Most firms have now built something. A prompt chain over a foundation model that pulls figures out of capital account statements, or summarises a diligence pack, or reconciles a report. It works in the demo, it works on the first fifty documents, and then it stops improving. This session is about that wall. 

What causes it, whether better prompting gets you past it, and what the second eighty percent actually costs in engineering time, evaluation infrastructure and ongoing maintenance. We will walk through the same workflow twice, built both ways, and be specific about where building is the right answer and where it is not. And what leading firms are doing today across different uses cases including due diligence, CIM to IM workflows and data room analysis. 

• What can a small internal team realistically stand up in a quarter, and where does it hit a wall?
• At what point does accuracy stop improving with better prompting?
• How do you measure extraction accuracy before an output reaches a report or an investor?
• What is the real maintenance cost once the person who built it moves on?
• When an extracted figure is wrong, who is accountable?