The AI deal desk that does the work, and learns how you do it.
Cited analysis from SEC filings, and an agent that tracks every email, answers what it can, asks you what it cannot, and earns the right to send on its own. For boutique advisors, emerging managers, founders, and the students who will join them.
No credit card. Nothing is sent from your mailbox unless you switch that on.
- 328
- tools across 8 careers
- 18,037
- startups indexed
- 230
- AI workflows
- 3
- learning loops per account
How it works
One loop, from the filing to the follow-up: signal, analysis, action, memory.
Signal
New SEC Form D raises every business day, 8-K events decoded, insider trades, and funding news matched to the people in your inbox.
Analysis
Comps, precedents, capital structure, memos and QoE flags built from XBRL and filing text in minutes. Every figure links to its filing.
Action
An agent that tracks every thread, drafts the reply, runs outreach and keeps the pipeline current. It sends on its own only where you allow it.
Memory
A playbook of your answers, lessons from your edits, and trust earned email by email. It needs less of you every week.
The relationships agent
Your inbox, handled the way you would handle it.
A founder sending cold emails, answering prospects and replying to the team. An investor screening pitches. The agent works in your mode, your voice and your rules.
Tracks every email
Connect Gmail, Google Workspace, iCloud, Yahoo, Zoho or Fastmail with an app password. It reads new mail every five minutes, files who wrote and what they want, and keeps a pipeline: lead, contacted, engaged, meeting, proposal, won.
Answers what it can
Prospects, customers and your own coworkers get a reply drafted in your voice. You choose, per kind of email, whether it waits for you or goes on autopilot.
Asks what it cannot, and remembers
When a reply needs something only you know (a price, a date, a yes), it asks you one short question. Tick “remember” and the next person who asks gets the answer without bothering you.
Runs outbound
Build a list from 18,000+ startups or your own, qualify it against your ideal profile, and work through a personalised sequence that stops the moment someone replies or opts out.
Your queue, this morning
The agent needs your input
What pricing should we quote for 10 entities on NetSuite?
Your answer finishes the reply to Dana, and is remembered for the next person who asks.
Reply to coworkerhigh confidence
Tell them we integrate natively with NetSuite and QuickBooks Online…
Autopilot sends this in 4 min unless you stop it.
Replymedium confidence
Autopilot left this for you: The agent's confidence is medium, not high; It touches money, terms, legal or something sensitive.
Studio
The model and the deck, built live while you watch.
Analysts spend their nights spreading comps, rebuilding models, pasting tables into slides and tying out every number. Studio does that work in a live workbook and deck that stay linked, and checks it the way a reviewer would.
Watch it build
Ask for a valuation pack and watch the DCF, the comps and the pitch book fill in from SEC filings, a cell at a time, with the agent's cursor on screen. Stop it at any moment; undo any run in one click.
Model and deck, linked
Tables, charts and figures on slides are links into the model. Change the WACC and every page moves with it; the tie-out flags any number on a slide that does not trace back to a cell.
Models a VP would sign
DCF, LBO with a real debt schedule and its circularity solved, trading comps, merger and cap table templates. Blue inputs, black formulas, green links, sourced historicals.
The checks you do by hand
Every run ends with an audit: numbers typed into formulas, overwritten formulas, broken row patterns, errors, circular references. Banker formatting and data tables in one click.
Page turns, turned
Leave comments on cells and slides, or photograph the MD's marked-up printout and every pen mark becomes a comment on its cell or slide. The agent makes each change and resolves each comment, saying what it did.
Inside Excel and PowerPoint
Install the add-in and the agent works in your own workbook: each cell lands in Excel as it writes, your edits sync back, and decks in PowerPoint refresh in place when the model moves.
Data room to model
Drop in a CIM or audited accounts and its financial tables arrive as a sheet of inputs, every figure copied as printed and tagged with its page. Upload a seller's model and get an intake report first.
Before it goes out
A brand check for stale cover dates, missing source lines, text off the page and off-brand colours, with one-click fixes. Checkpoints show exactly what moved since the version the MD saw.
