Your contract review agent can read an NDA and summarize the key terms. But ask it "what are the risky clauses?" and you get a different answer every time.
Neither gets solved by a bigger model.
Here's a deep dive of the Box AI Extract NDA risk workflow.
The agent calls POST /2.0/ai/extract_structured with a fixed five-field schema: governing law, term, termination, confidentiality period, liability cap. Confidence scores and source references included. A transparent rule set scores each clause automatically.
The Deal Risk Summary saves back to Box as a companion file alongside the source document. Same output shape every time, regardless of how the document is written.
Extract, don't prompt.