AI Usage & Ethics
We build with AI for a living, so we hold ourselves to a clear standard for how it's used. This is what runs your systems, what it never does, and where a human always stays in the loop.
/1.0Your Data Never Trains AI Models
We use enterprise and API tiers from our AI providers, which contractually prohibit using your inputs or outputs to train their models. Your business data is processed to do your work, and nothing else.
/2.0Which AI Models We Use
We are model-agnostic and pick the right tool for each task. Our primary providers:
- Anthropic (Claude), our default for reasoning, agents, and anything touching judgment. Accessed via Anthropic's API or AWS Bedrock so data can stay in your cloud region.
- OpenAI, used for specific tasks where it's the best fit, under API business terms (no training on your data).
- Other specialized models (e.g. transcription, vision) are selected per engagement and disclosed to you. The full list of AI sub-processors lives on our Data Governance page.
/3.0Human-in-the-Loop
AI does the work; humans own the decisions that matter. For any action that is irreversible, externally visible, or financially or legally consequential (sending communications to third parties, moving money, deleting data, filing documents), a person reviews and approves before it happens. We design systems so the AI proposes and a human disposes wherever the stakes warrant it. You decide, per workflow, where that line sits.
/4.0Error & Hallucination Handling
Language models can be confidently wrong. We engineer against it rather than hope:
- We ground outputs in your real data and sources, not the model's open-ended recall.
- Confidence thresholds and validation checks gate automated actions; low-confidence results route to a human.
- We prefer reversible, idempotent operations and keep an audit trail so any action can be traced and undone.
- We monitor live systems and correct the root cause when something drifts. We don't paper over bad output.
/5.0Responsible Use
We build automation to augment your team, not to deceive anyone. AI agents that interact with people are identifiable as automated when that matters. We don't build systems designed to mislead, surveil unlawfully, or discriminate. If a use case crosses an ethical or legal line, we'll say so and decline it.
/6.0Compliance Posture
We want to be precise here. Fellowship AI is a consulting firm and does not itself hold SOC 2, ISO, or HIPAA certifications. The strong compliance baseline we operate under is inherited through our providers. We build on infrastructure and models that carry independent, audited certifications:
- Anthropic: SOC 2 Type II · ISO 27001 & ISO 42001 (AI management) · GDPR-aligned · HIPAA-eligible under a BAA · no training on API customer data
- AWS: SOC 1 / SOC 2 / SOC 3 · ISO 27001 · PCI DSS Level 1 · HIPAA-eligible · GDPR-compliant data processing
For GDPR- or HIPAA-exposed work, we structure engagements to use the right tiers (e.g. Bedrock with a BAA) and execute the necessary agreements. A DPA is available on request, and our own security controls are detailed on the Data Security page.
Questions about how AI is used in your engagement? Email [email protected].