Research foundation · B2B fintech · Institutional investment research
Designing for people who get paid to be right.
A venture-backed AI research platform for hedge funds and asset managers wanted to move from chat-first to workflow-native — from a surface an analyst visits when they have a question to one they live inside because their work is already in it. The founder is a former buy-side operator who can build a surface overnight, which raises the stakes on the surface being right. Four documents came first: a seven-category market map, a user model, a five-phase workflow map, and a synthesis built to force decisions in a room.
The premise
You don't skip the research. You pay for it in rebuilds.
The fast-looking version is to open the product, sketch a better IA, and ship something on Friday. Then an analyst uses it, and it reads as professional to a designer and amateur to a practitioner. In this domain the vocabulary carries status — and getting it wrong isn't a polish problem, it's a credibility problem that sends the whole surface back to the start with build cycles already spent on it.
The vocabulary audit
This was the highest-confidence output of the entire user document, and the cheapest thing to get wrong. Every label in the interface is downstream of it.
Words with no acceptable substitute — they signal consumer software and cost credibility on contact
- discover
- explore
- insights
- unlock
- all-in-one
- smart
Hedging or softening a practitioner's own vocabulary reads as condescension. V — consistent across analyst writing, fund letters and every incumbent interface in the competitive set.
The shape is the argument. One path is a spiral, the other is a line — and the spiral costs build cycles, not just design ones. That converts a craft argument into a cost argument, which is the one a founder actually feels.
Foundation 01 · Market & competitive landscape
Not one market. Seven, and they're converging.
The difference between “this competes with Bloomberg” (mostly wrong) and “this unbundles specific workflows from the terminal stack while creating a layer that didn't exist” (closer to true) is the whole positioning of the product. So the first document was a taxonomy with verified per-seat pricing in every row.
| Category | What it sells | Canonical players | Per seat / yr |
|---|---|---|---|
| 1 · Terminals | Real-time multi-asset data, messaging, trading | Bloomberg, FactSet, Refinitiv, S&P Capital IQ | $12K–$32K V |
| 2 · Modern data platforms | Equity fundamentals, charting, screening | Koyfin, Fiscal.ai, Godel Terminal, YCharts | $360–$3,600 V |
| 3 · Document intelligence | Search and synthesis across filings and transcripts | AlphaSense, Hebbia, BamSEC | $12K–$25K+ V |
| 4 · Expert networks | 1:1 expert calls, transcript libraries | GLG, Third Bridge, Guidepoint, AlphaSights | $20K–$100K + per call I |
| 5 · Earnings & IR content | Calls, transcripts, IR materials | Quartr, Stream by Mosaic, AlphaStreet | Free–$10K V |
| 6 · AI-native platforms | Agentic research, end-to-end workflows | Rogo, Hebbia, Fiscal.ai, Perplexity Finance | Free–$50K+ V |
| 7 · Data infrastructure for AI | Clean structured fundamentals exposed to LLMs | Daloopa, Snowflake Marketplace, FactSet API | $5K–$50K+ I |
The categories are collapsing into each other. One incumbent consolidated rows 3, 4 and 5 in a single $930M acquisition, then launched structured financial data that reaches into row 1. Another bought a slide-generation company to move from inputs to outputs. Everyone is racing at the same surface from a different starting point.
The price collapse is structural, not promotional
Verified per-seat annual pricing, on a linear scale. The compression at the bottom is what makes the unbundling economically rational for a fund — not a discount cycle, but a data layer that anyone can license.
Koyfin is a 95% cut from Bloomberg for roughly 80% of equity-focused workflows. Fiscal.ai is 99%. Perplexity Finance is 100%. The moat left the data layer the moment the primary providers started licensing to anyone who'd pay — which means it has to come from the application layer, and that is a design problem before it is an engineering one.
The migration is measurable, and the incumbent is already installed
| What the market says | Figure | Why it changes the design brief |
|---|---|---|
| Funds planning to cut or drop terminal seats within 18 months | ~34% V | The demand for an alternative is real and quantified — this is a tailwind, not a bet. |
| AlphaSense ARR, and YoY growth | ~$500M · 73% V | The category leader is compounding faster than a startup can out-ship it on breadth. |
| AlphaSense penetration of the top 50 hedge funds | 70% V | They are already in the room during the sales conversation. The product is not being compared to nothing. |
| AlphaSense penetration of the S&P 100 | 88% V | Interface conventions are already set. Breaking them costs credibility unless the break is deliberate. |
Sorting the moats from the things that close in a year
The point of this table was to tell the design phase where to spend its best work — a moat gets flagship treatment, a closing advantage gets shipped fast and cheap.
| Capability | Verdict | Reasoning |
|---|---|---|
| Sector-templated analyst skills | Moat | Hand-curated by domain practitioners. Thins if a competitor commoditises it with generated templates, but the curation is hard to fake. |
| Proprietary content corpus | Moat | Genuinely differentiated content, conditional on the licensing holding up under scrutiny. |
| Generative custom dashboards | Moat now, feature in 12–24 months | Nobody in the competitive set ships persistent, live, natural-language-generated dashboards today. Several will. The clock is the design brief. |
| Custom data ingestion | Feature | Architecturally replicable by any well-funded competitor. |
| Structured financial model types | Feature | The maths is well known; only the integration into a conversational surface is mildly novel. |
| Third-party data integration | Feature | Anyone can license the same source. It also means data quality is bounded by a competitor's supplier. |
The flagship surface isn't the biggest one. It's the one with a 12–24 month window on it.
