AI Disclosure
Last Updated: March 11, 2026 · Version 1.0
How AI is used across Bespoke Mentis products, what models power each system, training data categories, known limitations, and our responsible AI principles.
§ 1 — Training Data Disclosure (California AB 2013)
California AI Training Data Transparency Act — Effective January 1, 2026
California AB 2013 requires any developer deploying a generative AI system to California users to publicly disclose high-level information about training data. The following applies to AI systems operated by Bespoke Mentis, Inc.:
Mentis AI Chat (Public Website)
Powered by OpenAI GPT-4 and successor models
Training data source: Bespoke Mentis does not train the underlying language model. Training data for OpenAI models is disclosed by OpenAI at openai.com/policies.
Fine-tuning by Bespoke Mentis: System prompts and persona instructions are applied. No fine-tuning of the base model. No client data used for training.
Data collection for training: Conversations are not collected or used for model training by Bespoke Mentis or OpenAI (as governed by our API agreement).
Synthetic data: Not used in model training by Bespoke Mentis.
Personal information in training: Not applicable — Bespoke Mentis does not train models.
MIOS Intelligence Modules
Powered by OpenAI API with proprietary governance layer and prompt engineering
AI model training: Bespoke Mentis does not train the base AI models. AI capabilities are provided via API from OpenAI.
Client data use: Client data processed within MIOS is not used to train, fine-tune, or improve any AI model.
Proprietary components: The MIOS governance layer, constitutional enforcement logic, routing system, and module architecture are proprietary software built by Bespoke Mentis. These do not involve model training.
Agent Conexus
Multi-agent pipeline with Thompson Sampling ML engine and AI-powered email generation
Email generation: AI email variants are generated via OpenAI API. Client prospect data is not used for model training.
Thompson Sampling ML: The self-learning optimization model (Thompson Sampling) is a first-party Bespoke Mentis implementation. Training data: aggregate engagement metrics from the client's own outbound campaigns only. No cross-client data sharing for model improvement.
Lead enrichment: Provided by third-party commercial data providers whose data practices are governed by their own policies. Bespoke Mentis does not collect or train on this data.
Project Foresight
Governed intelligence platform for pharmaceutical and biotechnology organizations
AI model source: Foundation model capabilities provided via API (OpenAI and specialty biomedical models). Bespoke Mentis applies proprietary governance, routing, and citation-verification layers.
Training data: Foresight ingests publicly available biomedical data (published literature, regulatory databases, clinical trial registries). Proprietary client pharmaceutical data is not used for model training.
Data use prohibition: Client compound data, pipeline data, and strategic information processed within Foresight is contractually prohibited from being used for AI model training by Bespoke Mentis or any sub-processor.
Mentis Console
Constitutionally governed AI engineering platform with 15 specialist AI Cores
AI model source: Foundation models via API (OpenAI and others per Core specification). The 15 specialist Cores are Bespoke Mentis–engineered agents built on top of these foundation models using proprietary system architecture.
Client engineering data: Not used for model training. The SHA-256 hash-chained evidence ledger records decisions for the client's own auditability — not for model improvement.
Governance architecture: The constitutional enforcement layer, routing logic, and evidence chain system are proprietary Bespoke Mentis software. Not trained on client data.
§ 2 — AI System Identification (FTC Disclosure)
Bespoke Mentis discloses the use of AI systems in the following contexts on this website and in our products:
You are interacting with an artificial intelligence system. A visible "AI" identifier is displayed in the chat interface. This is disclosed in the Terms of Service § 6 and on the chat page itself.
Every blog article generated by AI includes a visible "AI Generation Disclosure" box in the article sidebar, identifying Mentis Intelligence as the author, the model used, and the retrieval date.
All outputs are clearly generated by AI within the product interface. Each product includes constitutional governance disclosures in its user interface and documentation. Product-level AI disclosures are included in Order Forms and DPAs.
Where marketing content is substantially generated by AI (including outbound email templates), Bespoke Mentis labels such content appropriately per the California SB 942 best practice standard.
