The Next Generation of Investment Research: Human + AI Agent Teams

TL;DR

Investment research is moving beyond the one analyst + one AI assistant model.

The next step is Human + AI Agent Teams: specialized agents take responsibility for different parts of the research process—financial analysis, news, valuation, competitive research, risk, cross-checking, and drafting—while humans set the objectives, inspect the evidence, challenge conclusions, and make the final investment decision.

The real opportunity is not simply to generate investment memos faster. It is to redesign the research organization itself so humans spend more time on judgment and less time on repetitive research and coordination.

In the Wrightery demo, this workflow is built without writing a single line of code.

Video demo:
https://youtu.be/B1C64h9iMpE?si=3Ce68MToohGKrOpH

Investment research has always been a team effort.

A strong investment decision rarely comes from one person asking one question and producing one answer. It usually involves different people looking at the same opportunity from different angles: financial analysis, industry research, valuation, news, competitive positioning, risk, and portfolio context.

Generative AI has already made many of these tasks faster. Analysts can summarize filings, review transcripts, draft research notes, and explore investment ideas in minutes.

But most of today's AI workflows still follow an old pattern:

One person + one AI assistant.

That is useful, but it is only the first stage.

The next generation of investment research will increasingly be built around a different model:

Human investors working with teams of specialized AI agents.

Instead of asking one general-purpose AI to do everything, different agents can take responsibility for different parts of the research process, work in parallel, challenge one another, and bring their findings together for human review.

The result is not simply a faster chatbot.

It is a different way to organize research.

From AI Assistant to AI Research Team

The first wave of generative AI in investment research has largely been about individual productivity.

An analyst reads a filing and asks AI to summarize it.

An analyst reviews an earnings call and asks AI to extract the most important changes.

An analyst drafts a memo and asks AI to improve the language.

These are valuable use cases. But the analyst still owns almost every step of the workflow.

They still decide what to research, move between data sources, ask follow-up questions, compare conflicting information, organize the evidence, challenge the thesis, and assemble the final report.

AI helps the analyst work faster.

A Human + AI Agent Team changes the model.

Now, rather than using AI only when a person asks a question, the research process can be divided among specialized agents with clearly defined responsibilities.

One agent can focus on financial statements.

Another can monitor news.

Another can analyze competitors.

Another can examine valuation.

Another can deliberately build the bear case.

Another can review the work of the other agents and look for unsupported claims, contradictions, or missing evidence.

The human analyst or portfolio manager is no longer responsible for manually executing every research step.

Instead, the human increasingly defines the objective, sets the standards, reviews the evidence, challenges the conclusions, and makes the final judgment.

What a Human + AI Agent Research Workflow Looks Like

I recently built a demo in Wrightery to explore what this model can look like in practice.

The workflow starts with investment idea generation and continues through research, challenge, cross-checking, and memo creation.

A candidate investment is first identified for deeper investigation.

Specialized agents then examine the opportunity from different perspectives and use different data sources. Rather than asking one model to generate a single broad answer, the work is distributed across agents designed for specific research responsibilities.

The next step is especially important.

Some agents are not asked to produce more analysis. They are asked to challenge the analysis that already exists.

They look for weak assumptions.

They search for evidence that contradicts the thesis.

They identify risks that may have been overlooked.

They ask whether an apparently strong claim is actually supported by the underlying data.

Finally, the research is brought together into a structured investment memo for human review.

The portfolio manager or analyst can inspect the supporting evidence, challenge the conclusions, and decide whether the thesis deserves capital.

Video demo:
https://youtu.be/B1C64h9iMpE?si=3Ce68MToohGKrOpH

The important point is not that several AI agents appear on the screen.

The important point is that research responsibilities are being divided between humans and AI in a new way.

Why One AI Should Not Have Every Job

Large language models are general-purpose systems. A single model can perform many tasks.

But being capable of many tasks does not mean every task should be combined into one prompt.

Investment firms do not normally ask one employee to be the financial analyst, industry expert, news researcher, valuation specialist, risk officer, and portfolio manager at the same time.

There is a reason research teams specialize.

Different responsibilities require different data, tools, instructions, standards, and ways of thinking.

The same principle can be applied to AI.

A financial-analysis agent may be connected to structured financial data and instructed to focus on revenue quality, margins, free cash flow, balance-sheet strength, and earnings trends.

A news agent may focus on discovering recent developments and separating facts from interpretation.

An industry agent may examine competitors, market structure, and structural trends.

A valuation agent may compare multiples, historical ranges, and alternative scenarios.

A bear-case agent may be explicitly designed to disagree.

And a reviewer agent may have a very different job:

Do not create a new thesis. Find what is wrong with the existing one.

That specialization can make AI research more useful because it introduces structure, accountability, and deliberate disagreement into the workflow.

Challenge May Be More Valuable Than Generation

Much of the current discussion around generative AI focuses on its ability to produce things.

