Skip to main content

AI Governance

Document statusPublic documentation
Target audienceEnterprise security & compliance teams
Last updated2026-01-28

Scope of this document

This document provides a technical and architectural overview of Phoenix AI governance controls. It describes current system design, configurations, and operational practices.

This document is not:

  • A contractual agreement (see your HG Insights MSA for binding terms)
  • A substitute for HG Insights Corporate InfoSec documentation (SOC 2, penetration testing, etc.)
  • A guarantee of specific SLAs (see service agreement)

Specific contractual commitments, audit rights, and liability terms are negotiated separately in customer agreements.

Executive summary

Phoenix is HG Insights' AI-powered platform that enables enterprise customers to run intelligent agents for account research and GTM intelligence. Phoenix agents query HG Insights' proprietary third-party data to answer go-to-market questions about target accounts, markets, and technologies.

Key differentiator: Unlike AI tools that process your sensitive internal data, Phoenix's core function is to make HG's data accessible through natural language. Customer data exposure is limited and well-defined.

1. What Phoenix agents actually do

1.1 The core value proposition

Phoenix agents answer GTM questions by accessing HG Insights' data assets:

Data sourceTypeExamples
FirmographicsHG proprietaryCompany size, revenue, industry, locations
TechnographicsHG proprietaryTechnology installations, spend, vendors
Intent signalsHG proprietaryBuying signals, technology changes
Contact dataHG proprietaryDecision-maker information
SEC filingsPublic10-K, 10-Q analysis
News/webPublicRecent company announcements

1.2 Customer data: what Phoenix receives

Phoenix receives the following categories of customer data:

Data categoryDescriptionHow it's usedLogged in traces
Target account identifiersDomains, HGIDs, company namesInput to research queriesYes (see section 4)
Branding assetsLogo URL, brand colorsOutput stylingNo
LLM API keys (optional)Customer-provided OpenRouter/OpenAI keysRoute LLM calls through customer accountNo (redacted)
Enablement content (optional)GTM playbooks, industry messaging, problem-solution mappingContextualize outputs for customer's marketYes
Target account lists

We recognize that strategic account lists may be competitively sensitive. Target identifiers are logged in execution traces for auditability. Trace access controls are described in section 4.3.

Phoenix does NOT receive or process:

  • Customer PII databases
  • Customer financial systems or transactions
  • Customer internal documents (unless explicitly provided as enablement content)
  • Customer CRM credentials (OAuth tokens are used; credentials are never transmitted)

1.3 Composable architecture: bring your own data and models

Phoenix is designed to be highly composable. Customers can swap components to maintain full control:

ComponentDefaultCustomer option
Data sourcesHG Insights MCP toolsSubstitute customer's own MCP tools
LLM providerOpenRouter (Claude/GPT)Customer's own API keys (OpenRouter or OpenAI direct)
LLM routingVia OpenRouterBypass OpenRouter with direct provider keys
Agent executionPhoenix SaaSOn-premises deployment (roadmap)

Key benefit: If a use case requires first-party customer data, the customer provides their own MCP tool. Phoenix orchestrates the agent, but the sensitive data flows directly from the customer's tool to the LLM. HG never receives or stores it.

Complete OpenRouter bypass: Customers can completely remove OpenRouter from the architecture by providing their own OpenAI API key. When configured:

  • LLM calls go directly from Phoenix to OpenAI
  • OpenRouter is not involved in any data flow
  • Customer's OpenAI contract governs data handling

For the list of HG Insights subprocessors, refer to HG Insights Corporate InfoSec documentation.

2. AI model governance

2.1 LLM provider architecture

Customer Request


┌─────────────────┐
│ Phoenix Agent │
│ (Orchestration)│
└────────┬────────┘


┌─────────────────┐ ┌─────────────────┐
│ OpenRouter │─────►│ Claude / GPT │
│ (or direct) │ │ (Inference) │
└─────────────────┘ └─────────────────┘
ComponentProviderRegion
Agent orchestrationHG Insights (Railway)US
LLM routingOpenRouter (or direct to provider)Varies by provider
Model inferenceAnthropic / OpenAIVaries by provider

2.2 Data sent to LLMs

When an agent runs, the LLM receives:

  • Prompt instructions (HG-authored agent prompts)
  • HG Insights data (firmographics, technographics, etc.)
  • Target account identifiers (domain, company name)
  • Optional customer context (enablement content, persona settings)

The LLM does NOT receive:

  • Customer CRM credentials
  • Customer API keys
  • Bulk customer databases

2.3 Model training policy

No customer data is used for model training.

