How AI Is Reshaping Investment Research: Evidence from 4.2 Million API Queries
The dataset analysed in this piece is drawn from anonymised query traffic on the Valyu API from Q1, on free and individual tier plans. All account, customer, session, and IP-level identifiers are stripped at ingestion under our standard data-handling policy; figures below describe category-level patterns only, and no individual prompt, response, or customer is reproduced. Enterprise tier traffic is excluded entirely. Valyu operates a strict Zero Data Retention policy on enterprise plans, so those queries are never stored or available for aggregate analysis.
What follows is a breakdown of what the data tells us about how AI is reorganising the desk-level research workflow in finance.
Public discourse on AI in financial services has fixated on the wrong layer. The headlines describe a future of autonomous trading desks and machine-managed risk, both of which assume the model is the agent of action. The activity we observe at the API tells a different story. The visible change is happening one stage earlier, inside the research that produces every position those layers eventually act on.
Trading and risk both sit downstream of an analyst's view. If the way views are formed changes, the rest of the system rearranges around it. We pulled six months of financial query traffic from the Valyu DeepResearch API and broke it down by category and sender. The shape of that traffic is the cleanest signal we have on how the buy-side and sell-side are actually using AI at the desk level today.
Research as the unit of analysis
Finance, viewed from sufficient height, is an information industry. Every position, deal, and capital allocation traces back to someone consuming a body of evidence and forming a defensible view. The structure of the industry reflects this. The sell-side, primarily the bulge-bracket and mid-market investment banks, employs roughly 9,000 to 11,000 publishing equity analysts globally who cover companies and distribute notes to institutional clients. The buy-side (hedge funds, asset managers, private equity, sovereign wealth) consumes those notes, pressure-tests them, develops a variant view, and converts that view into capital deployment.
Trading and risk both sit downstream of this work. Quant strategies are not exempt. They are research compressed into code, with human judgment exported into rules. So when the production function for research changes, every layer above it adjusts.
The desk-level research workflow has six stages that have been broadly stable for three decades: source the names, do the work, build the thesis, debate it at investment committee, execute, monitor. AI has now intruded on five of the six. The first instinct on what to look at and the conviction call at the end are the two stages still owned by the analyst. Everything between them is being rebuilt around retrieval and inference.
This pattern, machine intrusion in the middle and human persistence at the edges, tracks the broader theory of technological substitution. Citadel Securities, in a February 2026 macro note authored by global macro strategist Frank Flight, framed it directly: "The economy contains a vast array of tasks: physical, relational, regulatory, supervisory – that are costly to automate. Even cognitive automation faces coordination frictions, liability constraints, and trust barriers. It seems more likely that AI will be a complement rather than a substitute for labor in many areas. Historically, technological revolutions have altered task composition rather than eliminated labor as an input." The investment research workflow is now living through that composition shift, and the query data below is one of the cleaner empirical pictures of where the line between complement and substitute is actually falling on the desk.
The mechanics of machine-driven research
Underneath every credible AI deployment in this space sits the same architectural pattern: retrieval-augmented generation. This is because every claim in financial research has to be traceable to the source. An unattributed number cannot be defended at IC and cannot survive compliance review. So instead of letting the model produce text from training memory, the system retrieves source documents at query time and constrains the answer to be grounded in them.
The pipeline is now standard across the firms running it in production. Filings, transcripts, news, fundamentals, and market data are parsed, chunked, and embedded as vectors in a search index. A user query is embedded the same way, the closest chunks are returned, and the model writes the answer alongside citations.
The newer layer is agentic
Rather than running a single retrieval call, the system decomposes the original prompt into a tree of sub-questions, decides what to retrieve next based on what came back, calls tools when numerical work is required, and continues until the answer holds together. A prompt as simple as "build a thesis on UBER" expands into 14 to 22 sub-queries covering peer comps, segment-level growth, margin trajectory, balance sheet composition, recent guidance revisions, and consensus deltas. The output is approximately what a competent first-year analyst would produce in three to five working days.
Two constraints set the engineering
- Every claim has to carry a citation
- The underlying data ages in hours rather than weeks
Both push the same direction: retrieval has to be live, ranked, and tight. Static document stores break on Monday morning, when fresh filings and transcripts hit the wire and yesterday's vectors are already stale.
This is what the Valyu DeepResearch API provides: a customisable research agent that returns ranked, cited, structured results across filings, transcripts, fundamentals, market data, web sources, and academic literature, formatted directly into a dense research report with your required deliverable. With different levels of depth depending on the desired mode; Fast, Medium, Heavy, Max. Building this internally typically takes 6 to 12 months, a dedicated full team of data engineers, and on going maintenance. The economic effect is that differentiation moves up the stack, away from infrastructure work and toward agent design.
A taxonomy from the wire
Across the last quarter, a meaningful share of the queries hitting our API were financial in origin: from analysts, from agents running unattended on their behalf, and from product features embedded in firm-built copilots. We classified roughly 4.2 million such queries by intent and sender type. Six categories cover 96% of the volume.
The percentages below describe what we observe at the API as a single user-facing query. They understate the underlying compute. Each one expands, on average, into 11 to 19 sub-queries that the agent runs to answer the prompt. The numbers describe demand at the prompt layer, not the work being done beneath it.
