Section 01

AI adoption has moved beyond experimentation, but maybe not yet to standardization

Participants described active AI use across both investment and operational functions. On the operational side, examples included reconciliation support, report preparation, drafting communications, and speeding up recurring administrative work.

Within investment teams, AI is increasingly showing up in the first review of manager documents, meeting summaries, investment materials, valuation comparisons, performance attribution, and risk reporting. A common comparison was that AI can handle something like the first pass a junior analyst might traditionally have done: processing a large information set, organizing it, and preparing an initial output for someone else to review.

That doesn't mean these workflows have become institutionalized. In a number of organizations, individuals appear to be using AI well without sharing their prompts, agents, or lessons learned with anyone else. Teams may end up solving the same problem more than once. The gap looks less like access at this point, and more like repeatability.

Section 02

The harder constraint may be shifting from the model to the architecture around it

The discussion moved fairly quickly past comparisons among individual models. Most participants expected firms to use several models, chosen by task, required reasoning, data sensitivity, cost, and internal controls. The more open-ended questions were about what sits around the model.

Some organizations are building data and integration layers using platforms such as Snowflake, Databricks, APIs, and internally developed connectors, with several approved sources already feeding enterprise AI environments. Others are more cautious, with connector access restricted over security concerns. A rough shape emerged: trusted data → governed access → appropriate model → investment workflow → human review. The model matters. But increasingly, it's the system around it that seems to determine whether it can be used reliably at institutional scale.

Section 03

Model choice is expanding, and governance may need to keep pace

Participants noted that the set of viable models is widening, including capable international entrants alongside established US providers. Broader competition was generally seen as good for performance, cost, and innovation, though this wasn't a universal view.

The more consequential question, for several in the room, wasn't which model performs best. It was where data goes once a model is chosen. Some organizations were unwilling to send institutional data to infrastructure outside approved jurisdictions, regardless of how capable the model itself was. A model's origin and its deployment environment are separate questions, and firms may need controls that address both.

One takeaway, offered tentatively: governance may work better if it supports model choice rather than assumes any single provider or geography stays dominant indefinitely. That's consistent with the broader theme of the conversation, that the architecture around a model may matter more, over time, than the model itself.

Section 04

Institutional memory could become AI's most durable asset

One recurring concern was that useful knowledge tends to stay attached to individual users. An analyst develops a strong prompt. Another builds an agent. A third finds a reliable way to evaluate a document. Unless those methods are captured and shared, the organization may capture the productivity benefit without building anything that outlasts the person who created it.

Participants talked about the value of preserving approved prompts, reusable agents, evaluation criteria, research frameworks, output templates, workflow instructions, model-selection guidance, and examples of acceptable and unacceptable outputs. This seemed to matter most when people change roles or leave. The goal isn't just recovering the prompts someone used. It's retaining the reasoning behind them.

We explored a version of this idea in The Allocator's Moat: Defining Your AI IP: that the more durable advantage may not be the model itself, but the context, frameworks, and institutional knowledge built around it.

Section 05

Governance seems to be moving from policy toward operating practice

Most organizations represented were either drafting an AI policy or expanding an existing one. Policy on its own was generally seen as a first layer, not a finished answer. A few principles came up often enough to be worth naming.

A related idea: matching model strength to task. Several felt the strongest available model isn't always necessary, and that firms may benefit from a rough rubric for which tasks warrant lower-cost models versus which call for deeper reasoning or tighter controls.

Section 06

Senior professionals may need to help shape the AI standard

Prompt development is sometimes treated as a technical or junior-level task. Several participants pushed back on that framing. Senior investment professionals tend to have a clearer sense of what a good output should contain, which facts are material, which comparisons are misleading, and where an answer that looks complete is actually missing the point.

That judgment, several felt, needs to inform the prompts, templates, and evaluation criteria used more broadly, not just sit with whoever happens to be most comfortable with the tools. Analysts may become highly skilled at working with models. Senior practitioners still seem essential for defining what the standard should be. If that's right, training probably needs to run in both directions: younger team members helping senior professionals get comfortable with the tools, and senior professionals encoding the judgment that makes the tools worth using.

Section 07

Is part of the analyst budget becoming a token budget?

The hiring conversation was more nuanced than a simple prediction of job losses. Most participants didn't expect AI to eliminate investment teams outright, and several were explicit about wanting to keep junior analysts and make them more effective with AI, not replace them.

