AI & Data Disclosures

1. AI Models Used

Model	Provider	Purpose
GPT-4o	Azure OpenAI Service	Document data extraction — reads uploaded PDFs and images, returns structured JSON with field values and confidence scores
GPT-4o	Azure OpenAI Service	Content enrichment — normalizes vendor names, dates, currencies; generates categories and keywords
GPT-4o	Azure OpenAI Service	Insights — answers natural-language questions about your document data
text-embedding-ada-002	Azure OpenAI Service	Generates vector embeddings for semantic search and duplicate detection

2. How Documents Are Processed

When you upload a document, the following pipeline runs:

1. Upload — Your file is received over HTTPS and temporarily held in server memory
2. AI Extraction — The document is converted to base64 and sent to Azure OpenAI GPT-4o with a structured extraction prompt. The model returns JSON containing detected fields, values, and confidence scores
3. Storage — The original document is stored in encrypted cloud storage (Supabase Storage). The extracted JSON is saved to the database
4. Embedding — A text summary of extracted fields is sent to Azure OpenAI's embedding model. The resulting vector is stored for semantic search
5. Enrichment — GPT-4o normalizes and enriches the extracted data (vendor names, dates, categories). This runs asynchronously and does not block the extraction response
6. Anomaly Detection — The system checks for duplicate documents, unusual amounts, new vendors, and stale invoices using statistical baselines. Alerts are generated when anomalies are detected

3. What Data Is Sent to AI Models

Document Extraction

The full content of your uploaded document (as base64-encoded PDF or image) is sent to Azure OpenAI. This is necessary for the model to read and extract data from the document.

Content Enrichment

Extracted field names and values (not the original document) are sent to Azure OpenAI for normalization and categorization.

Insights Queries

When you ask a question in the Insights interface, your question and relevant extracted text from matching documents are sent to Azure OpenAI to generate an answer.

Embeddings

A plain-text summary of extracted field names and values is sent to Azure OpenAI's embedding model. Only extracted text is sent, not the original document.

5. AI Accuracy and Limitations

AI-powered extraction is not perfect. Important limitations include:

• Accuracy varies by document quality — handwritten text, poor scans, unusual layouts, and non-English documents may produce lower accuracy
• Confidence scores are estimates — a high confidence score does not guarantee correctness
• Field detection is probabilistic — the model may miss fields, hallucinate values, or misinterpret ambiguous content
• Enrichment is AI-generated — normalized vendor names, categories, and summaries are generated by AI and may contain errors
• Insights are grounded but not infallible — answers are based on your document data, but the AI may misinterpret or miscalculate when synthesizing across multiple documents

You should always review and validate extracted data before relying on it for financial, legal, or operational decisions.

6. Automated Decision-Making

The Service performs automated analysis that may generate alerts or recommendations:

• Duplicate detection — documents with >95% semantic similarity to existing documents are flagged
• Amount anomaly detection — invoice amounts >2 standard deviations from a vendor's historical average are flagged
• New vendor detection — first-time vendors are flagged for awareness
• Stale invoice detection — invoices dated more than 6 months ago are flagged

These are informational alerts only. No automated decisions are made that produce legal or similarly significant effects. You retain full control over all business decisions.

Under GDPR Article 22, you have the right to request human review of any automated processing. Contact privacy@aclearly.com.

7. AI-Related Data Retention

Data Type	Retention
Original documents	Until you delete them or close your account
Extracted JSON	Until you delete the extraction or close your account
Vector embeddings	Until you delete the extraction or close your account
Enrichment data	Until you delete the extraction or close your account
Anomaly alerts	Until dismissed or account closure
Vendor baselines	Until account closure
Data sent to Azure OpenAI	Not retained by Azure after API response

8. Opting Out

AI processing is core to the Service. You cannot use aclearly without AI-powered extraction. However:

• You can delete any document and its associated extraction, embedding, and enrichment data at any time
• You can dismiss anomaly alerts
• You can close your account and all data will be permanently deleted within 30 days
• You can request a full data export before deletion by contacting privacy@aclearly.com

9. Changes to AI Systems

We may update the AI models, prompts, or processing pipeline used in the Service. If we make material changes to how your data is processed by AI (such as switching providers or adding new categories of AI processing), we will notify you via email at least 30 days before the change takes effect.

10. Questions

For questions about AI processing, data handling, or these disclosures:
privacy@aclearly.com