Key Takeaways

  • Jenny Heller at Brandywine Group Advisors built an internal operating model that moves AI adoption past sandbox experiments into core allocator workflows.
  • Brandywine paired draft performance reviews with internal "How I Operate" personal profiles inside Claude, asking the model to reframe manager feedback to match each employee's communication style.
  • A test to generate mid-diligence updates between early issue memos and formal investment memos within one hour collapsed because the model could not synthesize unstructured deal judgment.
  • Institutional investment offices maintain underwriting integrity by deploying Heller's Three Layers of AI Trust.

Heller's Three Layers of AI Trust

  • Layer 1: Perimeter & Confidentiality Trust: Verify enterprise-grade data boundaries where proprietary data is walled off and never used to train public foundational models.
  • Layer 2: Output Verification Discipline: Enforce mandatory human review and cross-examination on every model-generated deliverable before it informs consequential decisions.
  • Layer 3: Team Provenance Transparency: Maintain full transparency across the organization regarding what proportion of an investment memo, note, or analysis was drafted by an AI versus a human investor.

When This Works (and When It Doesn't)

This framework works when institutional allocators need to move past ad-hoc employee experimentation without risking proprietary portfolio data or regulatory breaches. Allocators managing billions across private equity, hedge funds, and real assets process reams of unstructured data, including manager tear sheets, meeting notes, and quarterly letters. Securing the enterprise perimeter (Layer 1) lets teams connect models to tools like Slack for manager meeting prep without leaking track records or pipeline intelligence.

The system breaks down when investment teams treat AI synthesis as an analytical shortcut for intermediate deal analysis. Heller learned this firsthand when Brandywine tested automated mid-diligence updates between initial issue memos and formal investment committee memos. “The thought was that the lead or associate director could pull together this update using AI in an hour,” Heller noted. "It was a failure." Large language models excel at retrieval and style adjustments, but they struggle to synthesize intermediate investment conviction where the real analytical work happens in the gaps between the data points. If a junior investor cannot defend the underlying assumptions in a cross-examination, provenance tracking (Layer 3) exposes the workflow failure immediately.

Why It Matters

Top allocators are shifting from asking what models can write to mapping where model output pollutes diligence records. As LPs interrogate GP operational efficiency, institutional allocators are establishing the boundary between mechanized data retrieval and human investment judgment. Heller's model signals that institutional capital values data provenance as highly as the analytical output itself. Firms that fail to track provenance risk presenting automated consensus as proprietary conviction, diluting their underwriting edge.