The platform's roughly 120,000 members trusted its daily recovery and strain scores enough to change how they trained, slept and ate that day. That level of influence over daily decisions came with no matching documentation: the models behind the scores had no written statement of what they were intended to be used for or where their limits sat.

Oversight had not kept pace with the models' reach either. There was no bias or drift review across member demographics, so nobody could say whether a score behaved consistently for every group of members relying on it. And there was no formal release sign-off, meaning a new or retrained model could reach members with no defined check beforehand.

Scrutiny of consumer health AI was growing at exactly the moment this gap was most exposed. Product, legal and support teams were each carrying risk they could not fully see: product could not point to documented limits, legal could not point to a bias review, and support fielded questions about scores with no plain-language explanation to offer members.

We started with a complete model inventory, cataloguing every production health-insight model with a model card and a documented intended-use and limitation statement. For the first time, product, legal and support could all point to the same written record of what a given model was meant to do and where it stopped.

We then made review a standing habit rather than a one-off exercise. Quarterly bias and drift reviews now examine every model across demographic slices, checking that scores behave consistently for the different groups of members within the roughly 120,000-member base rather than assuming they do without any supporting evidence at all.

Finally we closed the loop with members and with governance itself. Plain-language in-app score explanations now tell members why a score looks the way it does, and a cross-functional review board runs a defined sign-off workflow for every new or retrained model before it ever reaches a single subscribed member.