An investment research assistant that answers across your notes, broker research and filings — every sentence cited, every license checked
Days of cross-sector reading answered in minutes, every sentence cited to its page.
See one case, screen by screen ↓An investment research assistant is AI software that answers an investment team's questions across its sources and shows where each answer came from. Investment Research Assistant answers across your own notes, the broker research each reader is licensed for, filings and transcripts in their original languages, and expert calls that compliance has cleared. Eight agents plan, read, compare and verify, every sentence is cited to its page, and the analyst decides what the answer says. It serves equity research and credit alike: the same answer shows a holding's margin hit and, for bonds the funds hold, what it does to leverage.
AI in asset management: a buyer's guide to agentic AI for investment managers →
Updated
A portfolio manager asks one cross-sector question: which of our names does a proposed tariff hit, and how hard? The analyst starts opening broker portals one at a time, and a manager takes research from about 20 providers. Then she searches the research management system for the firm's own notes and meeting minutes, reads the 10-Ks and the annual reports in German, Japanese or Vietnamese, and checks the expert-call library. Reconciling all of it by hand typically takes days, and the research budget has stopped growing.
General AI investment research tools don't solve this, because broker research is licensed to named readers and may not be redistributed. A chat tool that reads every note and answers anyone breaks the agreement. Even when answers come back, the broker's estimate and the firm's own get averaged into one number, a foreign filing is read late or not at all, and nothing records who saw which licensed text.
Estimated split for one cross-sector question across about 20 broker providers, by hand and with the solution. Most of the time left is the analyst's own reading, choices and approvals.
Eight agents plan, read under each reader's licenses, compare and verify; the analyst decides, and a second person approves anything leaving the team.
Friday, October 16. A proposed 25% duty on footwear made in Vietnam would start January 1, 2027, with comments due November 16. Marcus Bell wants Quillfield Capital's exposure ranked before the November 6 investment committee. Here is what happens next in the working solution.
Lena covers 41 names across four funds. Her home screen opens on what the Planner did overnight. It re-ran her 7 watched questions and read 214 new broker notes and 38 filings in her coverage, and 2 watched answers changed and are waiting for her. Marcus Bell asked at 08:41 for a ranking of the tariff exposure. The Halden Brothers portal failed at 06:10, so its reports come from Thursday's copy.
Planner: “I drafted a 5-question plan — nothing is read until you run it.”
Lena types her question in plain words: “Which of our consumer holdings are most exposed to the proposed 2027 tariff on Vietnamese footwear?” The Planner replies with five sub-questions, each naming its sources: who sells footwear, how much is made in Vietnam, what the proposal says, how much can be passed on or moved, and what it means for gross margin and our positions. She can switch any of them off, add one or change the scope. The scope is her coverage: four funds, all four sources.
“Here is how I will answer. Nothing has been read yet — switch off what you don't need, add what's missing, then run it.”
Each source gets its own lane, with live counts. The Notes Retriever reads only notes her team can see (41 read). The Broker Research Retriever checks her licenses (18 of 22 providers), reads 72 reports and withholds 1 Ostrander Partners report she isn't licensed for. The Filings Retriever reads 58 filings and transcripts, 3 of them in German, Japanese and Vietnamese, each anchored to its page. The Expert Call Retriever reads 15 cleared transcripts and skips 1 on compliance hold.
“1 report withheld — Ostrander Partners, not licensed for you.”
The answer ranks Stridewell Brands, Vogtmann Sport and Harbor & Finch. Together they are $1.23 billion of positions across four funds, from portfolio accounting at the October 15 close. One table shows each name's footwear share of sales, share made in Vietnam, US share of sales and FY27 margin hit: −90bp, −60bp and −45bp in our base case. All 14 sentences are cited, with 23 footnotes from 17 documents. By citation, 26% of the answer comes from internal notes, 26% from broker research, 39% from filings and transcripts and 9% from expert calls.
Under the source mix: “1 not licensed”, “1 on compliance hold” and “Halden Brothers · Thursday's copy”.
The Dissent agent found 3 disagreements and set each one side by side. Pellam & Co. puts Stridewell's FY27 gross-margin hit at −180bp before any mitigation. Lena's own note says −90bp, net of the 4% price rise and a 12-point shift to Indonesia: the same facts under different assumptions. Lena chooses Show both, with the reason code “Different assumptions, both valid”. The memo, and anyone it's shared with, will show the same.
