AI Solutions for the Buy Side: Research

Part 2 of a four-part series on the AI stack for the buy side.

Hello, this is part 2 of a four-part series on the AI stack for the buy side. The first post, covering screening, is here: AI Solutions for the Buy Side: Idea Generation.

We've worked with hundreds of buy-side clients over the past four years, and I'd like to share some of what we've learned: the current landscape for AI workflows on the buy side, and partly where I think things are headed. It's written for anyone at a public market equity fund searching for and comparing AI solutions for different parts of the research process.

Today I'll be covering research. I'm defining research as point-in-time report synthesis or fact-finding.

Job to Be Done

The job of research is hard to define in that there are so many purposes. An analyst conducting research could be learning, validating, or generating reports to communicate their findings. All of these jobs are being sped up by AI workflows, but AI-generated content — i.e. slop — can also slow you down and obfuscate important information.

How often is the general intuition surrounding a complex system the right one? When a teammate comes to you with a 20-page report, how long does it take to grok the underlying concepts? How much research, and what types, LLM-driven systems can be trusted with will vary between funds and between individuals inside them. The one truth I can say without a doubt is that things are changing quickly.

Solutions

AI solutions are as varied as the jobs they solve. Since the explosion of Claude Code in late 2025, agent harnesses have become the favored architecture for most workflows. I will focus on agent harnesses here; I don't think less complex systems like RAG pipelines will have a place in the future of workflows, except as tools for your agent. Agents' flexibility and direct access to data sources make them the obvious choice.

With flexibility comes greater security concern. AI models have been shown to disregard legality, or even instructions, in pursuit of completing their directed task. It's not hard to imagine a swarm directed to gather information for an investment decision getting into all kinds of trouble, akin to the recent OpenAI/Hugging Face incident.

Claude Code is the first that comes to mind. It's great for analyzing data, building reports, and many other things. It doesn't necessarily require much expertise, but as a long-time engineer I will say it's much easier to direct CLI agents if you have some inkling of what they're doing. If you don't understand your file system or other common concepts, it will be easy to fall into misunderstandings with the agent. Anthropic started MCP, an open standard, as a way for you to connect to your data providers. MCP connectors are great as an open standard, but in testing we've noticed they lead to model degradation because of context bloat and decision paralysis (which of 10 MCPs is the right one when their tools overlap?). Claude Code is highly customizable through skills, which are plain text (markdown) files that define workflows for the agent.

AlphaSense provides agents that are less flexible, in that they require a template input, defined sources, and a desired output file type. AlphaSense's benefit is its access to sell-side research and Tegus expert network calls. There are many product offerings from smaller companies similar to AlphaSense's, without the same data access, that offer templated generative outputs. The AlphaSense agents exist in rigid environments where the model can't be directed to produce whatever the user asks for, or to use classical engineering workflows to accomplish tasks the way Claude Code or Portrait's Easel can.

Easel is Portrait Research's agent. We built Easel to be the right combination of flexibility, low time to value, data connection efficiency, and security. Easel runs in our cloud, but users can connect their own data sources. Our native data connections are built to ensure efficient agent traversal and low context usage, so each session can be fully utilized without sacrificing model intelligence. Our Portrait CLI allows Claude Code, or any other desktop-native harness, to easily fan out research to Easel without worrying about Claude's local swarm limits. We support writing flexible skills and can work with any file formats commonly used.

Implications

The time it takes to understand a business, dig into the details of any situation, and build reports has certainly come down. Now, instead of spending time getting up to speed on each data or research provider's GUI, time is spent learning where and when to trust model output. The sheer volume of research that can be produced makes this harder: the boundary is only learned by spotting bad output, and spotting it requires scrutiny. Analysts now need to learn where and when they can trust their model output, not just how to do the synthesis.

Grounding in sources helps, but if the model doesn't understand your unique viewpoint or understanding, it can lead to misalignment in research direction. Systems you build or purchase must be highly auditable, trustworthy, and aligned.

In closing, I'll offer my very brief recommendation on how to approach implementing AI in your fund. Use what adds value today: demo products and assess all the solutions side by side. After you understand what works, learn how to deploy it to all your analysts in a way that is reproducible and doesn't leave a trail of overly unique configurations at each desk. The most important thing for a fund at this point is ensuring their investment frameworks, principles, and process are codified, and that the integration of AI supports them.