At the same time, some firms are rethinking where the next marginal dollar of research capacity should go. Historically, wanting to review more managers might have meant adding another analyst. For some participants, that capacity increasingly gets purchased through models, tokens, data infrastructure, and connectors instead. It's not that the analyst budget disappears. It's that a portion of its future growth may show up as an AI budget instead. A few participants already expected their analyst pools to stay roughly static even as the organization grows.

Section 08

Alpha may be earned around the edges

One idea that came up, worth sitting with rather than treating as settled: alpha tends to be earned around the edges, not in the consensus middle. If AI can help a firm automate and systematize the parts of research that are closer to consensus, that may free up capacity for the harder, more differentiated decision-making that's less amenable to automation in the first place.

That framing doesn't resolve the questions raised elsewhere in this piece, about succession, about judgment formation, about where the human needs to stay in the loop. If anything, it sharpens them. If the consensus work is what AI absorbs first, the differentiated work is what's left for people to do, and it may be the harder work to learn from a machine-generated first pass.

Section 09

A thinner analyst pool may raise a succession question

Junior investment work has traditionally done two things at once: produced analysis and developed the next generation of portfolio managers. Analysts learn, in part, by reading documents, building comparisons, chasing down inconsistencies, and occasionally getting things wrong. Much of that work is repetitive, but it also seems to be how pattern recognition and judgment get formed.

If AI takes on a large share of that foundational work, the industry could end up weakening its own development pipeline without quite meaning to. Where do future portfolio managers come from if fewer junior analysts are hired? What does succession planning look like once the traditional apprenticeship model shifts?

Nobody suggested simply preserving junior tasks for their own sake as the answer. A more likely path, several felt, is redesigning analyst development on purpose, perhaps rotating analysts into more decision-oriented work earlier, asking them to challenge AI conclusions, and weighing the reasoning behind a recommendation rather than just the speed of getting to one.

Section 10

What happens when AI starts investing: toward a diligence framework of its own

A related question, and a genuinely open one: not how investment teams use AI for research, but how allocators should diligence a strategy where AI materially influences investment decisions. There was little appetite in the room for treating "AI-driven" as a sufficient description of a process on its own.

Participants naturally reached for comparisons with systematic and quantitative managers. Much of it seemed to transfer: economic thesis, data quality, backtesting methodology, overfitting risk, model monitoring, key-person dependency. The difference, a few noted, is that traditional systematic strategies are usually built around code expressing a fairly stable investment philosophy. Generative AI and autonomous agents may introduce more uncertainty, since outputs can shift with model versions, prompts, and orchestration choices that the manager doesn't fully control.

"The objective isn't replacing investment philosophy. It's using AI to build and maintain code that stays consistent with it."

One systematic investor put it this way, and it seemed to land with the room. The philosophy, in this view, should precede the technology rather than follow from it. There was general caution about strategies that are entirely systematic without a meaningful human in the loop, though what "meaningful" means in practice wasn't fully resolved. It doesn't necessarily mean approving every trade. It probably means that identifiable people understand the system, govern how it changes, and remain accountable for the portfolio it produces.

Section 11

Wealth management doesn't look ready to remove the human, at least not yet

Participants drew a line between using AI inside an investment organization and placing clients into fully AI-run investment programs. The wealth segment came across as the most cautious of the group. Wealth clients, whether high net worth, ultra high net worth, or family offices, don't only buy portfolio construction. They tend to rely on a person to interpret changing circumstances, explain hard decisions, and offer some reassurance during uncertain periods. Trust there was described as relational, more than computational, though reasonable people may weigh that differently as the tools improve.

Participants seemed more comfortable with AI supporting advisors than replacing them: preparing meeting materials, flagging portfolio issues, summarizing client history, and cutting administrative work, freeing the advisor to spend more time on judgment and the relationship itself. The model that seems most likely for wealth, at least for now, isn't autonomous advice. It's a better-equipped human advisor.

Cost came up throughout, as a related thread rather than a standalone topic. Token limits and model pricing matter, but a few participants pointed out that the more useful question isn't what a token costs. It's closer to what it costs to produce a trusted, reviewable, reusable output, once data, connector infrastructure, human review, and the cost of fixing errors are all counted in.