Your talent and skills modules make workforce-affecting recommendations. Where is your bias testing — disparate impact, protected attributes, fairness metrics? Show me.?

The full executive answer

Directly: we have not run formal statistical fairness testing — no disparate-impact ratios, no demographic-parity metrics — and I will not pretend otherwise. The primary control is architectural, and it is the strongest kind: the workforce modules operate on organisational aggregates only — skills-gap patterns, team-level adoption, aggregate sentiment. No individual employee records, no protected attributes, no named-person scoring goes in, and no hire, fire, promote or evaluate recommendation on any individual comes out. The classic disparate-impact failure — a model scoring people differently across demographic groups — has no input surface here, by design rather than by testing.

That is a real mitigation but not a complete one, and I know the sophisticated version of your question: aggregate-level advice can still encode bias — a recommendation pattern that systematically disadvantages certain functions or geographies, or biased framing absorbed from the underlying model. Current compensating controls: a fixed low temperature and structured output schemas that constrain free-form judgment; provenance validation that stops invented justification figures; the consensus mechanism putting a second provider’s model — different training data, different biases — against the first on high-stakes output; and the feedback loop capturing flagged outputs for review.

The committed roadmap, in order: qualified counsel review of these modules against EU AI Act Annex III — they are explicitly flagged for that in our model cards, with a pre-built geo-gate that can disable them for EU organisations until conformity; then structured bias evaluation of aggregate recommendations across synthetic organisation profiles as our eval sets deepen; and external audit as we scale. This is exactly the gap our own product would flag in a customer, and I would rather show you the honest model card than a fabricated fairness report.