// scope
This disclosure describes our own use of AI in delivering services to clients, and the choices a client has. Where we deploy AI on your behalf (private LLMs, agents), the processing is governed by the Data Processing Agreement and the Statement of Work, not by this page. This page is current as of the date above; material changes are reflected in the consent banner.
Last updated: August 16, 2026
1. Why this disclosure
We use AI to do our work: drafting, researching, reviewing, coding, testing, and operating parts of our infrastructure. We believe you deserve to know which parts of our work involve AI, which models we use, where your data flows, and how to opt out where alternatives exist.
2. Where AI is used
| Surface | AI involvement | Human review |
|---|---|---|
| Marketing site chatbot | Front-line answers to common questions; falls back to a contact form or a human for sales topics. | No human review of individual chats unless you ask for a human. |
| Blog, marketing copy | Drafting, research, and editing; all published copy is reviewed and edited by a human before publication. | Yes, human-in-the-loop before publication. |
| Code in client deliverables | Drafting, refactoring, testing, code review; final commit is reviewed by a senior engineer. | Yes, code review by a human engineer before deployment to production. |
| Operations runbooks | Drafting and updating; reviewed by the on-call engineer before adoption. | Yes, human in the loop before adoption. |
| Customer support triage | Categorization, suggested responses, and routing; a human handles edge cases and approvals. | Yes, human approves before any outbound response is sent. |
| Internal research and analysis | Yes, with human synthesis before any decision is made. | Yes, human in the loop before any decision is made. |
3. Inference posture
The AI we use in delivering our services runs on a mix of:
- Local / private inference on hardware we own (Mac mini + DGX-class) for client work that requires it, including any engagement covered by a DPA with a no-public-cloud instruction.
- Dedicated hosted inference on single-tenant GPU pools where local hardware is not feasible, governed by the engagement SoW.
- Public frontier models (where the engagement permits it, and only with the consent required by the SoW) for tasks where they are clearly the best tool.
We prefer open weights (Qwen, Llama, Mistral, DeepSeek) for any work that touches client data. The specific model used on a given engagement is named in the SoW or in the runbook attached to it.
4. Training posture
We do not train foundation models on client data. For fine-tuning engagements (the "Training & customization" track on our consulting page), we may fine-tune a model on data you provide; that fine-tuning is governed by the SoW and the DPA. In all cases:
- Client data is used only for the documented training objective.
- Resulting weights are delivered to you and not retained by us except as you direct.
- Aggregated, anonymized evaluation metrics may be retained for service-improvement purposes.
- We do not contribute client data to upstream open-weight training corpora.
5. Humans in the loop
For every workflow in Section 2 with a "human in the loop" marker, a named human is responsible for the final review and approval. We document the reviewer in the commit, ticket, or approval system. AI never ships to production without human approval.
6. Data flows
| Activity | Where the prompt goes | Where the response goes | Where it is logged |
|---|---|---|---|
| Client work, default | Local inference on client-isolated hardware | Local | Local audit log |
| Client work, hosted tier | Dedicated GPU pool in named region | Dedicated GPU pool | Client-scoped audit log |
| Public model use, with consent | Public model endpoint named in the SoW | Same | Provider's log under their DPA terms |
| Marketing chatbot on this site | Hosted AI endpoint under our vendor DPA | Returned to your browser | Encrypted application log, 30 days |
| Internal research and ops | Internal AI endpoint, no third party | Internal | Internal audit log |
We do not use your inputs to train public models unless the SoW expressly permits it.
7. How to opt out
- Marketing chatbot. Decline non-essential cookies in our Cookie Policy; the chatbot also has a "talk to a human" path that bypasses AI entirely.
- Internal AI use during an engagement. Tell your project sponsor and we will scope the work to local inference only. The SoW's data-routing terms will be updated to reflect the constraint.
- Customer support triage. Reply "human only" to any triage response and a human will take over without further AI involvement.
8. Accuracy and limitations
AI outputs are probabilistic. They can be wrong, biased, or unsafe. Where a fact, a calculation, or a regulatory statement is material, a human verifies it. The Company does not warrant the accuracy of AI outputs, and the MSA / Terms include a specific disclaimer for AI outputs.
9. Bias, safety, and red-teaming
We red-team the systems we deploy for prompt-injection resistance, jailbreak resistance, data exfiltration, and bias on the dimensions relevant to the engagement. Our Responsible AI Policy describes the principles we apply. For high-risk deployments we recommend a pre-launch review with the client's compliance team; that review is included in Sovereign-tier engagements.
10. Retention of AI prompts and outputs
| Class | Retention | Purpose |
|---|---|---|
| Chatbot transcripts on the marketing site | 30 days | Service improvement and abuse response |
| Internal R&D prompts | 90 days | Service improvement and reproducibility |
| Client-engagement prompts (local tier) | Per the SoW; default 30 days unless extended for audit | Audit, debugging, reproducibility |
| Fine-tuning datasets | Until you direct deletion, plus 30 days | Re-running jobs and audit |
11. Changes
We may update this disclosure as our tooling and the law evolve. Material changes that affect where data flows are reflected in the consent banner, in your SoW (where applicable), or by email notice to clients with a DPA.
12. Contact
Questions: ai-policy@empowered.guru. For security research on our AI systems, see our Responsible Disclosure policy: AI-specific vulnerabilities (prompt injection, model exfiltration, jailbreaks) are in scope.