Over the last two decades, the investment research stack has grown by addition. A search tool for documents, a terminal for data, a research management tool for notes, a spreadsheet for modeling. Each of these speeds up a part of the process, but they were never built to work together. So far, AI has been no different.
While investors now have access to a chat interface that connects public filings and internal notes, most are using it as an enhanced search and synthesis tool. Ask what management said about quarter-to-date trends, and the chatbot answers in seconds. But harder questions require more data and context which most AI systems don’t have. Investors are left to fill in the gaps themselves: alternative data in one tab, macro numbers in another, and an analyst stitching the pieces into a cohesive analysis. And because chatbots have no context of the company, every question starts from scratch. AI saved time on basic search tasks, but now that time is filled with new manual work.
The mistake is treating AI as one more tool. It should be the first real opportunity to rethink the entire investment process. Instead of asking “how do we add AI to what we already have,” the right question is “how would we build this process from scratch?”
One System, Not One More Tool
An AI-native research platform requires a few things: the data, a single place for all research needs, an interface for humans and AI alike, and true agency.
Thoughtful Data Integration
Most funds have connected the basics, like filings, transcripts, internal notes and emails. But there’s a much broader set of data that investment decisions rely on: consensus estimates, macro inputs, alternative data, internal expert network calls, positioning, risk data, and more. If it exists inside the fund, it should be part of the system.
Even harder is integrating critical information that doesn’t exist in a clean data format. Clinical trial notes on a website, how a graph in a quarterly filing changed over time, how the tone of a CEO’s voice sounded in their earnings call. In order to use these as inputs, the system must have the ability to browse, listen, see, and understand. These are actions a human analyst does naturally, but a basic chatbot cannot.
Getting the data layer right requires a meaningful buildout of a fund's data infrastructure. Done well, the agent has access to everything an analyst would, and can process far more of it. Done poorly, the agent floods its own context with noise, burns unnecessary tokens, and returns outputs that seem reasonable but are quietly wrong and difficult to untangle. (And no, MCP does not solve this, and we'll dig into why in another post.)
Every Step of the Process, One Place
Data integration solves what the system can pull. But querying data is only one step in a sequence of work that currently lives in a half dozen disconnected tools. Listen to an earnings call in one app. Send conference notes into another. Pull alternative data from a third, then bounce between Excel files for correlation analysis and model updating.
Tool switching wastes time, but the deeper cost is lost context. Each tool holds a fragment of the analyst's thinking, and the only thing connecting them is the analyst. The AI, if it's involved at all, only sees whatever got pasted into it last.
The true test of one system is simple: can an analyst run their research process there, end to end? That's the only way an agent can follow the thread of the work instead of catching fragments of it. It also breaks a limit analysts have always lived with: a person can only do one thing at a time. An agent with the full thread isn't bound by that. Analyses run in parallel, on schedules, in the background, while the analyst reviews output instead of babysitting the process.
Built For People and Agents
A single workspace solves where the work happens. The harder part is who it serves. A lot of energy is going into building for AI agents: MCP servers, CLI tools, agent-ready data sources. We believe the system should be built for both people and agents, because human judgment in investing isn't going away. And what a human needs is different from what an agent needs. Most teams building for humans default to chat, but chat isn't always the best way to get an answer. Sometimes it's a dashboard read at a glance, a real-time alert, or a document that's one click away.
Building for humans and agents doesn't mean bolting two systems together, it means everything inside is usable by both. A dashboard a human reads at a glance is structured data an agent can pull into its next analysis. A memo an agent drafts is something a human can edit, question, and trace back to its sources. The right form is a workspace where humans and agents work side by side.
An Agent That Knows, Learns, and Acts
Everything so far builds a strong system on its own. But we think it can go further. Every agent will soon write memos, analyze filings, and update models. The difference lies in three things that don't come from the model layer.
The first is domain knowledge. A good agent’s context builds over time. It needs to understand the companies an investor covers. It needs to know what their key drivers are at a given moment, what mattered in prior quarters, and all the subtleties that investors know but don’t write down. An analyst should never have to explain the basics to it.
The second is learning. Investing is unusual in that new information is constantly changing the ground truth, meaning a good financial agent cannot rely on its knowledge cutoff. It has to learn from new information as it comes in live, and it must constantly update its understanding of the overall environment.
The third is proactivity. A good agent doesn't wait to be asked. It starts earnings previews a few weeks before the earnings date, actively monitors live commentaries for key topics, and checks an investor thesis against incoming data. Multiple workflows happen in the background before investors have to ask.
The three reinforce each other: an agent that knows the company learns faster from new information, and an agent that stays current does better work on its own. Faster research has a ceiling. An agent that can generate insight doesn't.
Building For The Fund
Everything we've described works for a single team, but it should be built at the fund level. The data infrastructure alone makes the case: the buildout is too foundational to repeat team by team.
Building centrally also means building for the long run, something that will thrive as AI rapidly changes. That means infrastructure that adapts as the stack evolves: a model-agnostic design so the fund always runs on the best available model, with cost optimized through prompt caching and context management. It also means dev-tooling that lets the fund's developers trace what an agent actually did, evaluate whether the output improved, and iterate quickly. Just as important, a central system lets institutional knowledge be retained, shared, and compounded.
This is a lot to build, and doing it from scratch could mean a multi-year effort. One system consolidates vendors and simplifies contracts, but the biggest win is that finishing it means a multi-year head start.
Implied: Your Foundation to Build On
AI is already reshuffling the tech industry, and it will do the same to investing. The edges that came from headcount, data access, and algorithms are getting easier to replicate. What can't be replicated is what's already inside each fund: the proprietary data, the analysts' accumulated judgment, the institutional knowledge that never made it into any vendor's dataset. The winners of the next decade will be the funds that integrate that edge into a working AI system first.
There's a lot to build, and a race to build it. Where your fund spends its effort is the real decision: on the harness that surrounds the models — agent frameworks, evals, model routing, context management — or on integrating the data and judgment that only exist inside its walls?
We built Implied so you don’t have to choose. Everything this piece describes, from the infrastructure, to the workspace, to the agent that knows, learns, and acts, already exists as one system. Your fund can run it as-is, or take the foundation into your own environment and make it your own. Either way, the starting point is a proven working system, already used by investors at five of the ten largest discretionary hedge funds, so the effort can be focused on the only part that's an edge: your own data, workflows, and judgment.
We have a solution for funds at every size and stage. If you've been thinking about the best way to build an AI-native platform, talk to us.