Trust & Security

AI Model & Provider Cards

These cards make important AI systems, dependencies and controls visible at a safe level of detail. RaeburnAIos is The Raeburn Group's proprietary AI system and the primary AI layer used by Raeburn Consulting; underlying model and provider dependencies are documented separately where they are relevant and evidenced.

Last reviewed: 27 August 2026

RaeburnAIos

RaeburnAIos is The Raeburn Group's proprietary AI system, developed as the Group's primary intelligence and orchestration layer for its AI-enabled consulting, analysis, discovery and automation services.

Provider / service layer: Proprietary Raeburn AI system operated for Raeburn Consulting; approved underlying model/provider dependencies are governed separately and may vary by deployment.
Model / AI layer: RaeburnAIos is the primary AI system used by consulting.theraeburngroup.com rather than a single third-party foundation-model identity.
Purpose: Provide the AI layer for Raeburn Consulting workflows, including analysis, discovery support, structured recommendations, automation assistance and other approved consulting use cases.
Data categories: Authorised consulting and client context required for the relevant workflow. Credentials, secrets and unnecessary sensitive data are prohibited unless explicitly required, approved and protected for the use case.
Retention: Retention is governed by the relevant Raeburn service and deployment configuration. Prompt, output and supporting-data retention is minimised and must have an operational, contractual, security or legal purpose.
Training setting: Client information is not designated as a general-purpose RaeburnAIos training corpus. Any underlying provider training or data-use setting must follow the approved production provider account, contract and configuration.
Jurisdiction: Deployment and provider specific. Hosting and model-processing locations are taken from the actual approved production configuration rather than assumed from a generic model name.
Limitations: AI output may be incomplete, inaccurate, stale or misleading. Material claims, recommendations and consequential outputs require verification proportionate to their impact.
Human oversight: Human accountability remains with the responsible Raeburn consultant or service owner. Consequential client-facing, legal, financial, security or other high-impact outputs require appropriate human review.
Security concerns / controls: Least-privilege access, separation of privileged instructions from untrusted content, prompt-injection resistance, controlled tool/data access, output validation where applicable, and secure-development/vulnerability-management controls.
Fallback: Workflows should degrade safely when an AI capability or downstream provider is unavailable. Non-AI/manual review, deterministic processing or another approved pathway is used where the service design supports it.
Monitoring: Provider/model changes, security findings, abnormal failures, unsafe-output events, validation failures and material AI incidents are reviewed through the Raeburn security and AI-governance process.

AI Proposal Generator

A supporting Raeburn AI-enabled service for producing structured consulting proposal content from authorised discovery and client context.

Provider / service layer: Supporting Raeburn service; current repository implementation can use the OpenAI API as a downstream model provider.
Model / AI layer: GPT-4.1 mini is the repository default where this component is configured to call OpenAI; deployment-controlled model override is supported.
Purpose: Generate structured consulting-proposal content from authorised discovery and client context when used as part of an approved workflow.
Data categories: Proposal-relevant client and discovery context. Credentials, secrets and unnecessary sensitive data are prohibited inputs.
Retention: Raw prompt retention is not required by the application merely to generate a proposal. Provider/account retention settings and contractual terms are verified separately for the production deployment.
Training setting: Client information is not designated as a Raeburn general-purpose training corpus. Provider-side training/data-use treatment follows the applicable production API account terms and settings.
Jurisdiction: Deployment/provider specific; the production provider contract and account configuration are authoritative.
Limitations: Can hallucinate, omit context, make incorrect calculations or return malformed content. Material claims require appropriate verification.
Human oversight: Human approval is the documented default for consequential/client-facing use.
Security concerns / controls: Client context is treated as untrusted data; prompt-injection and secret-exfiltration instructions are blocked at the privileged prompt layer; outputs are schema-validated.
Fallback: Deterministic non-model proposal generation is used when the model is unavailable, unconfigured or produces invalid output.
Monitoring: Model/provider changes, generation failures, validation fallbacks, security findings and material AI incidents are monitored through the service assurance process.

DiscoveryOS

A Raeburn discovery and evidence platform for structuring tenant-scoped operational information and supporting authorised AI-assisted interpretation.

Provider / service layer: No direct foundation-model provider declared in the current service configuration
Model / AI layer: Optional downstream handoff to an approved Raeburn AI service; a future direct model requires a new approved provider/model card.
Purpose: Collect and structure tenant-scoped discovery evidence and operational observations, with optional AI-assisted downstream interpretation/proposal generation.
Data categories: Tenant-scoped discovery evidence, observations, system metadata and authorised operational/client context.
Retention: Documented default DiscoveryOS retention is 90 days; deployment-specific schedules may be approved. Downstream AI retention is governed separately.
Training setting: Discovery/customer evidence is not designated as a general-purpose model-training corpus.
Jurisdiction: Deployment specific; hosting and downstream AI-processing locations are taken from actual provider/account configuration.
Limitations: Discovery evidence can be incomplete, stale or inaccessible. AI-derived interpretations can be wrong and do not replace source evidence or authorised human judgement.
Human oversight: Operators validate findings and remain accountable for consequential recommendations and customer-system actions.
Security concerns / controls: Tenant-scoped authorisation, encrypted application data, bounded ingestion/rate limits, operator authentication/OIDC support and explicitly configured downstream access.
Fallback: Core discovery and evidence capture do not require a direct foundation model; structured evidence remains available if an optional AI downstream service is unavailable.
Monitoring: Tenant-boundary/access failures, encryption/retention controls, evidence-ingestion limits, downstream handoff failures and AI/security incidents are monitored.

Governance

A model/provider card is reviewed when a provider, model or orchestration layer changes, new data categories are introduced, retention or processing location changes, the service becomes higher impact, or material security/privacy terms change. See our AI Responsible Use & AI Security Standard and Security & Responsible Disclosure Policy.