AI may assist. Defined software policies assess supported rules. Humans make regulatory decisions. Where AI is used When enabled, AI assists recipe drafting, kitchen questions and advisory shelf-life estimates. The verifier pilot can use AI to select relevant server-generated evidence facts for a draft summary. Feature availability depends on the deployment. The current implementation uses Anthropic’s Claude API with the model configured for the service. We do not advertise a fixed model version or a provider substitution without updating the relevant disclosures. Information sent to the provider Depending on the feature, requests may include your prompt, recipe/ingredient information, selected operational context and relevant food-safety guidance. The verifier pilot sends the question and a limited set of evidence facts, which may include attribution appearing in that evidence. Do not enter passwords, payment details, unnecessary personal identifiers or medical details into AI prompts. Staff health declarations are not automatically supplied as general assistant context; a person can still disclose sensitive information by typing or pasting it, so review inputs before submitting. Provider processing locations, retention and training permissions depend on the applicable service terms and account settings. We do not make a blanket “never used for training” or zero-retention promise. Ask us to confirm the applicable terms before submitting restricted information. Review and traceability Treat generated recipes, explanations and estimates as drafts. Check ingredients, allergens, quantities and food-safety advice against reliable sources and your registered plan. Manager confirmation is required before an advisory shelf-life estimate is used for a label. AI does not decide that food is safe. Supported food-safety assessments use defined software policies. A record without a supported assessment must not appear to pass merely because it was saved. Verifier draft assistance is labelled, cites underlying evidence and cannot author an official verification decision. The authorised reviewer remains accountable. Logs, limitations and feedback The hardening records AI access and outcomes without including prompt or response bodies in its general security audit. Some features retain an associated business record, such as an estimate, its rationale or cited fact references. Provider request retention is separate from our audit logging. AI can produce plausible errors and can misunderstand incomplete or adversarial inputs. Food-safety evidence, source references and human review take priority. Report inaccurate or unsafe output with the feature and approximate time; avoid emailing health data or credentials.
This page needs JavaScript to display in full.