We've been running AI across a team for about two years. Expected the hard parts to be the models. They weren't.
The problem that cost us most early on was context. We had a system making customer-facing recommendations without access to the business-specific knowledge it needed to answer accurately. Spent too long trying to fix it at the prompt level. The context layer didn't exist, and prompting didn't fill that gap, it just made it less obvious until something downstream failed badly enough to trace back to it.
That failure pushed us to map the other places where AI builds break structurally rather than technically. We found five more, and they kept showing up across different stacks and different team sizes in roughly the same order.
The first is identity, when you move from one person's AI to a team's AI, shared context without role-based permissions either creates noise or recreates the same knowledge silos you were trying to escape.
The second is decision memory, records of what was decided aren't the same as memory of why, and that gap compounds quietly until a new team member gets a confident wrong answer from a system referencing reasoning that was abandoned months ago.
The third is attention. Dashboards only work if someone looks at them, and the failure mode of every dashboard ever built is the same: critical things slip through when life gets busy.
The fourth is write-back. Manual logging is a tax on the busiest moments, and the more important the work, the less likely anyone stops to document it.
The fifth is governance, when the same agent that builds something also evaluates it, that's not a check, it's a loop grading its own homework.
The sixth is economics, at solo scale AI cost is a rounding error, at team scale you're looking at a vendor invoice with no way to connect spend to specific workflows or outcomes.
Which of these have you hit? And did they show up in this order or did something else surface first? If you're interested, we turned these into a diagnostic with 14 questions. Takes about five minutes, link in the first comment if you want to run through it.
[link] [comments]