My uploaded board packs contain names, emails, salaries. What personal data actually reaches Anthropic and OpenAI, and what is your legal basis story?

The full executive answer

The design rule is data minimisation enforced in code: personal identifiers are redacted before egress, on every path. The same hardening function that neutralises prompt injection also scrubs high-confidence personal data — email addresses, phone numbers, card numbers, national identifiers, network addresses — from any untrusted text before it is included in a prompt. This is declared per model in our registry as the uniform posture: redacted before egress. Our logging is under the same discipline — the logger masks personal data, so identifiers do not leak into operational logs either.

What the providers receive is therefore organisational context — scores, industry, maturity patterns, redacted document excerpts — which is the point: the analysis is about the organisation, not about persons, so the product does not need personal data at the model layer to do its job. Combined with the providers’ no-training API terms and available zero-retention options, the residual personal-data exposure at the model layer is designed toward nil.

Two honest boundaries. Pattern-based redaction catches structured identifiers with high confidence; free-text references — "our CFO Jane said" — are mitigated by context design and length caps rather than perfectly caught, which is exactly why the per-workspace no-third-party-model switch exists for tenants whose risk appetite demands zero egress. And the formal GDPR paperwork — records of processing, DPIA for the document pipeline — is being completed with counsel as part of launch readiness; I will show you the technical controls today and the paperwork status as it stands, not as I wish it were.