By Will Goodwin, Co-founder, Tumelo.
I was recently in Boston for the Council of Institutional Investors' fall conference, where I had the honour of speaking on one of the mainstage panels about AI in proxy voting and stewardship. Michael McCauley of the State Board of Administration of Florida moderated, and he asked great questions, ones we hear most often from clients. I've written up how I answered them here, along with a few things I'd have said if we'd had longer.
Tumelo started in pass-through voting, so we spent years learning how votes travel through the proxy system and get submitted around the world. During that time, investors kept asking whether we could help them apply their own voting policies at scale. Our first AI prototypes were exciting, but they weren't accurate or reliable enough for decisions investors make under a fiduciary duty. It was only in late 2025 that the models got good enough for us to build something clients could depend on. That's when we launched ProxyBeacon, which provides AI infrastructure for stewardship.
"Where do you draw the line between deterministic tasks that AI can handle autonomously—such as flagging over-boarding or dilution threshold breaches—and complex qualitative evaluations that strictly require human analyst oversight?"
Honestly, that's more our clients' decision than ours. Our part is keeping the AI's job separate from the investor's. The AI extracts data from company filings, and every data point comes with a citation. The investor's own rules then decide the outcome. Something objective, like a limit on how many boards a director sits on, can be written as a rule and applied the same way every time. Where a policy leaves room for interpretation, the vote is referred to the stewardship analyst. If the model would have to make a judgement call, we don't let it make one.
Shareholder proposals show how this works. They're often nuanced and depend heavily on context. A few clients have written rules for them, but most use Beacon to pull out the relevant information and get it in front of the right analyst, who makes the call.
"In a fiduciary environment, hallucination risk is fatal. What specific technical guardrails—such as source-document visual citations or deterministic verification layers—do you demand from AI tools before trusting their output for voting records?"
There's a balance here between speed and reliability. You can pull information out of a filing almost instantly and present it to the user, but it won't always be right. We've been on a journey with this, in partnership with our clients. Really, our conclusion has been that there is zero appetite for error.
So rather than putting data on screen the moment it's extracted, we check every data point first. We only retrieve from the company's own filings, and software checks run on every output. A second AI agent then opens each citation and checks it against the source, and anything flagged goes to one of our data analysts.
When the data lands, you can click on any citation and the filing opens at the passage it came from. Every vote can be traced back to the rule that was triggered, the data point it used, the filing it came from and the person who approved it. The team can change any vote before it's sent.
What's been really interesting is how much this makes teams get precise about what they mean. Take a rule that says "refer the vote to the team if CEO pay is excessive". What does excessive actually mean? Is it relative to a peer group, and if so, which one? Does the company's performance come into it? How do you think about one-off awards? We see clients working through questions their written policy never had to answer, and they come out with a policy they can explain rule by rule.
Once it's written, they can backtest it, which means running it across previous proxy seasons to see how it would have voted. That tells a team roughly how many votes it will need to review before the season starts, which makes planning much easier.
"Standard proxy advisory reports often accept company-selected peer groups that can distort pay-for-performance alignment. How are you leveraging AI to dynamically evaluate compensation packages against investor-defined, custom peer groups across thousands of global issuers?"
Our first approach was to let users pick a peer group in real time. It turned out that's a lot to ask of someone in the moment, and it needs a great deal of specialist knowledge. So now clients define peer groups by rules instead, using GICS codes (the Global Industry Classification Standard), relative performance or whatever criteria suit them. We call this a custom framework: the client's own definition, set once and applied wherever their rules need it.
A nice side effect is how useful it has become for engagement. Before meeting a company, a team can run a report on a peer group it has defined and see how that company compares. It means a team can say to a board: this company is doing it, so why aren't you?
For me, the answer is earlier engagement. When policy outputs arrive soon after a company files, a team can raise questions with the company well before the vote, instead of rushing in the last few days of a busy season. Several of us on the panel made the point that those conversations are where human judgement counts most.
ProxyBeacon is also designed to work alongside investors' own systems and AI environment, so a team preparing to meet a board can draw on its voting records and research alongside its own notes.
It's still early days, but I can see this helping stewardship teams do more for their clients. Often a client only wants a few things reflected differently in the policy. In ProxyBeacon, an asset manager can copy its own policy and add, remove or edit rules for that client's mandate.
Imagine turning up to a client meeting having already compared their voting policy with yours, and being able to say, "here's how we could adjust things to match the outcomes you're looking for".
This was the last question. Companies are starting to use AI to draft their filings, so more and more, AI is reading what AI has written. For us, keeping extraction accurate as filings and models change is part of the job. For a stewardship team, I think it comes back to build or buy. Do you want to be worrying about this yourselves, or would you rather someone else did?
Thanks to Michael and CII for having me, and to my fellow panellists for a good discussion. If you'd like to see how a policy is written as rules, or how the checks work on a live meeting, get in touch.