Case study
The model already worked. Nobody was using it.
We moved the AI into the workflow legal researchers already had, instead of asking them to come to it.
The situation
AltaML and Jurisage had built something genuinely good: a working AI classifier for legal research, technically sound, doing real work on real documents.
The problem wasn't the model. Using it meant leaving the tools researchers already lived in, going somewhere else, and coming back. That's a small friction. Small frictions are what kill adoption of good AI.
This is one of the most common patterns I see. The AI is real. The delivery vehicle is wrong.
What happened
We redirected the product from a destination to an extension.
Instead of a place researchers had to visit, the capability became a browser extension that surfaced inside the documents they were already reading. The intelligence stayed the same. The distance between the researcher and the intelligence went to zero.
That sounds like a packaging decision. It's a product decision, and it's usually the one that determines whether an AI capability gets used or quietly gets abandoned.
The result
The capability landed inside the existing research workflow rather than beside it, and the engagement led to more than $1M in follow-on work.
Most failed AI features aren't failed models. They're working models placed where nobody will reach for them.
When I score a use case, “who will actually use this, and where will they be standing when they do” is part of the assessment. A technically excellent feature with an adoption problem scores worse than a modest one that lands in an existing habit.
The hard part was that nothing looked broken. The model worked, the engineering was sound, and it was doing real work on real documents. A team that has built something genuinely good does not immediately hear "the product is in the wrong place" as a product problem — it sounds like a criticism of the thing they got right.
The friction we were describing was also small. Leaving one tool, going somewhere else, coming back. Nobody would call that a blocker if you asked them. It is still what determined whether the capability got used, and small frictions are the hardest to argue about precisely because everyone agrees they are small.
Common questions
If the model worked, what was the problem?
The delivery vehicle. Using it meant leaving the tools researchers already lived in, going somewhere else, and coming back. That is a small friction, and small frictions are what kill adoption of good AI.
What changed in the product?
It moved from a destination to an extension. Instead of a place researchers had to visit, the capability surfaced inside the documents they were already reading. The intelligence stayed identical — there were no changes to the model itself.
Is this a packaging decision or a product decision?
A product decision. Where a capability sits relative to the work it supports usually determines whether it gets used or quietly abandoned, and that is not a question of presentation.
Why not just train people to use the separate tool?
Because training works against a friction people notice and fails against one they do not. Nobody would have called leaving their workflow a blocker if asked directly, which is exactly why it went unaddressed while adoption stayed flat.
Your turn
Bring me the list nobody's willing to question.
Twenty minutes, no pitch. If I'm not the right person, I'll say so and point you somewhere better.