{"slug": "ai-in-stewardship-a-strategic-framework-for-asset-managers", "title": "AI in Stewardship: A Strategic Framework for Asset Managers", "summary": "Will Goodwin, co-founder of Tumelo, argues in a new white paper that AI is reshaping asset manager stewardship as regulatory pressure mounts on the proxy-advisory duopoly. The report contends that firms building AI-integrated infrastructure can deliver cheaper, faster, and more bespoke voting policies, undermining the economic justification for the conflict of interest in proxy advice.", "body_md": "Will Goodwin is the Co-founder & Head of US Sales at Tumelo. This post is based on his Tumelo report.\n\n*This post is adapted from the Tumelo white paper, “AI in stewardship: a strategic framework for asset managers”. It argues that AI is reshaping the stewardship function just as the proxy-advisory duopoly comes under regulatory pressure. Written by co-founder Will Goodwin, it sets out how the firms that build the right infrastructure now — and connect AI to it the right way — will run functions that are cheaper, faster, more bespoke, and deliver better outcomes for clients.*\n\n## Introduction\n\nEvery industry is currently working out what AI means for it.\n\nStewardship is more exposed than most. The function is under increased operating pressure from several directions at once — complex ballot issues, expanding regulatory disclosure, RFPs and client-reporting work absorbing senior analyst time. On top of that, there’s a structural shift on the supply side: the proxy advisors that most asset managers have relied on are now under serious regulatory scrutiny — and for the first time, credible alternatives are emerging.\n\n### Why now: the forces reshaping stewardship\n\nThe pressure is partly internal — more disclosure, more client reporting, more demand on a fixed analyst headcount. But the sharper change is in the proxy-advisory market itself. Regulatory and political pressure on the established advisors has intensified over the past two years.\n\nIn December 2025 [an executive order](https://www.whitehouse.gov/presidential-actions/2025/12/protecting-american-investors-from-foreign-owned-and-politically-motivated-proxy-advisors/) directed the Securities and Exchange Commission (SEC), the Federal Trade Commission (FTC) and the Department of Labor to increase their oversight of the firms that control more than 90% of the proxy-advisor market: the SEC to reconsider the rules that govern proxy advisors, and to decide whether they should be made to register as investment advisors; the FTC to examine them for anticompetitive or deceptive practices; and the Department of Labor to revisit the fiduciary rules that govern their use. Together with in-house counsel increasingly treating reliance on a proxy advisor as an operational risk on the asset manager’s own register, this has turned a procurement decision that used to be routine into a strategic one.\n\nThe supplier relationship most functions have built around is being questioned — in some cases displaced — and the pressure is coming from regulators and clients alike.\n\n### AI promises to solve the conflict of interest in proxy advice\n\nFor two decades, the proxy advisory industry has operated as a duopoly. A small number of firms produce the voting recommendations that more than 90% of asset managers rely on, the custom voting policies those managers implement, and — critically — the research, data and services they sell to the very issuers whose proposals are being voted on. The firms advising funds on how to vote are also earning money from the boards being voted on; in fact, it is their profit centre.\n\nBut now regulators have stopped treating this conflict of interest as a structural quirk, and started treating it as an enforcement target. In-house counsel at large asset managers are now classifying proxy advisor reliance as a material risk. The exposure used to sit on the proxy advisors; it now sits, at least partially, on the asset managers who depend on them.\n\nWe’ve heard from the market that the proxy advisors’ historic defence has been an economic one: producing custom voting policy at scale was expensive, and the issuer-side revenue is what made it viable. Take that revenue away, the argument runs, and the custom service managers actually want would not exist.\n\nAI changes this. The cost of delivering a custom policy collapses by an order of magnitude. The economic argument that justified the conflict of interest disappears.\n\nThis is relevant for stewardship leaders in two ways. One, it reframes the AI conversation: this is not only an efficiency story, it is the way out of an increasingly untenable dependency on conflicted vendors. Two, the asset managers who build their internal capability now — clean voting infrastructure, addressable data, constrained retrieval over filings, evaluation infrastructure — will be operating a stewardship function that is cheaper, more effective and more defensible than one held together by proxy advisor subscriptions.\n\n### The capability shift in AI models\n\nAI has developed significantly since the launch of ChatGPT in 2023. The models have reached a stage where meaningful work can now be accomplished.\n\nThe shift that most impacts stewardship is retrieval and verification. The accuracy that makes AI usable here does not come from a frontier model on its own — ChatGPT or Claude out of the box will not reliably answer a multi-document stewardship question, and treating “point an LLM at it” as the strategy is how firms get it wrong. It comes from the system built around the model: constrained retrieval, multi-stage retrieval, automated evaluations and source verification, working together. Anything the system is unsure of is elevated to a human for review rather than taken on trust.