§ 3 — Known Limitations & Failure Modes
We disclose the following known limitations of AI systems used in our products. This disclosure is required by EU AI Act Article 13 (transparency obligations) and is consistent with responsible AI practice.
Hallucination
Large language models (LLMs) may generate factually incorrect information with apparent confidence. Bespoke Mentis mitigates this through citation requirements, constitutional constraints, and human review gates — but cannot eliminate hallucination risk entirely. All AI outputs should be reviewed by qualified humans before high-stakes use.
Knowledge cutoff
Foundation models have training data cutoffs and may lack awareness of recent events, regulations, or publications. Foresight mitigates this through live data ingestion; other products rely on the model's training data plus any retrieval-augmented generation configured in the system.
Domain limitations
AI models may perform less reliably in highly specialized domains, novel situations, or contexts outside their training distribution. Bespoke Mentis's specialist Cores are designed to address specific domain limitations — but always require human expert review in regulated contexts.
Bias
AI models may reflect biases present in their training data. This is particularly relevant for Agent Conexus (outbound targeting decisions) and any application involving classification or scoring of individuals. Human oversight is required.
Prompt sensitivity
AI outputs may vary significantly based on how inputs are framed. Constitutional enforcement layers reduce but do not eliminate this variability. Users should treat AI outputs as first drafts requiring human review.
Third-party model dependency
Core AI capabilities are provided by third-party model providers (primarily OpenAI). Changes to those models' behavior, availability, or pricing may affect Bespoke Mentis products. We maintain architecture designed to accommodate model substitution.
§ 4 — Human Oversight Mechanisms
Bespoke Mentis's constitutional governance architecture is designed around the principle that AI should augment human judgment, not replace it. The following oversight mechanisms are implemented across products:
§ 5 — Automated Decision-Making Technology (ADMT)
California CPPA regulations effective April 1, 2027 require pre-use notices for automated decision-making technology (ADMT) used in significant decisions. This section provides advance disclosure of Bespoke Mentis ADMT systems.
Agent Conexus — Lead Scoring and Sequence Optimization
The Thompson Sampling ML engine in Agent Conexus learns which email variants, send times, and sequence structures perform best for each customer segment. This optimization affects which messages are sent to outbound prospects. Key factors: engagement history, segment classification, prior A/B test results. Human operators can override all ML recommendations.
MIOS — Content and Campaign Intelligence
MIOS Intelligence Modules use AI to recommend content strategies, campaign actions, and operational priorities. These are recommendations presented to human operators — not autonomous decisions. Humans approve all significant actions.
To request human review of any ADMT-influenced output, or to exercise opt-out rights, contact with subject line "ADMT Request."
§ 6 — EU AI Act Risk Classification
The EU AI Act (effective August 2, 2026 for high-risk provisions) applies to Bespoke Mentis products used by EU-connected entities. Preliminary risk classification:
Informational AI assistant with no autonomous action capability. Voluntary use with clear AI disclosure.
AI system generating business content and operational recommendations. Human approval required for significant actions. Not listed in Annex III high-risk categories.
AI-generated commercial communications and ML-driven optimization. Human oversight on all campaigns. Transparency requirements apply to AI-generated content.
Clinical intelligence for regulated healthcare/pharma sector. Potential Annex III classification (AI in health). Full technical documentation, conformity assessment, and human oversight documentation in preparation for August 2026 deadline.
AI-assisted engineering tool with mandatory human review gates. Not used for decisions affecting individuals' rights directly.
EU AI Act compliance documentation for Foresight will be published by July 2026. Enterprise customers requiring EU AI Act technical documentation packages should contact .
§ 7 — Contact for AI Concerns
For questions about AI systems, to report AI-related concerns, or to request AI system documentation:
Email:
Subject lines:
- "AI Disclosure — [Question]"
- "ADMT Request" — opt-out or human review request
- "EU AI Act Documentation Request"
- "AB 2013 Training Data Inquiry"
Bespoke Mentis, Inc. · AI Disclosure v1.0 · Effective March 11, 2026