why most enterprise AI projects die before reaching production

spent the last few months working on agent deployments for non-tech businesses and the biggest takeaway is that prompt engineering is kind of a waste of time if your data is trash. everyone wants to talk about model switching or fine tuning, but the real nightmare is always legacy data. most traditional companies have their actual core knowledge split between random excel sheets, legacy crm setups, and whatever is inside a senior manager's head. the moment you connect an agent to an actual database, security gets nervous because there are zero permission gates. so instead of 5 minutes to deploy, you spend 4 weeks just cleaning data pipelines, setting up warehouses, and building strict authorization layers so the bot doesn't pull restricted records or hallucinate on a customer call. anyone else dealing with this? how are you guys handling authorization gates and schema cleanup before letting an agent touch live data?