AI Usage and Risk Policy
1. Purpose and Scope
This policy describes how Docwright Software LLC ("DocWright") manages risk associated with the optional AI-assisted features within the DocWright application (alt text, equation descriptions, and scanned-page transcription). It applies to DocWright's use of third-party AI provider APIs: the Anthropic Claude API, the Google Gemini API, and Azure OpenAI. Each install uses one provider, chosen by the institution. DocWright does not develop, train, host, or operate any AI or machine learning model itself.
2. Two-Layer Risk Model
DocWright's AI risk posture rests on two distinct layers, and this policy is deliberate about not conflating them:
Layer 1 — Underlying infrastructure (the AI provider's responsibility)
The AI provider the institution has configured is responsible for its own infrastructure, security program, and data handling. Anthropic publishes SOC 2 Type 2, ISO 27001, and HIPAA compliance certifications covering the Claude API. Google (Gemini API) and Microsoft (Azure OpenAI) publish their own compliance documentation, which DocWright has not independently assessed. DocWright relies on the provider's infrastructure rather than duplicating it.
Layer 2 — DocWright's own integration risk (DocWright's responsibility)
Separately, DocWright is responsible for how it integrates and constrains its use of that API. This policy addresses that second layer specifically — it is not a restatement of a provider's certifications, but a description of DocWright's own design choices that limit risk in how the API is used.
3. DocWright's Risk-Limiting Design Choices
- Minimal data exposure: Only a single image is sent per AI request: an extracted figure or equation for descriptions, or one page for scanned-page transcription. The document file and its metadata are never sent. For transcription, everything visible on that page goes to the provider. This is a fixed architectural constraint, not a configurable setting that could be loosened.
- Opt-in by design: The AI features are never invoked automatically. A user must explicitly click "Generate AI Descriptions" for each document, and transcription is a separate action ("Transcribe Pages"). An institution can disable the features by not providing an AI provider API key. DocWright does not currently enforce this centrally: a user can enter their own key or switch provider in Settings, so institutions that need to prevent this should address it in their acceptable-use policy.
- Institution-controlled API relationship: Each institution supplies and manages its own API key and account with its chosen provider. DocWright does not provide, proxy, or have visibility into this key, and is not a party to the data-handling terms governing that institution's own API usage — those terms are between the institution and the provider directly, under the provider's commercial (not consumer) terms. Anthropic's and Microsoft's commercial terms state that data is not used to train their models; Google's Gemini terms say the same for paid services but allow unpaid (free-tier) use to be used to improve its products, and DocWright cannot see which tier a key belongs to, so institutions that choose Gemini should issue paid-tier keys. Azure OpenAI retains prompts and outputs for up to 30 days for abuse monitoring unless modified abuse monitoring is approved.
- No model training or fine-tuning: DocWright does not train, fine-tune, or otherwise modify any AI model. It is solely a consumer of the providers' existing commercial APIs.
- No plugin or multi-step automation: Each AI request is a single, direct API call with no chained actions, plugins, or downstream automated decisions based on the model's output beyond inserting the generated description text into the document.
4. Human Oversight
AI-generated descriptions are inserted into the processed document automatically, without requiring approval prior to insertion. Scanned-page transcriptions work differently: each is shown to the user for review and editing before it is embedded. However, the user retains the original file alongside the processed file and a fix report listing all AI-generated content, allowing review and manual editing of any description before final use (e.g., before uploading to the institution's LMS). DocWright considers this post-hoc review workflow appropriate given the low-stakes, easily reversible nature of the output (descriptive alt text, not a decision affecting any individual).
5. Incident Response for the AI Feature
Because the AI feature depends entirely on an institution-supplied API key, it can be disabled immediately by removing or not renewing that key in the application's Settings panel (or by not providing one at deployment) — no DocWright-side action or reinstall is required. This is the primary control available in the event of an identified issue with the feature.
6. Current Limitations
DocWright does not yet have a formal, written AI risk-management framework modeled on a published standard (e.g., NIST AI RMF), nor a dedicated AI governance function, consistent with the company's current single-founder structure. DocWright's provider choice is not centrally locked, and DocWright cannot see which service tier an API key belongs to; both are described above. This policy represents DocWright's current, good-faith documentation of its actual AI risk-limiting practices, and will be expanded as the company grows.
7. Related Documents
- DocWright Data Privacy Policy
- DocWright Information Security Policy
- DocWright Data Privacy Impact Assessment
8. Contact
Gregory Buthusiem, Founder & Managing Member
Docwright Software LLC
(856) 817-5100