\n\n### The core idea: stewardship infrastructure, a central LLM, and the outputs it generates\n\nTo benefit from models that keep improving, asset managers can deploy a central LLM, such as Claude or ChatGPT. But an LLM is only as good as the context it can reach, so it needs access — through the Model Context Protocol (MCP) — to the systems holding data and carrying out decisions. For stewardship, this means connecting that central LLM to the infrastructure that lets you decide, execute and report on votes. There are three layers:\n\n**Stewardship infrastructure**— the durable, regulated systems the function runs on: voting-policy rules, investment-grade data, voting and trade pipelines, audit records. You build this once and keep it. This is the layer Tumelo provides.**The LLM orchestration layer**— the central LLM itself, connected through MCP to that infrastructure, deciding what to retrieve, reasoning over it, and doing the work.**Outputs**— the RFP responses, regulatory submissions, engagement notes and vote rationales the LLM generates on the fly from that infrastructure. Nothing here is built and stored; it is generated when asked for, and regenerated when the data moves.\n\n**The stewardship infrastructure: what you build once and keep**\n\nThere are four components that make up the infrastructure:\n\n**The voting policy rules engine**. It must be flexible enough to mirror your policy and robust enough to reproduce the same answer every time. Every decision must be reconstructable: which policy applied, which rule fired, which data point was used, and which source that data came from. This is non-negotiable for Stewardship Code disclosure, SRD II reporting, and the rising bar of client scrutiny. An audit trail bolted on afterwards is not acceptable; it has to be a property of the system.\n\nThe design principle that makes this compatible with regulated infrastructure is the separation of objective from subjective: custom policies are expressed as objective rules, so when a meeting comes up the system does not ask the LLM how it should vote — it applies the rule, deterministically and the same way every time. Three things follow:\n\n**Model risk.** Because the decision logic is deterministic, auditors and regulators can fully inspect the audit trail from triggered rule to source document.**Defensibility.** A contested vote can be traced end to end: the rule that fired, the data that triggered it, the filing the data came from.**Change management.** When a policy needs to evolve, the change is to the rule, not to a prompt. It can be tested, reviewed, and versioned in the way any other piece of regulated infrastructure is.\n\n**Investment-grade data** — the information you need to decide how to vote: accurate, with every figure traceable to the source; timely, refreshed at a cadence that matches the work and its deadlines; and carrying enough metadata, including stable identifiers, to be joined across systems and reconstructed years later. When this data is extracted by AI rather than assembled by analysts reading filings every season, it can be built at a fraction of the cost and more accurately — but only if managers can rely on it, and reliance comes from infrastructure, not from the model. Four layers provide it, and each is independently auditable: constrained retrieval over the filings; automated evaluations that run on every output; an independent AI review, in which a second agent opens each citation and checks it against the source; and a human reviewer, to whom anything the system is unsure of is escalated.\n\nThis is the area of AI engineering that has matured most rapidly in the past 18 months, and it is where the gap between the best systems and the weaker ones is widest. A system without sophisticated retrieval engineering will produce accuracy figures in the low-to-mid 80s on multi-document analytical questions; a system with it can produce accuracy figures above 99%. For a buyer of stewardship data, the implication is that “AI-powered” is now too broad a label to be useful — two providers using the same marketing language can be 15 accuracy points apart.\n\n**Voting pipelines** — the route a vote travels to the company. Global vote submission requires integration with custodians whose systems vary by market, jurisdiction and client, and once connected they must be monitored and maintained. None of this is work an LLM should be doing ad hoc. It is infrastructure you build, test rigorously, and trust to keep working.\n\n**The vote records database** — the system of record for every ballot cast and the rationale behind it. Each entry is date-stamped and linked to the rule that fired, the data that fed it, the source filing, and the person who submitted or edited it. This is what makes the rest of the infrastructure defensible; regulatory and client scrutiny both depend on it.\n\nThese four components are built once, carefully — with the care that goes into anything that must withstand a decade of scrutiny — then owned and maintained, not regenerated.\n\n**The central LLM and MCP**\n\nConnecting an LLM to internal systems poses a real challenge. Unrestricted API access is too permissive — the equivalent of handing the LLM a developer’s login, and most legal teams will rightly refuse it. Screen-driving, letting the LLM click through an app like a person, is too fragile and inherits everything a logged-in user can do. MCP is the standard that has emerged: the LLM is given a defined list of permitted actions — read filings, read voting records, write engagement notes — and can do those and nothing else. Every action is logged; every permission can be revoked, and the firm can show after the fact exactly what the LLM was allowed to do and what it did. That is a position regulated firms can live with.