“I, Lena Whitfield, make this decision; it is recorded with my name, the time and the evidence.”
Vogtmann's Vietnam share comes from its Geschäftsbericht 2025, page 47 of 214: “Rund 34 % unseres Schuhvolumens wurde 2025 in Vietnam gefertigt”, or about 34% of its footwear volume. The Verifier matched the sentence to the passage at 97%. A broker's estimate of about 22%, from March 4, was written two weeks before the annual report, so the answer uses the company's newer figure and says why.
Original · German · page 47, beside the English checked by the Verifier.
The Verifier drafted 16 sentences and kept 14. It deleted “Stridewell plans a further 3% price increase in autumn 2027” because no passage in any source supports it. “Harbor & Finch has already moved its boot lines to Mexico” overstated the source, which says the company is qualifying a plant for 2028, so the sentence was rewritten in those words. Lena can still see both deletions.
“2 sentences removed by the Verifier — it could not cite them.”
Lena asks why the broker sees twice our hit. The Dissent agent answers with footnotes. The broker models the duty before any mitigation, and our note nets the price increase and the Indonesia shift. The expert's estimate of 18 to 24 months to move a production line is the reason to keep −180bp as the downside. Next she asks about Kestrel Foods. It is on the restricted list, so nothing is read, and the question is logged.
Verifier: “I can't search or answer about Kestrel Foods.”
One click saves the answer as memo M-26-0918, “Vietnam footwear duty: exposure of our consumer holdings”, with its 23 footnotes and 3 disagreements. Nothing is shared until Lena approves the final text. Then it goes to Marcus Bell and Kenji Mori inside the team. Before it can go to Ana Sato in Global High Yield, London, the agent checks her licenses. She isn't licensed for Pellam & Co., so 1 quotation and 1 sentence repeating its figure are replaced with a reference.
“Outside the team needs the Director of Research.”
Request AP-3318 lands in Helen's approvals. It shows what the agents checked and puts what licensed readers see beside what Ana Sato will see. Helen asks why a passage was replaced. The agent answers that Ana's team holds no Pellam & Co. seats in this morning's entitlement sync, and that the other 16 documents are licensed for her or public. Helen approves with the reason code “Research use, internal”. Lena can't approve her own request.
“A reason code is required; your decision is recorded with the evidence above.”
The same tariff question, seen by the five people who carry it: what their week looked like, and what it looks like now.
Eight agents plan, read under each reader's licenses, compare and verify; the analyst decides, and a second person approves anything leaving the team.
Turns a question into sub-questions, picks the sources and the scope (funds and dates), shows the plan and waits for approval. Re-runs watched questions when new sources arrive.
Searches the firm's own research notes, models' commentary and meeting minutes in the research management system and the document store.
Searches broker research from 22 providers and checks the reader's license with each provider.
Reads filings, annual reports and call transcripts in the issuer's own language, and translates passages anchored to their page.
Searches expert call transcripts that compliance has cleared.
Finds claims where sources disagree, sets both quotes side by side with their pages, and explains in one line why they differ.
Re-reads every cited passage on its own, deletes sentences it cannot support, scores confidence, and checks the restricted list and how fresh each source is.
Turns an answer into a cited memo, checks each recipient's licenses, proposes redactions and routes the approvals.
Approves what leaves the team. The agents propose; a named person decides.
Under every answer the analyst can ask back or tell the agents what to change. Each reply cites its sources, and the agents act on what she asks.
Why does the broker see twice our hit?
The broker models the duty before any mitigation. Our note nets the 4% spring price increase and a 12-point shift of units to Indonesia. Both assume brands absorb a third to a half of the duty, so the gap is the Indonesia shift. The expert's 18 to 24 months for a line move is the reason to keep −180bp as the downside, not to discard it.
Show the credit view
Only Stridewell has bonds in our funds: $96M of the 6.25% notes due 2031 in Global High Yield. Even at a 180bp hit, net leverage moves from 2.4x to 2.8x, inside the 4.0x incurrence test (Ana Sato, Sep 29). I added a credit paragraph to the memo draft.
Add Aoyama to the ranking
I added sub-question 6 and read 1 filing in Japanese. Aoyama makes 52% of its shoes in Vietnam but sells only 9% to the US, so it ranks fourth and low, with $58M in Global Consumer Equity. The ranking table now shows Aoyama.
What about Kestrel Foods?
