AI in Proxy Voting and Governance: Resolution Analysis at Scale

Proxy season is a compression problem. A stewardship team of five might face 3,000 meetings and 25,000 ballot items inside eight weeks. The honest description of how most institutions cope: follow the proxy advisor's recommendation on everything except a shortlist of high-profile names. That is outsourced judgment wearing a policy's clothes. We build AI operations systems for document-heavy teams, and proxy analysis is one of the most natural fits we've worked on — because a voting policy is already a written standard, and reviewing documents against a written standard is exactly what agents do well.

The Core Pattern: Resolution vs Policy

In production, we run document review against standards as a routine automated action — a draft checked against a playbook, clause by clause, with deviations flagged. Proxy voting is the same architecture:

1. Ingest the meeting agenda, proxy statement, and supporting filings.
2. Classify each resolution (director election, say-on-pay, share issuance, shareholder proposal, etc.).
3. Apply the house voting policy for that category — encoded as explicit, testable criteria.
4. Output a recommended vote with the policy clause cited and the relevant disclosure quoted.
5. Route to a human: routine items in a fast-review queue, exceptions escalated with full context.

The output that matters is not the recommendation — it's the citation. An analyst can verify a quoted policy clause against a quoted disclosure in 30 seconds. Verifying an unexplained "vote against" takes 30 minutes.

Where the Volume Actually Sits

Roughly 70% of ballot volume is mechanical policy application. Freeing analysts from it is not a marginal gain — it returns the majority of proxy season to the 5% of items where stewardship actually happens.

Consensus on the Contested Items

For contested votes — proxy fights, controversial shareholder proposals, major M&A — a single model's reading is not enough. We use multi-model consensus: the same resolution and policy analyzed independently by multiple models, conclusions compared. Agreement gives the analyst a strong starting brief; disagreement is itself signal that the item deserves committee time. In our experience the disagreement rate lands in the low single digits — a very effective triage filter.

The AI never votes. It applies the written policy, shows its work, and surfaces the cases the policy doesn't cleanly cover. The vote — and the accountability — stays human.

The Audit Trail Is the Product

Stewardship teams answer to clients, regulators, and increasingly to the public record. Every recommendation the agent produces carries: the policy version applied, the disclosures relied on, the model outputs, the confidence, and the name of the human who approved. When a client asks "why did you vote for this pay package," the answer is one query, not an archaeology project. In several engagements, this reporting capability — not the time savings — was what got the project funded.

Rollout Without Betting a Proxy Season

• Off-season: encode the voting policy as testable criteria. This exercise alone usually exposes ambiguities in the written policy worth fixing.
• Backtest: run the agent over last season's meetings and compare against actual votes cast. Investigate every mismatch — some are agent errors, some are inconsistencies in past voting.
• Shadow mode through the first live season: recommendations generated but analysts vote as before, agreement measured weekly.
• Assisted mode the following season: routine items batch-approved from the queue, analyst time concentrated on escalations.

The Point

The scandal of proxy voting was never that institutions vote wrongly — it's that most items are never genuinely analyzed at all. Agents make universal, policy-consistent, documented analysis economically possible for the first time. That is not a threat to stewardship teams. It is the first tool that lets them do the job as described in their own policy documents.