\n\n**The outputs the LLM generates**\n\nA stewardship team produces a large amount of outputs: an RFP answer, a Stewardship Code submission, a quarterly client report, a rationale for a contested vote. These have to be accurate, but they are not infrastructure — each is produced once, for one audience, and superseded by the next version. Once the infrastructure is solid and the LLM can access it and act on your behalf, producing outputs is a request, not a project: the analyst asks — answer these RFP questions — and the LLM assembles the document from the infrastructure and cites every figure back to its source. Next quarter, when the numbers have moved, they regenerate it rather than rebuild it. Outputs cost almost nothing, because the expensive part — trustworthy, cited, current data — was already built once.\n\nTwo examples show this in practice.\n\n**Answering stewardship questions for clients and reporting.** A wide range of recurring work — regulatory and Stewardship Code reporting, ad hoc client requests, and RFP questionnaires — comes down to the same task: *how many ballots did we vote, on what share did we vote against management, how does that differ for this client’s holdings versus the firm-wide book, how many engagements did we hold and with whom.*\n\nHistorically, an analyst exported voting records into Excel, rebuilding pivot tables, and answering slightly differently each time because the work is redone from scratch.\n\nThe new way: the analyst points the firm’s LLM at the stewardship infrastructure — rules engine, voting records, engagement data — and asks it to answer; the LLM pulls the data, applies the right cuts (firm-wide vs client, full-year vs year-to-date), and returns each answer with citations back to source. The analyst reviews and edits where judgement is needed.\n\nThe gain is consistency as much as speed: answers trace back to the records and don’t drift between analysts or need re-defending under scrutiny.\n\n**Refining a voting policy with a backtest.** Teams iterate on policy every off-season — adjusting thresholds, tightening guidance, responding to client preferences.\n\nThe question that should sit at the heart of every change — *what would happen if we treated any director with more than four other public-company boards as overboarded, instead of five?* — has historically been very hard to answer with evidence; a real backtest meant a request to the proxy advisor, weeks of waiting, and a bill.\n\nWith the rules engine and voting records connected to the LLM, the analyst describes the change in plain language and the system generates two or three variants, runs each against last season, and returns a short comparison against the firm’s actual votes, broken down by sector, geography and issuer size.\n\nA backtest that used to take months and an external supplier now takes an afternoon and an internal decision.\n\n**The opportunity**\n\nThe pay-off from getting this distinction right is substantial, and the strategic decisions that have always been hardest for stewardship leaders become straightforward. Build or buy decisions all flow from one test: is this part of the stewardship infrastructure, or is it an output? Stewardship infrastructure is where you buy the best provider you can find and integrate it deeply. Outputs sit on top of it, generated on demand by the LLM with the right access. Answer that question for each part of the function and the operating model largely writes itself.\n\nGet that line wrong and you fail in one of two familiar ways. The first is paralysis — treating every output as if it had to meet the standard of a regulated system: three-year roadmaps, stacked procurement and compliance cycles, nothing shipped. The second is fragility — handing the LLM a job it should never hold, like building the voting database, owning vote-execution logic, or keeping the audit trail. That system sparkles in a demo and fails on the audit. Drawing the line between the infrastructure and the rest is what avoids both.\n\n**Closing**\n\nAI is transforming how an asset manager operates. Strategies to take advantage of it involve deploying a central LLM connected across the firm’s data and providers, enabling efficiencies, removing the drudgery, and ultimately elevating teams to deliver better outcomes for clients. With the shift in regulation happening at the same time, stewardship teams now need to think about how to position themselves for the next five years.\n\nThe asset managers who draw the line in the right place will run stewardship functions that are cheaper, faster, more bespoke, and deliver better outcomes for clients. The architecture is already being built, by providers who can see what the next decade of stewardship will look like.", "url": "https://wpnews.pro/news/ai-in-stewardship-a-strategic-framework-for-asset-managers", "canonical_source": "https://corpgov.law.harvard.edu/2026/07/15/ai-in-stewardship-a-strategic-framework-for-asset-managers/?utm_source=rss&utm_medium=rss&utm_campaign=ai-in-stewardship-a-strategic-framework-for-asset-managers", "published_at": "2026-07-15 11:30:30+00:00", "updated_at": "2026-07-15 11:51:10.804200+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-policy", "ai-tools"], "entities": ["Tumelo", "Will Goodwin", "SEC", "FTC", "Department of Labor"], "alternates": {"html": "https://wpnews.pro/news/ai-in-stewardship-a-strategic-framework-for-asset-managers", "markdown": "https://wpnews.pro/news/ai-in-stewardship-a-strategic-framework-for-asset-managers.md", "text": "https://wpnews.pro/news/ai-in-stewardship-a-strategic-framework-for-asset-managers.txt", "jsonld": "https://wpnews.pro/news/ai-in-stewardship-a-strategic-framework-for-asset-managers.jsonld"}}