Q: What is in-house AI for proxy research and why are firms adopting it? A: In-house AI for proxy research uses retrieval-grounded models to convert long proxy statements and filings into fast, traceable votes and to apply firm policies consistently across large voting universes. Firms are adopting it to increase speed, accountability, and consistency while keeping final judgement with stewardship teams. Q: How does retrieval-augmented generation (RAG) improve accuracy in proxy research? A: RAG retrieves the exact passages from filings and generates answers only from that evidence, providing explicit citations and returning a no-source-available result when information is missing. This grounding reduces hallucination and improves reliability compared with general-purpose models that can produce fluent but inconsistent outputs on long, complex documents. Q: What audit and logging elements should an auditable in-house AI for proxy research platform include? A: To meet compliance expectations every output should link to the triggered policy rule and parameters, the specific filing sections used as evidence, the model version and prompt template, and the user, timestamp, and any overrides. Immutable decision logs and citation-quality links that land at the right clause, table, or footnote are also necessary to demonstrate traceability. Q: How can systems reduce “context noise” when processing long proxy filings? A: Implement multi-stage retrieval that classifies the question first, constructs a query plan, and ranks and filters results so the model receives only the most relevant evidence. Structured extraction that converts tables, charts, images, and footnotes into text further reduces irrelevant context and improves downstream accuracy. Q: What end-to-end evaluation metrics should stewardship teams track? A: Teams should measure groundedness, temporal awareness, citation quality, and policy fidelity across the entire pipeline rather than evaluating only the model component. Building a benchmark set from real stewardship questions and running regression tests in CI/CD helps surface inconsistencies and prevent performance regressions. Q: Which proxy-voting rules can be automated by in-house AI for proxy research and which require human judgement? A: Objective, measurable rules—such as pay increases above a defined threshold, attendance or overboarding limits, or combined CEO/chair checks—are well-suited to automation because they are testable and auditable. Subjective calls about materiality, controversy, or strategic judgement should remain with analysts, with the system surfacing facts and voting history to inform their decision. Q: How should teams monitor and control model drift and system changes over time? A: Maintain versioning for models, prompts, retrieval settings, and policies, and set up drift alerts when accuracy, groundedness, or latency moves outside defined bounds. Use shadow runs to compare new and old systems before release, keep clear change logs, and have rollback plans to recover from failures. Q: How can custom voting policies be applied at scale and validated before going live? A: Allow teams to define plain-language rules that compile into precise checks, enable market or sector overrides and rationale templates, and backtest those rules across issuers, markets, and years to see historical impact. Comparing backtest results against benchmark advisors and historic votes helps governance assess likely outcomes and reduces surprises.