I can't search or answer about Kestrel Foods. Grace Liang in Research Compliance added it to the restricted list on Oct 2 because the firm is wall-crossed on a private placement. Nothing was read, and this question is logged.
13 screens from the working solution, on its sample data. Pick one to see it large.
Answer rules, sharing service levels and retention are settings. Every change is versioned and logged, and a change to an agent needs a second approver.
| Setting | Default | Choose from |
|---|---|---|
| Default answer length | Standard | Short · Standard · Full |
| Minimum sentence confidence | 0.80 | 0.75 · 0.80 · 0.85 |
| Flag sources older than | 12 months | 6 months · 12 months · 24 months |
| Maximum broker quotation | 40 words | 40 words · No quotation — references only |
| Always show disagreements | On | On · Off |
| Checker SLA for sharing | 4 business hours | 4 business hours · 1 business day |
| Past SLA | Escalate to the Chief Compliance Officer | Escalate to the Chief Compliance Officer · Escalate to the CIO |
| Keep questions, answers and logs | 7 years | 7 years · 10 years |
Every figure is labelled: a target the solution is built to, an estimate, a typical published result, or a proven one.
“demo” = seen in the working solution, on its sample research data · “target” = the design goal, measured in the live solution · “estimated” = our estimate · “typical” = published figures (Coalition Greenwich study, 2021: about 20 research providers per manager, some 40; Substantive Research survey of 35 large managers, July 2026: 69% name broker licensing the biggest barrier to research AI; Substantive Research, September 2025: research budgets −0.2% at the 44 largest managers in H1 2025; US Investment Advisers Act Rule 204-2 on books and records). Quillfield Capital and the people, brokers, issuers, funds and identifiers named on this page are fictional: characters and sample data in the working solution.
An investment research assistant is AI software that answers an investment team's questions across its sources and shows where each answer came from. Investment Research Assistant answers across your own notes, the broker research each reader is licensed for, filings and transcripts in their original languages, and expert calls that compliance has cleared. Eight agents plan, read, compare and verify, every sentence is cited to its page, and the analyst decides what the answer says. It serves equity research and credit alike: the same answer shows a holding's margin hit and, for bonds the funds hold, what it does to leverage.
It can when it reads as the person asking, and that is how it reads here. The Broker Research Retriever checks the reader's license with each provider, uses only licensed text, quotes at most 40 words and declares what it withheld. In the demo that is one Ostrander Partners report, with a one-click access request but never its text. When a memo is shared, each recipient is checked again and unlicensed passages are replaced with a reference.
This one cites every sentence. Each footnote opens the passage on its page, with the source, date and license. A Verifier re-reads every cited passage independently, deletes any sentence it cannot support and shows what it deleted. A sentence below 0.80 confidence goes to the analyst instead of the answer.
Yes. The Filings Retriever reads filings, annual reports and transcripts in the issuer's own language (German, Japanese and Vietnamese in the demo). Every translated passage is shown beside the original, on its page, and machine-translated text is labeled.
The restricted list comes from the order management system, read-only. A question about an issuer on it is blocked with a plain reason, and the question is logged. Research Compliance places expert calls on hold, for example when a speaker named an unannounced customer contract, and the Expert Call Retriever never reads a held transcript.
People stay in control. The Planner waits for the analyst to approve its plan, and the analyst chooses in every disagreement and approves the final memo text. The Director of Research approves anything shared outside the team, and Research Compliance approves any client use. Nobody approves their own request, and the agents may not trade, change a position, contact an issuer or broker, or approve anything.
Every question, plan, passage read and cited, license check, viewer, decision and reason code is kept for 7 years by default, and Rule 204-2 requires five for most records. An auditor can export an evidence file sealed with a SHA-256 hash. Reports show broker research used in answers by provider, as evidence of research use for the broker vote under MiFID II and FCA rules.
The Agentic Solution Engine builds and deploys it from your requirements and documents: your sources, broker license files, roles and sharing rules. It goes live once every quality gate has passed. One of those gates is a test question: in the demo it cited 6 of 6 sentences and declared the report it withheld. We will walk you through it on your own research first.
Every call read the day it reports, every claim tied to its transcript line.
1-min demoEquity Research AutomationEvery result read, modeled and noted before the open, every figure tied to its page.
1-min demoFinancial Data ExtractionEvery covered name's model refreshed the day it files, every number tied to its page.
1-min demoWe'll run Investment Research Assistant on a sample of your own notes, filings and licensed research.