A chatbot answers the question you asked. A research system works out which questions need answering first. The difference is planning and delegation — and it is what separates the Enterprise terminal from every AI ticker widget on the market.
Ask a general-purpose model "should I be worried about this position?" and it will produce something fluent. It will also do the whole thing in a single pass: one context window, one shot, no notion of which parts of the question it is actually equipped to answer.
Real investment questions do not have that shape. "Should I be worried about this position" decomposes into at least four independent enquiries:
Those need different data, different tools and different reasoning. Flattening them into one prompt is why so much "AI for finance" reads confident and lands shallow.
Research in Fincept Terminal Enterprise runs in two stages.
The planner reads your question and decides what actually needs answering. Not a keyword match — a decomposition. It works out which legs the question has, what order they need to run in, and which of them depend on the others.
The specialists then take one leg each. An equity agent with fundamentals and filings access. A macro agent with the economic series and central bank data. A cross-asset agent for exposure and correlation. Each has its own tool set and its own slice of the data, and each returns findings rather than prose.
The planner reconciles the results — including where they disagree, which is frequently the most useful output — and writes the answer.
There is also a fast path. Not every question needs a committee: "what were last quarter's margins" is a lookup, and routing it through a planner would waste your time and our compute. The terminal runs a lighter single-pass mode for questions that deserve one, and escalates when the question earns it.
An agent is only as good as what it can reach. Inside the Enterprise terminal, the research agents have live access to:
That last one is the difference between a research assistant and a research assistant that works at your firm. An agent that has read your investment committee memo answers a question about the position differently from one that has not.
The agents we ship cover the general cases. Your desk is not a general case.
Agent Studio lets you define your own — its instructions, which tools it may call, what data it can see, how it should format what it returns. If your process for screening a name has eleven steps and six of them are mechanical, that is an agent.
The Node Editor wires agents into workflows visually. Screen, then enrich, then check against the dataroom, then draft — as a graph you can see and change, not a script one person on the team understands.
MCP servers connect those agents outward. If you run your own systems — a positions database, an internal research archive, a risk service — an agent can reach them through MCP without that data leaving your control or being copied into ours.
The Report Builder takes what comes out and produces the artefact somebody actually has to send: a note, a memo, a committee pack, with the sources attached.
The failure mode of AI research is not that it is wrong. It is that it is wrong fluently, and you cannot tell which parts to check.
So the agents return what they used. Which series, which filing, which date, which document out of your dataroom. When a number in the note is not traceable to something the agent actually read, that is a defect and we treat it as one.
This matters commercially, not just intellectually. If you are accountable for other people's money, "the model said so" is not a defensible position, and a research tool that cannot show its work is a liability dressed as a productivity gain.
Be concrete about it. The workflow this compresses is: run a screen, export the list, open each name, pull the filings, read the last two quarters, check the macro backdrop, look at what the sell side said, cross-reference your own notes, write it up.
Thirty to ninety minutes per name if you are quick, and it is the part of the job that is least like judgment and most like assembly. That is the part worth handing over. The judgment stays with you — it should, and no amount of planning and delegation changes that.
The honest test of a research system is not a demo question. It is a name you already know well, where you can tell immediately whether the output is sound or merely fluent.
That is the walkthrough we would rather run. See what an Enterprise seat includes, or get in touch and bring a name you know cold.
Instead of one model answering a question directly, a planning agent decomposes the question into legs — valuation, macro exposure, competitive position, filings review — and delegates each to a specialist agent with its own tools and data access. The results are reconciled into a single sourced answer. It matters because most real investment questions have several independent parts, and a single pass over one context window handles them badly.
A general model answers from training data and whatever you paste in. The Fincept research agents run inside the terminal with live market data, fundamentals, news and any documents in your private dataroom already wired in as tools. They plan before answering, cite what they used, and reason over current numbers rather than a training cut-off.
Yes. Agent Studio lets you define agents with their own instructions and tool access, and the Node Editor wires them into a workflow visually. MCP server support connects those agents to systems you already run, so an agent can reach your own data without anything leaving your control.
No. The open-source edition includes a single, simpler market assistant. Planning, specialist delegation, Agent Studio, the Node Editor and MCP connectivity are Enterprise features, because they consume real compute per query and reach data that cannot be redistributed in a public build.
Multi-agent research, a quant lab with backtesting, derivatives and macro coverage, and a private dataroom your agents read alongside the market — in one private terminal, from $10 per user, per month.