Eighteen of Twenty Top Investment Banks Run Research Through an AI Agent on AlphaSense

Agents now run 99% of the actions on AlphaSense. Eighteen of the world’s twenty top investment banks pay for it.

AlphaSense launched SuperAnalyst on October 6 at its AlphaSummit conference. The agent runs without stopping.

It plans research, reconciles evidence, runs code and writes memos. The outputs are the artifacts a desk ships.

Financial models, pitch decks, comparison tables and dashboards. Not summaries of work, the work itself.

The corpus is the moat. The agent draws on more than 500 million documents, per AlphaSense on October 6.

Filings, earnings calls, broker research and 300,000 expert interviews. That last category is the expensive one.

Two parts sit underneath the product. A Context Graph turns raw text into structured, source-linked data.

Skills are reusable workflows built by former practitioners. Somebody who used to do the job encoded how they did it.

The money arrived first. A $350 million round in June 2026 valued the company at $7.5 billion.

The October launch broadened access and added tools. AlphaSense passed $600 million in yearly revenue by the first quarter of 2026.

Adoption is the number to sit with. The firm added over 10,000 banking users in the third quarter alone.

It now serves 18 of the top 20 investment banks, among more than 100 in total. Named customers include Bank of America, J.P. Morgan and Morgan Stanley.

The reach runs past finance. Some 8,000 firms use the platform, including Google, Intel, Nvidia and Pfizer.

AlphaSense says the agent tracks a thesis and updates its work without a new prompt. That is the claim that matters, because it describes persistence rather than retrieval.

CEO Jack Kokko said on October 6 that results hinge on data quality and tool strength. Read that as a sales argument about the corpus, not a caveat about the model.

No third-party audit of the output has been published. The accuracy claim rests on the vendor.

The usage figures are striking and they need unpacking. Agents run about 2 billion actions a year on the platform, against 20 million by humans, per AlphaSense on October 6.

That ratio produces the 99% headline. It also compares two things that are not alike.

An agent makes hundreds of tool calls to answer a question a human would answer with four clicks. Counting actions measures how the software works, not how much work got done.

Token use tells the same story. Consumption hit 54 trillion a year, five times the early 2026 rate.

Tokens are a cost line. A five-fold rise is evidence of spend, not of output quality.

The harder number is scheduling. Agents running on a schedule handle 19% of all activity.

That one is real. Scheduled work means nobody is sitting there prompting.

Third-party recognition exists but is not an audit. Forrester named AlphaSense the only Leader in market and competitive intelligence on August 27.

Customer accounts are the strongest evidence available. Aberdeen’s Tiago Rodrigues said on October 6 that it builds valuation models like a human analyst.

A user at a top-two investment bank built a gross-margin workbook through the agent. AlphaSense also showed SuperAnalyst Apps, where a user describes a tool in plain words and the agent builds it.

The vendor’s own documentation carries the honest limit. AlphaSense’s FAQ says financial data citations are still being rolled out.

It tells users to check every figure before use. That instruction is the product’s current ceiling.

Final sign-off on every deliverable stays with a person. So does the thesis.

A person decides what question matters. A person decides what to do with the answer.

The target is the sell-side research associate. That role builds models, writes earnings previews and assembles pitch decks.

The desk runs on junior labor. First-year analysts build the models, pull the comps and draft the books.

Volume shows where that work has gone. Users have made over one million slides across more than 100,000 decks, per AlphaSense on October 6.

The headcount evidence is already on the record. McKinsey said in June 2026 that banks cut junior analyst classes by up to two-thirds.

Bloomberg Intelligence said in January 2025 that up to 200,000 Wall Street jobs face cuts. Neither figure is attributable to one vendor.

On the scoreboard this is legacy fee pool compression, and the pool is a headcount budget. Research is a cost centre, so the saving shows up as fewer seats rather than as revenue.

There is a second loss nobody budgets for. Building the model badly three times is how an associate learns what the model is for.

The tool is live now. Eighteen of twenty top investment banks already use it, which removes the procurement delay entirely.

The gate is the citation gap the vendor disclosed itself. A number the agent cannot source is a number a human must verify.

That verification step is what currently keeps a person in the loop at all. Close it and the checking job shrinks with the drafting job.

AlphaSense has not said when that ships. Until it does, every figure gets checked by hand.

Call it twelve to eighteen months for drafting, and longer for the check. The constraint is engineering, not regulation, which is why this moves faster than credit or claims.

Client calls and final judgment stay with a person. Those are the last two rungs on the research desk, and the ones below them are already gone.

Territory: +machines on the output, +incumbents on the sign-off.

The junior research headcount budget is the pool compressing, while the citation check and the client call stay human.