Why your first AI project fails at the foundation

Most AI projects don't fail because of the AI, but because of the foundation underneath. We lived through this in our own companies before turning it into a method for others. The technology is rarely the problem – the problem is that the AI meets a house whose basement was never cleaned. Three things must be in place before the first prompt.

1. The AI has to reach the data

A model is only as good as its access to your numbers. If orders, receipts and customer data sit in separate systems without interfaces, even the best model won't help – it simply can't see what it's meant to judge. For us the first work was exactly here: not a clever prompt, but connecting six systems so a shared data basis even exists. Connect first, then automate. Reverse that order and you build on sand.

2. The process has to be clean

AI accelerates what's already there. Automating a chaotic workflow doesn't create order – just faster chaos, at higher throughput. So before automation comes the uncomfortable question: is this process even worth speeding up? Often the most honest first AI move is to delete a process, not automate it. Sounds paradoxical, but it saves more money than any model.

3. You need a human with sign-off

Early on, AI generates for approval, not autonomously. Someone reviews and corrects the first results – and the system learns from exactly those corrections. We run it in stages: from "system suggests" through "human approves" to, once quality is proven, "runs autonomously with spot checks." Trust is earned, not assumed. Letting an AI loose on your customers unsupervised on day one confuses courage with recklessness.

The pattern behind it

Data access, clean process, human sign-off – sounds unspectacular? It is. That's exactly why it works. The exciting AI demos fail in operations on precisely these three boring points. Settle them first and you'll need less AI to achieve more.