The method
Four documents before one pixel.
Sequenced, not a research dump. None could have been written first: the market map established what the product competes against, so the user work knew what to look for; the user work established who, so the workflow map knew whose day to model.
The destination is greyed on purpose. The work was pointed at screens the whole time — it just refused to start there.
Every claim carries its own confidence
Three grades, marked inline on every sentence that makes a claim. Not epistemics for its own sake — the flag tells the design phase how much design to spend.
Then I audited my own documents and published the result
The fourth document opens by grading the first three. This went to the founder, not into an appendix.
“Largely a structured hypothesis. No user interviews have informed it.”
— Foundation 04, section 1.1, on the workflow mapThe audit also names what each document doesn't give: the market map has no read on the client's own retention or customer identities; the user model contains nothing from a real user session; the workflow map can't say whether the five phases match how these particular customers work. Naming that is a scheduling decision — it puts the interview budget on the 15%, not the 75%.
Foundation 02 · the reframe
They aren't segments. They're modes.
The company was treating deep-conviction and generalist investors as two user segments — a targeting decision to be made, one to be picked. The user work found the same analyst operates in both, cycling between them within a single week: generalist screening on Monday, deep-conviction work on one name by Wednesday. Senior PMs run the inverse.
| Morning triage, by mode | Deep-conviction | Generalist |
|---|---|---|
| Names in scope | 5–20 I | 30–100+ I |
| Primary filter | Thesis-relevance I | Coverage-relevance I |
| Time budget | 30–45 min I | 60–90 min I |
| Alert sensitivity | Low — thesis-material only I | Medium — sector-relevant I |
| Sell-side reliance | Low — they're ahead of it I | Medium — useful as a first filter I |
| Idea generation cadence | 1–2 new names per quarter I | Weekly screening cycle I |
| Role of the screener | Confirmation, after a hypothesis exists I | Discovery — the screen generates the hypothesis I |
Two segments means picking one and building two products eventually. Two modes means one product where the switch is the thing you design.
A fork worth taking before the home screen exists, not after.
The home screen still has to greet someone in a default state — but that becomes one decision, not an architecture.Foundation 03 · Workflows & jobs-to-be-done
Five phases, five different clocks.
Not sequential steps — recurring loops at different cadences, each mapped with a formal job statement, its triggers, the tools in use today, its failure modes, and the white space a product could occupy. Nothing wraps all five, which rules out an entire family of navigation ideas in a single page.
The ragged, non-aligned rhythm is the finding. Anything that models the analyst's work as a linear process is dead on arrival — that's a week of exploration the design phase never had to run.
Deep research isn't one job. It's five sub-jobs, and only one has no tool at all
This table is where the hero insight came from. It wasn't a brainstorm — it fell out of columns.
| Sub-job | Core question | Tool coverage today |
|---|---|---|
| 1 · Understand the business | How does this company actually make money, and is that durable? | Partial — filings search exists; nothing synthesises a structured business model rather than a summary. I |
| 2 · Map the competitive landscape | Who does this compete with, and is the moat widening or narrowing? | Partial — document search is strong; synthesis into a structured competitive view is absent. I |
| 3 · Build the financial model | What's the range of intrinsic values across scenarios? | Partial — the model lives entirely in Excel; no natural-language layer connects to it. I |
| 4 · Primary research | What does primary evidence say that public data cannot? | Partial — expert networks and transcript libraries cover most of it. I |
| 5 · Write the thesis | What exactly is the market missing, and what would change my mind? | None. The most important sub-job has zero product support anywhere in the market. I |
The hero insight
The thesis is the connective tissue — and nothing supports it.
Every one of the five phases resolves back to the same object. Triage filters against it. Idea generation proposes new ones. Deep research builds it. Earnings processing tests it. Synthesis documents it. And it is the one object in the entire workflow with no dedicated product support anywhere in the market. It lives in Word, in Notion, or in the analyst's head.
A product that doesn't make the thesis a first-class, persistent, user-owned object will feel like five disconnected tools rather than one workflow. That sentence is the redesign's organising constraint, and it came out of a column in a table rather than a whiteboard — which is why it can be defended to a founder rather than pitched to one.
Two cross-phase patterns that constrain every surface
| What the analyst produces | Where it ends up | Who reads it |
|---|---|---|
| IC memo | Word document | Investment committee |
| Position brief | 1–2 page Word doc | Portfolio manager |
| Financial model | Excel — irreplaceable | PM, IC |
| Sector overview / heatmap | PowerPoint or Excel | PM, IC |
| Earnings update | Brief email or Slack note | PM |
The product is the path to the spreadsheet, not a replacement for it. V — every comparable product in the category ships an Excel add-in.
