions, product information, regulatory landscape
Website scraping (BeautifulSoup)About page, products page, services page content, plus meta tags, Open Graph data, and JSON-LD structured data
Two-pass LLM synthesisPass 1: per-source fact extraction; Pass 2: cross-source profile synthesis with conflict resolution

What Gets Suggested ​

  • Company description — A 2-3 sentence summary of the organization
  • Industry details — Sector, sub-industry, detected vertical, and whether the company is in a regulated industry
  • Products & services — List of key offerings
  • Data classification tiers — Suggested tiers appropriate for the industry
  • Sensitivity floor — Recommended sensitivity level based on industry regulation
  • Sensitive terms — Proprietary names, product codenames, internal terminology the AI identified, each with source attribution
  • Safe terms — Public brand names and industry terms that commonly cause false positives
  • Per-field confidence indicators — Each suggestion includes a computed confidence score so admins can prioritize review

Reliability ​

  • Retry/fallback logic — If a search provider or page scrape fails, the pipeline retries with alternative queries and gracefully degrades
  • SSRF protection — All scraping targets are validated against an allowlist to prevent server-side request forgery
  • Computed confidence — Confidence scores are derived from source count, agreement, and extraction quality rather than LLM self-assessment

Privacy & Security ​

Auto-Research only uses publicly available information:

  • Web search results from configured search providers
  • Content from the company's public website
  • No private data is accessed or stored during research
  • The company name and domain are the only inputs sent to external services
  • All scraping requests include SSRF protection to prevent internal network access
  • All suggestions require explicit admin approval before being saved

Availability ​

Auto-Research is available in two places:

  • Onboarding Wizard — Step 1 of the Company Profile setup wizard
  • Company Profile Tab — The "Auto-Research" button in the Admin Dashboard

TIP

Auto-Research works best when your company has a public website with an about page. For private or stealth-mode companies, manual entry may be more appropriate. When only a company name is provided, Ghost AI will attempt to auto-discover your domain from search results.

FAQ ​

How quickly do new terms take effect? ​

Immediately. Once a term is saved, it is active for all users in the organization on their next message.

Do terms match partial words? ​

By default, terms use whole-word matching to avoid false positives. A term Falcon will not match FalconX unless you explicitly register FalconX as a separate term or enable substring matching for that term.

Can users see the company term list? ​

Users see that a detection was triggered by a company profile term, but the full registry is only visible to admins. This prevents the term list itself from becoming an information leak.

How do company terms interact with standard PII detection? ​

They complement each other. Standard PII detection (Presidio, NER, AI) runs first, then company terms are matched. If both a standard detector and a company term flag the same span, the higher-confidence detection wins.

Is there a limit on the number of terms? ​

The practical limit depends on your deployment, but the system is designed to handle thousands of terms efficiently using pre-compiled pattern matching.

Can I re-run Auto-Research after the initial setup? ​

Yes. Use the Auto-Research button in the Company Profile tab at any time to refresh suggestions. There is a 5-minute cooldown between research requests per organization.

Next Steps ​

SOC 2 Type II Certified | BAA Available | HIPAA-Ready