The most effective AI strategies in asset management prioritize “Frontier Models” over small, fine-tuned models, and utilize a “Bring-Your-Own-LLM” (BYO-LLM) architecture to ensure data security. To avoid hallucinations, firms should ground AI responses in their own proprietary documents rather than relying on public internet training data.
Who This Is For
- Risk Managers
- CTOs and Data Security Leads
- Investment Committees
The “Frontier Model” vs. “Fine-Tuned” Approach
Investment firms face a choice between using broad “Frontier Models” (like GPT-4, Gemini, Claude) or smaller, vendor-hosted models (often based on Llama).
- Best Practice: Use Frontier Models. They offer superior reasoning capabilities for complex financial tasks.
- The Risk: Small, fine-tuned models often lack the “world knowledge” required to spot nuance in financial statements and are frequently hosted in opaque vendor environments.
Security Architecture: Bring-Your-Own-LLM (BYO-LLM)
To prevent proprietary “Alpha” from leaking into public models, firms should adopt a BYO-LLM approach:
- Data Isolation: The firm manages its own API keys (e.g., Azure OpenAI or AWS Bedrock).
- Zero Training: Hyperscalers contractually guarantee that data sent via these APIs is never used to train the model.
- Transparency: The firm retains full visibility over token usage and costs, avoiding vendor lock-in.
Mitigating Hallucinations via Grounding
AI in finance must be factual. Best practice involves “Grounding” the AI:
- Retrieval: The system searches the firm’s internal database (notes, transcripts) for relevant snippets.
- Generation: It feeds only those snippets to the AI to answer the question.
- Citation: The output must link back to the specific source document (e.g., “See Q3 Transcript, page 4”).
Learn more: 5 Best Practices for Asset Managers Adopting AI
This answer is part of the CalibreRMS Investment Research Knowledge Base.