Google Built Its A.I. Research Agents for Wall Street
Wall Street's dominant financial data providers — FactSet, S&P Global, and PitchBook — are now part of Google's research infrastructure. Google disclosed Monday that the three firms helped design Model Context Protocol servers for its new Deep Research and Deep Research Max agents, letting them query those firms' data feeds directly as part of an automated research workflow, according to VentureBeat. Whether those integrations are live or still design-stage is the question neither Google nor its partners have answered — and it's the one that determines whether this press release is also a genuine shift in how Wall Street does research.
The answer matters because Perplexity's CTO publicly called MCP impractical at scale, citing high context window consumption and authentication friction. Google appears to be betting the opposite direction. That disagreement is the actual story here, not the benchmark table.
For the data firms, the arrangement is a deliberate bet on survival rather than a passive surrender. Partner with Google and risk becoming invisible middleware — the intelligence layer moves up to Google's agent while the data subscriptions remain but the customer relationship shifts. Stay out and risk that analysts and investors route their workflows through Google's API instead. Neither option is comfortable, but both require being inside the architecture.
The MCP architecture is the mechanism that makes this plausible. Unlike a conventional API integration, a Model Context Protocol server lets a research agent interact with private data without that data leaving its native environment. A hedge fund could point Deep Research at its internal deal-flow database and a PitchBook feed simultaneously, asking the agent to synthesize findings across both without manually moving data between systems. Whether it works reliably at scale — and whether enterprise finance teams will trust an agent to produce analysis they previously paid a human researcher to generate — remains the open question.
Google's own benchmark figures for the new Deep Research Max are substantial: 93.3 percent on DeepSearchQA, a research accuracy test; 54.6 percent on Humanity's Last Exam, a harder multi-step reasoning benchmark; and 85.9 percent on BrowseComp, which tests the ability to locate hard-to-find facts, according to Google's data. All three beat the December 2025 preview and Google's stated GPT 5.4 Thinking comparison. Independent verification against the full model set has not been performed; the figures are self-reported, which is standard practice but worth noting when the numbers are this large.
Google open-sourced the DeepSearchQA benchmark alongside the release — a move that signals confidence in the figures, since a company expecting external scrutiny tends to publish its methodology. The same infrastructure already powers Gemini App, NotebookLM, Google Search, and Google Finance, per Google's announcement. The tiered product structure — Deep Research for speed, Deep Research Max for exhaustive analysis using extended test-time compute — is aimed squarely at enterprise buyers making a build-versus-buy decision on research automation. Sundar Pichai put it plainly on X: use Deep Research for speed and efficiency, Max for the highest quality context gathering and synthesis, as VentureBeat reported.
The next 90 days will determine whether any of the financial data partners has moved from design-stage to live deployment — and whether Google's agent workflow can sustain enterprise-grade reliability without the context overhead that Perplexity's CTO cited as a blocking issue. If FactSet, S&P Global, or PitchBook announces a customer pilot before that window closes, the partnership is real. If the press release language holds, the strategic bind remains theoretical.
