1. AI can identify where to look. Judgment is harder to automate.

Several participants drew a distinction between processing research and exercising judgment.

AI can help with the more structured parts of manager research: extracting information, completing checklist-style reviews, identifying missing information, and surfacing areas that warrant attention.

It can potentially reduce a universe into a shortlist.

But participants were much less convinced that it can replace the judgment that follows.

Negotiating with a manager, assessing the significance of an issue, understanding why something feels inconsistent, or deciding whether a risk is acceptable requires context that may not exist explicitly in the available materials.

One participant described an experiment using ChatGPT to conduct investment-process diligence on a cyclical company. The output looked convincing, but it missed an important risk because that risk was not contained in the materials the model had been given.

Such experiences reinforce the user view that AI can reason across what it can see. Investment judgment often depends on recognizing what is missing.

2. AI use may be moving through three different stages

One useful framework from the discussion was to think about AI adoption across three levels.

Efficiency

Meeting-note summarization, transcription, document extraction, financial-statement review, drafting, and memo preparation.

These are increasingly established use cases.

Analysis

Reviewing diligence materials, comparing information, identifying potential risks, finding inconsistencies, and directing an analyst toward the areas that warrant deeper investigation.

This is where many research teams appear to be experimenting now.

Decision-making

Recommending a fund, taking an investment action, or determining that a particular risk should change an investment conclusion.

This remains a materially higher threshold.

The further AI moves across those three stages, the greater the need for validation and governance.

3. The important AI diligence question may become: when did it fail?

One participant raised a particularly useful question.

If a human analyst makes a mistake, regulators and investment committees understand that professional judgment is imperfect.

What happens when a machine makes the mistake?

That led the group toward a broader question about how allocators should diligence AI use by asset managers.

It may no longer be enough to ask:

Do you use AI?

More revealing questions may be:

• How do you test AI-generated outputs?
• When has the system produced a materially incorrect answer?
• Which workflows require independent verification?
• Who determines whether an output is good enough?
• How do you monitor performance over time?

Where AI supports a critical workflow, evidence of testing may eventually become as important as the AI policy itself.

Our Generative AI Due Diligence Questionnaire includes questions across governance, data access, model use, validation, continuity, and human oversight.

4. Who teaches AI what good judgment looks like?

An interesting debate emerged around configuring AI to support investment research.

One allocator described teaching AI which factors to review and what should constitute a risk.

Another described a very different research philosophy: asking three analysts to assess the same manager independently because the diversity of their conclusions matters.

Those approaches expose an important design question.

If an AI research system is trained primarily on the judgment of one senior investor, does it institutionalize expertise - or institutionalize one person's biases?

As firms build research agents, prompts, scoring frameworks, and evaluation criteria, diversity of thought may need to become an explicit design consideration.

AI can make research more consistent.

That does not necessarily mean research should become uniform.

5. Critical thinking may become more valuable as AI does more junior work

The conversation also returned to a theme we have heard in other markets: what happens to the analyst development model?

Twelve months ago, there was significant discussion about using AI to increase research capacity without adding junior analysts.

Some participants now see that view reverting in some cases and becoming more nuanced in others.

Junior work has never only been about producing a spreadsheet or drafting the first memo.

It is how analysts learn to notice inconsistencies, understand business models, ask follow-up questions, and eventually develop judgment.

If AI increasingly performs the first pass, firms may need to become more deliberate about teaching those skills.

For new graduates in particular, knowing how to question an AI-generated answer may matter more than knowing how to generate one.

The risk is not simply that AI produces a bad answer.

It is that the person reviewing it does not know enough to recognize one.

6. Adoption still differs meaningfully by region and investment strategy

Participants saw meaningful geographic differences.

There was a perception that UK and European firms remain more hesitant about certain AI applications than US firms, particularly where confidential information, secrecy laws, and data residency are involved.

Luxembourg was mentioned as an example of a jurisdiction where confidentiality considerations can materially shape adoption.

US investment teams were perceived as somewhat more willing to experiment with AI in areas such as financial modelling.

There was also variation by investment strategy.

Systematic managers appear naturally positioned to use AI alongside existing quantitative processes.

Macro managers were cited as using AI to aggregate information, build dashboards, and monitor a much larger number of factors.

Some systematic teams are also beginning to experiment with task-specific agents, coordinated by an orchestration layer that effectively acts as a CIO agent.

For now, however, most of these use cases remain focused on expanding research capacity and improving efficiency rather than directly generating alpha.

7. Implementation and governance are critical to trust in AI

A recurring source of hesitation was not whether AI could produce impressive output, but whether that output could be trusted consistently.

Several examples illustrated the problem.

One ratings provider had used an AI research agent that produced a polished and compelling report on a fund that was, in fact, deteriorating.

Another participant described very poor AI-generated output from a technology provider that had moved quickly to deploy AI without sufficient governance, testing, or evaluation.

Others noted the amount of repeated prompting still required to get some AI tools to the right outcome.

These examples point to a broader issue: impressive output is not the same as reliable output.

This is where AI governance becomes critical. Firms need clear processes for testing AI-generated outputs, identifying potential errors, validating results, and establishing human accountability. Without these safeguards, even sophisticated AI tools can introduce risks into research and investment workflows.

As AI adoption matures, the focus is likely to shift from simply deploying models to understanding their failure modes, testing them systematically, and building the governance needed to use them with confidence.

8. AI outcomes should improve as models and users learn

Participants were optimistic that outcomes will improve as both the technology and the people using it become more sophisticated.

Industry-specific models and tools, including Bloomberg's work with financial-domain LLMs, were seen as one path toward better context and more relevant outputs for investment professionals.

The other side of that equation is user learning.

As researchers become more experienced in structuring prompts, providing context, testing outputs, and understanding where AI performs well or poorly, the quality of collaboration with AI should improve.

For now, much of that learning remains informal. Most firms do not yet appear to have structured AI learning programs, so capability is often built through experimentation and knowledge-sharing within teams.

That is likely to change as AI becomes a more established part of the investment research workflow.

9. Recording, transcription, and proprietary information are becoming practical governance issues

Not every governance question involves sophisticated AI.

Some of the most immediate issues are much simpler.

Participants discussed the rapid adoption of meeting transcription and recording tools. Voice input can provide remarkably rich context to AI systems, and transcription can make institutional knowledge searchable.

But practices remain inconsistent.

Some asset managers prohibit recordings because discussions contain proprietary information.

Others routinely enable meeting transcription without first asking everyone present.

As AI captures more conversations, the boundary between convenience and institutional recordkeeping becomes increasingly important.

Questions around consent, retention, ownership, and access may become part of everyday research governance rather than purely legal or technology policy.

The bigger question

The discussion reinforced something we have now heard across several markets.

The debate is moving beyond whether AI can impact research workflows.

The interest was to observe how far firms should allow it to move along the chain:

Each step requires more context, more testing, and more accountability.

The firms who get this right may be those that become most deliberate about where automation ends and judgment begins.