Trading and Investment-Research Agents: What They Actually Do, What They Don't, and Who Pays for Them

Table of Contents

• What a Trading or Investment-Research Agent Actually Is
• The Research Workflow, Decomposed
• Why "Trading Agent" Is Mostly a Misnomer Right Now
• The Economics: Per-Task, Per-Seat, or Per-Outcome
• Where These Agents Break
• The Buyers and How They Actually Deploy
• The Moat Question
• Insights Most People Overlook
• References

What a Trading or Investment-Research Agent Actually Is

Strip away the marketing and an investment-research agent is a coordinated stack of language models, tool calls, and data connectors pointed at one job: turning a vague analyst request, "give me a thesis on the regional banks after the rate cut", into something a human can act on. It reads filings. It pulls price and fundamentals from a market-data API. It runs a discounted-cash-flow model in a sandboxed Python environment. It cross-checks the earnings transcript against the guidance from two quarters ago. Then it writes the thing up.

This is the Agentic AI-as-a-Service pattern applied to one of the most labor-intensive, highest-paid corners of the economy. The analyst who used to spend Tuesday and Wednesday building a comp sheet now spends twenty minutes reviewing one the agent built overnight. That's the whole pitch, and it's a good one, because junior analyst hours are expensive and the work is brutally repetitive.

What separates a real agent from a glorified chatbot is autonomy across steps. A chatbot answers a question. An agent decides it needs three more pieces of data to answer the question, goes and gets them, notices one source contradicts another, flags the discrepancy, and keeps going. That multi-step planning, the same capability driving the broader financial-analyst agents category, is what makes the output usable rather than merely impressive in a demo.

The Research Workflow, Decomposed

The reason research agents work better than trading agents is that research breaks cleanly into discrete, verifiable tasks. Here's the chain most production systems run:

Ingestion. The agent pulls structured and unstructured data, 10-Ks, 8-Ks, earnings transcripts, analyst estimates, alternative data like credit-card panels or shipping manifests. The hard part isn't reading; it's knowing which filing footnote actually matters and which is boilerplate.

Normalization and modeling. It maps raw numbers into a model. This is where sandboxed code execution earns its keep: the agent writes and runs the spreadsheet logic instead of hallucinating the math. A model that runs is a model you can audit. A number a language model "remembers" is a number you can't trust.

Synthesis. It drafts the narrative, the bull case, the bear case, the catalysts, the risks. Modern long-context models from providers like Anthropic's Claude documentation describe handling this kind of multi-document reasoning, and it's genuinely where the technology has improved most in the last two years.

Verification. The best systems run an adversarial pass, a second agent that tries to break the first agent's thesis. This catches the confident-but-wrong failure mode that makes naive deployments dangerous.

Notice that none of these steps involve a trade. Every one produces a document, a model, or a flag for a human. That's not an accident. It's the only configuration regulators, compliance teams, and risk officers will currently tolerate.

Why "Trading Agent" Is Mostly a Misnomer Right Now

Here's the uncomfortable truth the category name papers over: almost nobody is selling you an agent that autonomously trades your money. The ones that exist are either (a) inside quant funds, built by the fund itself, never sold as a service, or (b) retail-facing "AI trading bots" that range from mediocre to outright scams.

The reasons are structural, not temporary. An agent that executes trades is a fiduciary actor the moment it touches client capital, and fiduciary duty doesn't bend to "the model said so." If an autonomous system blows up a portfolio, somebody is liable, and "the AI did it" has never been a defense the SEC accepts. The regulatory weather here is worth watching, the SEC has flagged concerns about AI in investment advice through proposals like its predictive data analytics rulemaking, which targets exactly the conflict-of-interest and accountability gaps that autonomous trading would amplify.

So the honest framing is this: research is a recommendation engine, and trading is an execution engine. The recommendation engine is where the GaaS money is, because a wrong recommendation costs you a bad memo and a human's judgment to override it. A wrong execution costs you the trade. The market has quietly sorted itself accordingly. When you see "trading agent" on a pitch deck, read it as "research agent that stops one step short of the order ticket", and if it doesn't stop there, ask much harder questions.

The Economics: Per-Task, Per-Seat, or Per-Outcome

This is where investment-research agents get genuinely interesting as a business, because the pricing models reveal what's actually being sold.

Per-seat is the lazy default inherited from SaaS. Charge each analyst $X/month for access. It's easy to sell and easy to budget, but it's misaligned: a tool that makes one analyst do the work of five should not be priced as if you still need five analysts. Per-seat pricing on a labor-replacing agent is a value-capture failure, and smart buyers know it.

Per-task aligns better. Pay per research memo, per model built, per company screened. The marginal cost of an agent run is real, API tokens, compute for the sandbox, data licensing, so per-task pricing maps to actual usage. The risk is that buyers can do the math on what each task costs and start squeezing. This is the same tension running through the broader debate over vertical agent pricing and value capture.

Per-outcome is the holy grail everyone gestures at and almost nobody implements cleanly in research, for an obvious reason: what's the outcome? In execution, outcome pricing is natural, you take a cut of alpha. In research, the "outcome" of a memo is diffuse. Did the agent's research make money? You can't isolate it from the human's decision to act or not act. Outcome pricing works beautifully in fraud detection or collections, where the result is unambiguous. In investment research it tends to collapse back into per-task with a fancy invoice.

