What's the hardest part of an AI project after the contract is signed?

In my experience, it is not the build.

It is getting close enough to the real workflow to understand what actually needs to be built.

The contract usually gets signed around an idea:

“We need an AI assistant.”

“We want to automate this process.”

“We need a chatbot for our clients.”

“We want agents to handle this work.”

But once the project starts, the real work begins.

You find out the process is different depending on who does it.

The knowledge is spread across Slack, emails, spreadsheets, old docs, and someone’s memory.

The exceptions matter more than the standard flow.

And the person who approved the project is often not the person doing the work every day.

That is where AI projects either become useful or become another tool nobody trusts.

The hardest part is turning the client’s original request into something operationally real.

Not just what they asked for.

What the business can actually use, adopt, and rely on.

Curious how others see this: After the contract is signed, what part of an AI project usually becomes the hardest for you?