The strongest theme at UpperBound 2026 yesterday: AI projects fail for very normal reasons.

And people usually do not talk about them.

I attended four very different sessions at UpperBound 2026 yesterday.

Image from the UpperBound2026 website.

They all landed on the same conclusion: the hard part of AI is not the model.

Debra Somani gave one of the clearest talks I’ve seen about AI teams. Her point was simple: most AI success comes from intentional human design, not technology choices.

How to drive innovation and AI adoption from strategizing your tea

She broke teams into three roles:

• champion
• builder
• connector

What I liked is that these are operational roles, not titles. The “connector” idea especially stayed with me. Someone has to translate between AI capability and business outcomes or the whole thing turns into disconnected demos. This resonates with the organization of a team to drive innovation, not necesarily AI.

Then Vinoth Babu from AltaML talked about the “reality gap” between demos and production agents.