HG Insights configures OpenRouter with all training/logging disabled:

OpenRouter settingStatusEffect
Paid endpoints training on inputsDisabledNo training on prompts
Free endpoints training on inputsDisabledNo training on prompts
Free endpoints publishing promptsDisabledNo public dataset exposure
Input/output loggingDisabledNo storage with OpenRouter
ZDR Endpoints OnlyAvailableRoute only to Zero Data Retention models

Provider terms:

  • Anthropic API: Data not used for training per API terms
  • OpenAI API: Data not used for training per API terms

Customer control: Customers can bring their own API keys for additional control over their LLM provider relationship.

2.4 Customer model choice

OptionDescription
Phoenix defaultClaude 3.5 Sonnet via OpenRouter (privacy settings enabled)
Alternative modelsSelect from any model available on OpenRouter
Bring your own keysUse customer's OpenRouter or OpenAI API keys directly
Bypass OpenRouterDirect OpenAI integration available with customer keys
Custom deploymentOn-prem agents with customer's LLM infrastructure (roadmap)

3. Agent architecture

3.1 Read-only by design

Phoenix agents are read-only. They query data sources and generate outputs. They do not:

  • Write to customer CRM systems
  • Send emails on behalf of users
  • Modify external databases
  • Execute webhooks or trigger downstream actions

Agents consume data via MCP tools and produce artifacts (briefs, reports). All actions are observable in execution traces.

3.2 Tool-based architecture

Each agent has an explicit tool allowlist defined in its configuration:

# Example: Account Research Agent
tools:
- company_firmographic # HG data - read only
- company_technographic # HG data - read only
- company_spend # HG data - read only
- sec_filing_section # Public data - read only

Agents cannot call tools outside their allowlist.

3.3 Tenant tool control

Customers control which tools their agent instances can access:

ControlDescription
Enable/disable toolsTurn off specific MCP tools per tenant
Tool substitutionReplace HG tools with customer's own MCP tools
Parameter defaultsSet default research depth, persona, output format

4. Execution tracing and auditability

4.1 What gets logged

Every agent run captures:

Trace elementDescriptionStored
Run metadataAgent version, tenant, user, timestampYes
Tool callsWhich tools called, with input parametersYes
Model usageProvider, model, token count, latency, costYes
OutputsGenerated artifacts (briefs, reports)Yes
ErrorsExceptions and failure detailsYes

Retention: 2 years default. Shorter retention available for Enterprise tier customers.

Minimal logging mode: Enterprise customers can enable minimal logging, which stores only run metadata (cost, user, timestamp, success/failure) without input parameters or output content. Note that even with minimal logging enabled, prompts are still transmitted to LLM providers per their data handling terms.

4.2 What does NOT get logged (redaction)

Sensitive fields are automatically redacted before storage (not after retrieval):

  • OAuth tokens and refresh tokens (Bearer *, access_token, refresh_token)
  • API keys (patterns: sk_live_*, sk_test_*, sk-*, sk-ant-*, phx_*, AKIA*, AIza*)
  • Passwords and credentials (field names: password, secret, credential)
  • JWT tokens
  • Database connection strings (PostgreSQL, MySQL, MongoDB, Redis)
  • Customer LLM API keys

Redaction events are logged for audit purposes (field redacted, pattern matched) without exposing the original sensitive values.

4.3 Trace access controls

RoleAccess
Customer adminOwn tenant's traces only
Customer end-userResults only (trace access configurable per tenant)
HG SupportRequires customer-initiated support ticket; audit-logged internally

HG Support access policy: HG personnel cannot access customer tenant traces without a customer-initiated support ticket. All support access is logged. For Enterprise customers requiring additional controls (e.g., explicit approval workflows), discuss with your HG representative.

5. Output governance

5.1 What agents produce

  • Account research briefs - HTML/PDF one-pagers about target companies
  • Structured data - JSON/CSV exports of research findings
  • Recommendations - Suggested talking points, competitive insights

5.2 Output quality controls

ControlDescription
Source groundingOutputs generated from HG proprietary data, reducing hallucination risk
Source attributionBriefs include citations to data sources
Output loggingAll outputs stored for quality auditing
Structured promptsAgent prompts enforce consistent output format

5.3 Output accuracy disclaimer

Phoenix outputs are generated by AI models and are intended for informational purposes. While outputs are grounded in HG Insights data:

  • HG Insights does not warrant the accuracy of AI-generated content
  • Human review is recommended before business decisions
  • Specific accuracy SLAs can be discussed in Enterprise agreements

Customers can configure output disclaimers to appear on generated artifacts.

