My executives will type colleagues' names, performance issues, and salary discussions into your platform, and you pipe that to OpenAI and Anthropic. Convince me that isn't a privacy disaster in progress.?

The full executive answer

It's the right fear — and it's the specific threat our AI pipeline was architected against, at a named chokepoint. Every piece of untrusted text — anything a user types, uploads, or that arrives from an external source — must pass through a sanitisation layer before any AI model sees it. That layer does two jobs: it strips prompt-injection patterns, and it runs PII redaction that detects and masks personal identifiers — emails, phone numbers, and similar direct identifiers — before transmission. The invariant is written into our engineering working agreement as a non-negotiable: never interpolate raw user text into a prompt. It's a codebase rule with a single sanctioned path, not a per-developer habit.

Second layer, contractual: both providers are accessed via their commercial APIs, whose terms exclude training on API inputs — your executives' text is processed to generate the response and is not absorbed into anyone's model weights. That's a real, enforceable DPA flow-down, and both providers are on our published sub-processor list so your team can diligence their terms directly. Third layer, containment of what we retain: PII is masked in our own logs by the logging layer itself, and the most sensitive stored fields carry KMS-managed field-level encryption.

The honest limits, stated plainly: redaction is pattern-based, and no pattern-matcher catches every contextual identifier — "our CFO in the Berlin office" identifies a person to anyone who knows the org chart, and no vendor can truthfully claim to catch that. Our layered posture assumes imperfect redaction: that's why the no-training contractual terms and the encryption exist as independent backstops rather than alternatives. And governance ultimately includes the controller's side: workspace guidance about what belongs in the platform is a control your DPO office sets, and ours amplifies. I'd also flag proactively that this intersects the EU AI Act workforce question — where executives assess employees through our modules, we've pre-built geo-gating pending counsel's classification, which is the same conservatism applied to a different statute.