McKinsey's work on generative AI's economic potential puts banking among the sectors with the highest value-add per employee from these tools, which is precisely why pricing power flows to whoever captures that productivity gain, not whoever ships the cleanest UI.

Where These Agents Break

I want to be specific here, because the failure modes are not theoretical and they're not the ones the demos warn you about.

Stale or wrong data poisons everything downstream. An agent that confidently models a company off a superseded filing produces a beautiful, wrong memo. The synthesis quality masks the input error. This is worse than an obviously broken output, because it passes the smell test.

Confident hallucination on numbers. Language models are pattern machines, and a plausible-looking earnings figure is exactly the kind of pattern they'll invent. The only real defense is forcing every number through executed code or a verified data source, never letting the model "recall" a financial figure. Systems that skip this are accidents waiting to happen.

Survivorship and lookback bias in anything backtest-flavored. The moment an agent starts evaluating strategies against historical data, it inherits every classic quant trap, and most general-purpose agents have no idea those traps exist.

The override problem. When an agent is right 95% of the time, humans stop checking. Then the 5% lands in a client portfolio. The reliability question that runs through the entire agentic AI-as-a-service cluster is sharpest here, because the cost of the rare error is measured in real dollars, not a re-run.

The Buyers and How They Actually Deploy

Three buyer profiles dominate, and they want very different things.

Asset managers and hedge funds want augmentation, not replacement. They deploy agents to compress the research cycle, more names covered per analyst, faster turnaround on a thesis. They're sophisticated enough to demand auditability and to keep a human firmly in the loop. They also tend to build rather than buy when the workflow touches their edge, which constrains how much of this market a vendor can actually own.

Wealth managers and RIAs want scale. An agent that drafts client-ready market commentary or screens for tax-loss-harvesting candidates lets a small shop serve more households. The fiduciary constraints here are heavy, this overlaps directly with wealth-management agents and fiduciary duty, but the documentation-and-summary use cases are low-risk and high-value.

Corporate finance, IR, and strategy teams are the quiet, underrated buyer. They don't trade at all, so the regulatory burden is lighter, and they have endless demand for competitive analysis, comp sets, and market sizing. This is where a research agent can run with more autonomy precisely because nobody's capital is on the line.

The deployment pattern that works is narrow and deep: pick one workflow, say, earnings-season transcript analysis, and own it completely, with integrations into the firm's existing data stack. The agents that try to be a general "AI analyst" tend to be mediocre at everything.

The Moat Question

Why won't a horizontal AI platform just absorb this? The defensibility argument for vertical research agents rests on three things, and only one of them is durable.

Proprietary data integrations are real but copyable, anyone can license the same market-data feeds. Workflow depth is stickier; an agent wired into a specific firm's models, conventions, and compliance rules is annoying to rip out. But the durable moat is the regulated-trust layer: the audit trails, the compliance-aware outputs, the documented human-in-the-loop controls that let a firm actually deploy this without their legal team having a heart attack. That's the same dynamic powering the argument for how vertical agents win regulated industries, and finance is the most regulated industry of all.

The blunt version: anyone can build an agent that writes a decent stock memo. Very few can build one a compliance officer will sign off on. The second thing is the business.

Insights Most People Overlook

The real product is liability transfer, not analysis. Buyers aren't paying for a better memo, they can get a decent memo from a generic model for pennies. They're paying for a system whose audit trail lets them defend the decision to a regulator or a client. The vendor who understands they're in the compliance-evidence business, not the prediction business, prices and builds differently.

Per-seat pricing is a tell that the vendor doesn't believe in their own product. If your agent genuinely replaces analyst hours, charging per human seat caps your revenue at exactly the headcount you're eliminating. Vendors who cling to per-seat are quietly admitting the agent is a productivity nicety, not a labor substitute.

The best research agents are deliberately bad at confidence. Counterintuitively, the most valuable systems surface their own uncertainty aggressively, flagging weak data, contradictory sources, and low-conviction conclusions. A polished, confident memo is a liability; a memo that tells you exactly where to look harder is an asset. Most demos optimize for the former because it sells better.

Execution will stay human longer than anyone predicts, and that's a feature. The market keeps expecting autonomous trading to be "two years away." It won't be, because the bottleneck isn't model capability, it's the fact that no firm wants to be the test case for who's liable when an agent's trade goes wrong. The smart vertical players aren't waiting for that to change; they're building the research layer that stays valuable regardless.

Alternative data is where the genuine edge hides, and it's the hardest part to automate. Anyone can summarize a 10-K. The agents that matter are the ones that can ingest messy, non-standard data, satellite imagery, app-download trends, hiring data, and normalize it into something a model can use. That ingestion problem is unglamorous, deeply unsolved, and precisely where the defensible work lives.

References

• Anthropic, Build with Claude documentation
• U.S. Securities and Exchange Commission, Proposed Rule on Conflicts of Interest Associated with the Use of Predictive Data Analytics (PDF)
• McKinsey, The economic potential of generative AI: The next productivity frontier