5.4 Output ownership

Generated outputs belong to the customer tenant. HG Insights:

  • Does not use outputs to train models
  • Does not share outputs across tenants
  • Retains outputs per retention policy (default 2 years; configurable for Enterprise)

6. Infrastructure and data residency

6.1 Current deployment (SaaS)

ComponentProviderRegion
Agent ServiceRailwayUS
DatabaseNeonDB (PostgreSQL)US
Artifact StorageAWS S3US

Tenant isolation: Logical isolation via per-tenant database schemas and scoped API access.

Data residency: Currently US-only. EU or region-specific deployment is not available in the current architecture. Hybrid/on-prem deployment (roadmap) would enable customer-controlled data residency.

6.2 Roadmap: hybrid/on-premises

For enterprises requiring additional control:

  • Agent execution on-premises - Run agents in customer infrastructure
  • Configuration from Phoenix - Manage agent definitions via Phoenix UI
  • Customer-controlled LLM - Use customer's own LLM deployment
  • Customer-controlled data residency - Data stays in customer environment

Contact your HG Insights representative for roadmap timing.

7. Risk profile summary

7.1 Why Phoenix has a favorable risk profile

Risk factorPhoenix postureRationale
Customer PII exposureLowAgents query HG data, not customer PII
Data exfiltrationLowNo bulk customer data processing
Model trainingNoneOpenRouter settings + provider API terms
Autonomous actionsNoneAgents are read-only; no writes to external systems
Output hallucinationMitigatedGrounded in HG data with source attribution

7.2 Security controls

ControlImplementation
Input validationDomain validation (TLD checks, blocklists), HGID format validation, input size limits (1KB/param, 10KB total), prompt injection pattern detection
Prompt architectureXML-delimited system instructions, immutable assistant identity, role-locking constraints, explicit boundary between system and user content
Output constraintsStructured prompts enforce format boundaries, source attribution required
Trace redactionAutomatic redaction of OAuth tokens, API keys (Stripe, OpenAI, Anthropic, AWS, GCP), passwords, JWTs, and connection strings before storage
Tenant isolationPer-tenant database schemas; scoped API access
Rate limitingPer-tenant concurrency limits (default: 10 runs)

8. Frequently asked questions

Does Phoenix have access to our Salesforce data?

Only if you explicitly provide an MCP tool that connects to your Salesforce. By default, agents only access HG Insights data.

Can Phoenix agents modify our systems?

No. Phoenix agents are read-only by design. They query data and generate outputs. They cannot write to CRMs, send emails, or trigger external actions.

Can we bypass OpenRouter entirely?

Yes. Customers can provide their own OpenAI API key. When configured, OpenRouter is completely removed from the data flow. LLM calls go directly from Phoenix to OpenAI. Your OpenAI contract then governs data handling.

Where is our data stored?

Currently US-only (Railway, NeonDB, AWS S3). Hybrid deployment enabling customer-controlled data residency is on the roadmap.

What if we need shorter data retention?

Default is 2 years. Shorter retention (or immediate deletion) is available for Enterprise tier customers.

Can we disable logging of our inputs?

Yes (Enterprise). Minimal logging mode stores only run metadata (cost, user, timestamp, success/failure) without input parameters or output content. Note that prompts are still transmitted to LLM providers regardless of logging settings.

Is our data used to train AI models?

No. OpenRouter privacy settings are disabled, and provider API terms prohibit training. Customers can use their own API keys for additional control.

Can we audit Phoenix?

Execution traces are available via Phoenix UI and exportable. Broader audit rights (infrastructure, subprocessors) are addressed in Enterprise agreements.

Can we run Phoenix in our own environment?

Roadmap. Hybrid deployment (agents on-prem, configuration via Phoenix) is planned. Contact your HG representative for timing.

9. Contact information

Contact typeDetails
Enterprise salesContact your HG Insights representative
Security & complianceRefer to HG Insights Corporate InfoSec
Contractual termsRefer to your HG Insights MSA

Document control

VersionDateAuthorChanges
0.12025-12-11Phoenix TeamInitial draft
0.22026-01-28Phoenix TeamClarified data categories, read-only architecture, residency, disclaimers
0.32026-01-29Phoenix TeamDetailed security controls implementation (input validation, trace redaction, prompt hardening)