Generate a summary.

Generate a report.

Generate a valuation.

Generate a memo.

But in investment research, producing more content is not necessarily the hardest problem.

The harder problem is often determining whether the content deserves to be trusted.

A convincing investment thesis can still be wrong.

A beautifully written memo can still rest on weak assumptions.

A confident answer can still be based on incomplete or outdated evidence.

That is why I think some of the most valuable AI agents may eventually be the ones designed to challenge rather than generate.

They can ask:

• What evidence does this claim depend on?

• What information would invalidate the thesis?

• Which assumptions have not been tested?

• Is the conclusion supported by primary data?

• What has changed since the original research was produced?

• What would a skeptical investor argue?

• Are different data sources telling the same story?

This begins to resemble one of the strengths of a good investment team: different people are encouraged to look at the same opportunity differently.

AI agent teams can make that kind of structured disagreement much easier to create and repeat.

Evidence Matters More as AI Does More Work

There is a danger in using generative AI for research: good writing can look like good reasoning.

When an AI produces a polished paragraph, it is easy to focus on the conclusion and forget to ask where the evidence came from.

That becomes more dangerous as AI is given more responsibility.

A serious research workflow should make it possible to move backward from the final conclusion:

Conclusion → Claim → Evidence → Source

If an analyst disagrees with the conclusion, they should be able to inspect the evidence.

If an agent identifies a risk, the analyst should be able to see why.

If two agents disagree, the difference should become visible rather than being blended into one smooth paragraph.

The goal should not be to hide the AI process behind a polished answer.

The goal should be to make the research process more transparent and inspectable.

That is especially important in investment management, where the final decision may involve real capital and real accountability.

The Human Role Becomes More Important, Not Less

It is tempting to look at a multi-agent workflow and assume the goal is to automate the investment professional.

I see the opportunity differently.

The purpose is not to remove human judgment.

It is to use AI so that human judgment is applied where it matters most.

A portfolio manager should not have to spend most of their time searching through documents, copying numbers, organizing sources, formatting reports, or repeatedly checking the same information.

Those are tasks machines can increasingly help with.

Humans bring something different.

They decide which questions matter.

They understand portfolio context.

They recognize when a company does not fit a historical pattern.

They interpret management credibility.

They understand risk tolerance.

They make judgments when the evidence is incomplete.

And ultimately, they are accountable for the investment decision.

So the model I find most compelling is not:

AI replaces the analyst.

It is:

AI expands the research capacity of the analyst.

The AI agents do more of the research, monitoring, analysis, cross-checking, and drafting.

The human investor focuses more of their time on judgment, questioning, portfolio decisions, and the things that are difficult to automate well.

The Workflow Can Be Designed by the People Who Understand the Work

There is another reason this model is becoming interesting.

Historically, building a custom research workflow with multiple data sources, AI models, and automation required a software development team.

That created a gap between the people who understood the investment process and the people who could build the technology.

In the Wrightery demo, the workflow is assembled without writing a single line of code.

That matters because the best person to design an investment research workflow may not be a developer.

It may be the analyst who has produced hundreds of research reports.

It may be the portfolio manager who knows exactly which evidence matters before making a decision.

It may be the risk professional who knows which assumptions repeatedly cause problems.

If those domain experts can define agents, assign responsibilities, connect data sources, and specify how work should flow between them, AI becomes much more than a general-purpose productivity tool.

It becomes part of the operating model of the team.

Investment Research May Be an Early Example of a Much Bigger Shift

The pattern is not unique to investing.

Many forms of knowledge work have the same characteristics: multiple data sources, specialized expertise, repeated research, analysis, review, challenge, and a final decision that still benefits from human judgment.

Due diligence, market research, competitive intelligence, management reporting, compliance, RFP preparation, sales research, and strategic planning can all follow similar patterns.

That suggests a broader evolution in enterprise AI.

The first stage was:

Human + AI Assistant

The next stage is increasingly:

Human + Specialized AI Agents

And the more important long-term shift may be:

Human + AI Agent Teams

Humans define objectives, standards, and decision boundaries.

Agents perform specialized research and execution.

Other agents review and challenge the work.

Humans inspect the evidence, deal with exceptions, and retain accountability.

At that point, AI is no longer simply another tool inside an existing workflow.

It becomes part of how the team itself is designed.

The Question Is Changing

For the past few years, companies have understandably asked:

How can AI make our people more productive?

That remains an important question.

But I think a second question is becoming more important:

If we were designing this research process today, knowing that humans and AI agents can work together, would we design it the same way?

Probably not.

And that is why I believe the next generation of investment research will not be defined by a better chatbot or a bigger prompt.

It will be defined by Human + AI Agent Teams: people and specialized agents working together, each taking responsibility for the parts of the research process they are best suited to handle.

The goal is not simply to generate investment memos faster.

The goal is to build a better research organization.