AI data-usage statement

Acess Limited, trading as Quantifiable uses hosted models through the Vercel AI Gateway. This statement is part of the privacy notice.

Which models see which data

• Claude (Anthropic) sees structured prompts for routine and complex reasoning: item descriptions, assembly text, report narrative drafts. It does not receive a licence to train on that content.
• Gemini (Google) sees drawing page images when a firm runs takeoff plan analysis, plus the prompt that describes the task.
• OpenAI text-embedding-3-small sees cost-item text used to build a retrieval index. It does not see prices.

Suggestions, not writes

AI output is always a suggestion a human in the firm accepts, edits or rejects. The product does not silently write a priced estimate, bill, programme or report from a model response.

Web research

When a firm uses research, Tavily receives a product description and a unit (for example a timber section and kg/m). The query must not contain a price, a project name, a firm name, an email, or a tenant id. The platform refuses the outbound call if those tokens are present. The linking prompt is price-free and tenant-identifier-free by construction.

Retention and training

We log the flow, token counts and a refusal or validation path for allowance and debugging. We do not use tenant data to train a foundation model, and we do not put tenant data into a global pattern library unless the firm has explicitly opted in.

Questions: privacy@quantifiable.co.uk.