Foundation 04 · Synthesis & open questions
Research built to force a decision.
The first three documents were intelligence for me. The fourth was built to be used in a room — an operating manual for a 60-minute founder conversation, with a minute budget, a question order, and a note on which answers are load-bearing and which are only directional. Most research is written to be agreed with; that's the failure mode this document was designed against.
Every assumption paired with the question that kills it
Seven open questions, each stating the current assumption, the surface it shapes, what a real answer would change, and the exact words to ask. Four of them:
Walk me through where you keep your thesis on your highest-conviction position right now. What does that document look like?Shapes: thesis as a first-class object. If they have a format they trust, the product mirrors it. If it's in their head, the product defines it. Different design problems entirely.
When you were trading, was triage solved, or was it always frustrating?Shapes: the morning brief. If triage is already solved by a custom alert setup, the brief is a lower priority than we've been treating it.
What happens between “this is interesting” and “I'm building a model”?Shapes: skill templates and the deep-dive entry point. If the first look isn't a discrete sub-job, the templates are mis-positioned.
Show me your screen — what tabs are open right now?Shapes: everything. Reveals the actual stack, the actual workflow, the actual friction — rather than the one the analyst reports.
The ask, sized so that any partial yes is still useful
The interview request was structured in tiers rather than as a single request that could be declined once. Ask at Tier 1, accept gracefully down to Tier 3, request Tiers 4 and 5 as low-cost supplements regardless of the answer.
The gold. Two to four calls with current power users, 30–45 minutes each, recorded with consent.
One or two power users, plus one churned trial user. The churned user is disproportionately valuable — they say why it didn't stick.
One call with one current user. Anything beats zero.
Recorded sales calls, wins and losses, where buyers describe their own workflow. Not equivalent, but it extracts real workflow data.
A product analytics dump. Even rough numbers correct major assumption errors.
Framed as derisking, not as a research nice-to-have: a single 45-minute call can collapse two weeks of design uncertainty, and the asymmetry between the cost of the calls and the cost of a wrong inference is severe.
And a pre-written recovery plan for each way the research could be wrong
Written before the design phase, so a mid-engagement pivot is a lookup rather than a crisis.
| If we're wrong about… | What breaks | Recovery | Early warning |
|---|---|---|---|
| Which mode leads | Home screen and nav lead with the wrong default. First-session value drops sharply. | ~1 week | Interview day one. The question “what do you do first when you sit down?” surfaces it in fifteen minutes. |
| The commitment to workflow-native | The redesign becomes additive rather than transformational; the competitive positioning softens. | Reframe scope | First 30 minutes of the founder meeting. |
| The five-phase cut | Phase boundaries don't match reality. Skills appear in the wrong contexts. | ~1–2 weeks | The “show me your screen” question. |
| Shipped vs. aspirational capability | Surfaces get designed against capabilities that don't exist. The redesign becomes marketing. | Cut affected surfaces | First 30 minutes of the founder meeting. |
| Getting no interviews at all | Every inference becomes a load-bearing assumption with no validation. | Invest in the confident, soft-pedal the fragile | None. This is the worst case — if it happens, we know. |
What the foundation settled
Eight decisions committed without waiting for anyone.
The final section of the fourth document is the one that mattered most to the next phase: the decisions that no longer needed founder input, with the evidence behind each. These are the arguments that normally consume the first month of a redesign in review meetings. They were closed, in writing, with sources attached — so week one could build against them regardless of how the meeting went.
coverage not watchlist, name not stock, thesis as a first-class concept, comps, IC memo. Institutional, dense, citation-rich, no consumer warmth. High confidence, and independent of every other open question.Which resolved to a provisional IA before design started
- Todaythe brief — what changed on your names since you were last here
- Coveragethe analyst's book, as the primary navigation object
- Researchthe deep-dive surface, with chat as a primitive inside it
- Workflowssaved processes the analyst runs, surfaced in context
- Outputssaved dashboards, memos and models — all with export as a verb
Whether the default landing is Today or Coverage depends on the mode question. Whether Research is prominent or backgrounded depends on the chat commitment. The tree itself is sound.
Close
The design phase started narrow.
Four documents, sequenced so each one could exist. Seven market categories with verified pricing in every row. A user model that turned a targeting decision into a design problem. A workflow map that ruled out a whole family of navigation ideas in a page, and produced the one insight the product could be built around. A synthesis that named its own weak spots in writing and pre-wrote the recovery plan for each. Eight arguments closed with sources attached. The design principles document was deliberately held until after founder alignment — a principle the founder has signed off on settles ten screen arguments in advance; one he hasn't is a preference with better formatting.
The point of the research wasn't to know more. It was to make the option space small enough that design becomes a decision instead of a search.
Four documents, every claim graded, eight arguments closed — and the guessing named as guessing before anyone drew anything.
All work