Evaluating AI in Regulated Environments.

I Compliance as a Recommender Filter, Not a Gate

The typical enterprise approach to compliance is to gate the output. The AI produces a recommendation; a legal or compliance review approves or rejects it. The review adds time, cost, and friction. Worse, it normalizes the idea that the system might produce non-compliant outputs, which it sometimes does, and which the reviewer sometimes misses.

We inverted the pattern. The compliance filter sits upstream of the recommender. Non-compliant outputs are never produced, because the recommender cannot reach them. The filter is part of the generation, not a post-hoc review. The legal team does not gate output. They define filter rules.

II What the Filter Prevents

A partial list, drawn from installs across regulated verticals:

• Recommendations that would require processing of protected health information outside the boundary the customer has authorized.
• Recommendations that automate a decision which the relevant statute requires a licensed human to make.
• Recommendations that move customer data across a jurisdiction boundary not permitted by the customer’s existing data residency posture.
• Recommendations that imply retention of data beyond the customer’s documented retention policy.
• Recommendations that surface a prediction where the regulation requires an explanation the model cannot provide.

In every case, the filter surfaces an alternative recommendation that stays inside the compliant surface. If no compliant alternative exists, the filter surfaces a no-op recommendation that says so, and routes the underlying need to human review.

III The Audit Trail Is Native

Every recommendation the recommender produces is paired with the filter rules it passed. Every recommendation the recommender suppressed is logged with the filter rule that suppressed it.

This log is queryable by the customer’s compliance team. It is the raw material for audit. It is also the raw material for improving the filter. When a suppressed recommendation turns out to have been over-suppressed (the filter was too strict), the compliance team can adjust the rule. The conversation is grounded in specific outputs, not in abstractions.

IV What Shipping in a Regulated Industry Actually Looks Like

A regulated install does not take longer than an unregulated install. It takes the same seven to ten days. What changes is the composition of the filter rules the install inherits.

The five regulatory frameworks above are not exhaustive. Industry-specific frameworks exist for finance, biomedical, critical infrastructure, and several others. We encode them on first install in the relevant vertical. By the third install in that vertical, the filter is mature. By the tenth, it is better than any single installation would have produced in isolation. This is the compounding loop, applied to compliance specifically.

The customer’s compliance team is not replaced. They are upgraded. Their work shifts from output review to rule authorship. It is the most-valuable version of their own role.