The traditional spine: Investment Thesis, Filings, Transcripts
These three categories account for 57% of total volume and represent the traditional single-ticker workflow done faster. An Investment Thesis prompt almost never lands as a single retrieval. It arrives as a fan-out across peers, segment growth, margin trajectory, balance sheet quality, recent guidance, and the delta to street consensus, which the agent then stitches into a memo. The output is structurally what a junior analyst would produce by hand. The economics differ. A complete thesis run that previously consumed 14 to 18 analyst hours now resolves in 75 minutes of agent time at roughly $50 per run.
Filings and transcripts queries are narrower in scope. A filings query usually targets a specific item in a 10-K or 10-Q (segment disclosures, related-party transactions, contingent liabilities). Transcripts queries are most often forward-looking: pulling out guidance language, capex commentary, or new product mentions across an earnings call. Both are bounded tasks that AI compresses but does not transform.
The new shape: Universe Coverage
Universe Coverage is the category where the workflow itself, not the speed, is novel. Roughly one in five financial queries we log is now this shape. 78% of queries in the category originate from buy-side senders: PE platforms screening portfolio targets, asset managers running mandate-level surveillance, and hedge funds watching for thesis breaks across long lists.
A typical universe coverage cycle spans 240 to 320 tickers and runs weekly, sometimes daily. No analyst desk in the historical sense has ever covered that many names that often. The shape is wrong for human throughput: wide and shallow on each pass, compounding over time as positions get added, dropped, or rotated. Universe Coverage is the first category in the dataset where AI does work no analyst desk had previously attempted, because the labor cost made it impossible.
78% of the queries in this category originated from individuals from the buy side. A representative case of a well known buy-side client base: individuals from investment firms run recurring thesis-level memo on the DeepResearch API across hundreds of names in their coverage universe. Each name carries a refreshed view, with peer comparisons and risk signals layered in. Produced by hand, the same exercise would consume dozens of analyst-weeks of work and would run at a fraction of the current cadence. They run it on a regular schedule, with no incremental headcount.
The persona split
Cutting the same dataset by sender type sharpens the picture. Persona is inferred from each sender's query composition rather than self-reported.
The inferred data suggests that Hedge funds and PE platforms are the heaviest users of Universe Coverage. Quants would run almost entirely on Financial Data: 91% of queries from quant senders fall into the Financial Data bucket, which is consistent with how those strategies are constructed. Sell-side equity research, by contrast, barely touches the breadth categories. 74% of sell-side queries fall into Investment Thesis, Filings, and Transcripts, which is what their institutional clients pay subscription dollars to receive.
The asymmetry is informative. The breadth-versus-depth split tracks the business model, not the technology. AI lets each side push further into its own posture. The sell-side gets deeper on names it already covers. The buy-side gets wider across a universe the sell-side cannot afford to follow.
The surprise: Consensus Synthesis
The category that genuinely surprised us was Consensus Synthesis, at 7% of volume. The shape: a portfolio manager, IC analyst, or CIO office pulls every available sell-side note on a single name (Goldman, JPM, Morgan Stanley, Evercore, Wells, the sector specialists) and asks the agent where the brokers agree, where they diverge, and where the consensus position is exposed. The output is one document compressing what is typically 7 to 12 broker reports and roughly 200 pages of analyst output into a structured diff, often with a probability-weighted view of catalyst skew.
The function this serves is IC prep and thesis stress-testing. Several buy-side desks now run it as a "devil's advocate" pass before committee. The agent ingests the consensus, builds the bear case from it, and surfaces the contradictions a PM would otherwise have to chase by hand.
The reason this surprised us: investment committee debate was the part of the workflow that most observers, including most buy-side practitioners we spoke to as recently as 2024, believed AI would not touch. Conviction is a human act and remains one. But the artifact that conviction is debated against (the consensus map, the variant view, the assembled bear case) is now machine-produced. The debate runs on AI output even when the final call does not.
Financial Data and Filings as middleware
Financial Data and the long tail of Filings work are plumbing. Together they account for 35% of volume but rarely surface as a user-facing answer. They are middleware: called by another agent or an internal pipeline that needs a specific number, a clause, or a row of structured data to keep moving.
Three years ago, every serious finance team building AI features was constructing this layer in-house. The standard pattern was a team of 4 to 6 engineers, a build of 4 to 6 months, and a maintenance burden that scaled with every new SEC filing format and every transcript provider's API change. The teams running this in production today have largely abandoned that approach. They consume the data layer as an API call and place their differentiation upstream, in the prompt and the agent, where the actual investment edge lives.
What the data implies
The composition of the queries reveals which work is being recreated faster, which work is structurally new, and which work is becoming infrastructure. Single-ticker research is being compressed in time but not in shape. Universe-level surveillance is genuinely new and represents the first category where AI is producing work nobody was doing before. Investment committee debate has acquired a machine-built substrate, even though the conviction call remains human. Financial data ingestion has gone the way of payments: solved by APIs, and no longer a place where firms compete.
The competitive surface is moving. Parser quality and data engineering muscle are becoming table stakes inside the next 18 to 24 months. The firms that win at AI-enabled research will be the ones with the best prompts, the most considered agentic logic, and the cleanest taste about which questions are worth asking in the first place.
If you want to run your own queries on the same DeepResearch API, the DeepResearch API is open at platform.valyu.ai.