MiQ Digital · 2025–2026
3 min read
A Healthcare AI stack
Rebuilding a healthcare advertising product line around AI features, where the hard constraint is not model quality, it is what you are allowed to know.
$90M+revenue across the product line
Healthcare is the domain where the interesting constraint is never the model. It is the fact that the most useful question a planner can ask is often the one you are not permitted to answer.
I have owned the Healthcare product line through two HCI releases, a segment rebuild, and an AI chatbot that took analysts out of the middle of measurement queries. The current work is rebuilding that line around AI features properly, which in practice means building a stack where compliance is a layer the system passes through, not a review meeting it happens after.
The stack, top down
The order below is the order a question travels, not the order you would build it. That framing matters: every layer exists because the layer above it asked for something it was not allowed to have directly.
Chat, planning tools, and reporting all funnel into the same resolution path. Building one path rather than three is what stops the answer to a question depending on which screen you asked it from.
- chatbot
- planning tool
- reporting
Why the compliance layer sits in the path
The tempting architecture is to let the agents answer freely and check the output afterwards. It is easier to build and it demos better.
It also fails in the one way you cannot afford here: by the time you are checking the output, the restricted thing has already been computed, logged, and very possibly cached. Putting the constraint in the path means the restricted result never exists. That is a meaningfully harder engineering problem and it is not optional.
What this is not
It is not a model achievement. Nothing here required a frontier model or a novel technique. It required knowing which questions a healthcare planner actually asks, which of those are answerable, and building the path that distinguishes them automatically instead of through a person's memory.