“Build a valuation pack for Snowflake with a deck”
- Building the model from SEC data · DCF, comps and a summary sheet, from XBRL company facts
- Writing cells · DCF!C5:G20: revenue, margins, free cash flowAgent
- Computing a data table · Implied price across WACC and terminal growth
- Adding slides · Overview, comps, DCF, football field, sensitivity
- Auditing the model · No errors; no numbers typed into formulas
- Tying out the deck · Every figure on every slide traces to a cell
Every change is logged with its undo. Work in the browser, or inside Excel and PowerPoint with the add-in; files export with formulas and native charts intact.
The terminal
Terminal-grade functions, with the models shown.
More than thirty-five functions on familiar codes: GP BETA RISK WACC IRAT QUAL FCST EE ANR DVD GC ECO WEI EQS PORT. Each analytics screen shows its method and sources, and the assistant uses the same models, so an answer about risk or credit names the model behind every number.
Risk with its error bars
GARCH volatility forecasts and a price cone, betas four ways with their standard errors (including Welch's, the best predictor of future beta), and value at risk three ways with a backtest that says whether it held.
Credit without a ratings licence
An implied rating from Altman's Z'', Ohlson's O-score and Merton's distance to default, with historical default rates, for every US filer. Where a model does not fit a company, it is left out and the screen says why.
Forecasts with honest intervals
Revenue and the economy forecast with the methods that won the M4 competition, and bands sized by how wrong the method has actually been, with the coverage shown. Set beside the Street's consensus.
Valuation as a range
WACC from its parts with a Monte Carlo range, and a Monte Carlo DCF with correlated inputs, in the terminal, the tool library and inside any Studio model: P10, P50, P90 and what drives the spread.
Rates and the economy
The Treasury curve with its Nelson-Siegel factors and the New York Fed's recession probability; BLS inflation and jobs with model outlooks and the Sahm rule. Markets boards read each move in its own volatility.
It learns what you know
Knowledge tracing follows how well you know each function: hints fade as you master them, and the next functions unlock when you are ready. Ask for a quick quiz, or an AI read of any screen.
The adaptive engine
Autonomy is earned, not assumed. Try it: this runs the product's own code.
A language model's confidence in itself is not evidence. YouBank decides what to automate from what you do: every draft you send unchanged or edit, and every reply your outreach earns. Three learning loops, each with its uncertainty shown rather than hidden.
Replies to prospects the agent rates “high confidence”
LearningThe curve is the engine's belief about how often you would send this kind of draft exactly as written. It starts flat, since only your own decisions count, and narrows with every one you make.
- Decisions
- 0
- Sent unchanged
- 0
- Estimate
- 50%
- 90% interval
- 5%–95%
- Would need edits
- ≤ 90%
Autopilot is offered only when, after at least 20 of your own decisions, the engine is 90% confident at most 10% would need your edits: about 22 clean drafts in a row. A light edit counts against it, as does any change to a number, date or link. Stop three automatic sends in a row, or start editing more, and it hands that kind of email back.
Earned autonomy
Your own decisions, per kind of email, certified with exact bounds: autopilot is offered once the engine is 90% confident at most 10% would need edits. Spot checks keep it learning; stopped sends, a rising edit rate or time hand it back.
Outreach experiments
Thompson sampling over opening angles and send times, rewarded by human replies and penalised by opt-outs, learning from day one by counting unanswered sends as partial misses. One decision in ten explores at random, and every choice is logged.
Lessons from edits
Durable preferences inferred from what you change, applied once seen twice or confirmed, plus your own most similar emails as examples of tone. Lessons shape wording only; they never authorise a fact or relax a rule.
Methods: W. R. Thompson, Biometrika (1933); O. Chapelle and L. Li, NeurIPS (2011); D. Russo et al., “A Tutorial on Thompson Sampling” (2018); C. Vernade, O. Cappé and V. Perchet, UAI (2017), arXiv:1706.09186, for delayed rewards; A. Angelopoulos et al., “Learn then Test”, Annals of Applied Statistics (2025), arXiv:2110.01052, for certified error rates; G. Gao et al., PRELUDE/CIPHER, NeurIPS (2024), arXiv:2404.15269, and PROSE, ICML (2025), arXiv:2505.23815, for learning from edits.
Guardrails before autonomy
An email sent in your name cannot be taken back, so everything doubtful comes to you with the reason written down. Mailbox passwords and tokens are encrypted with AES-256-GCM; disconnecting deletes them.
- Autopilot is off until you switch it on, and each kind of email has its own setting: off, ask me, or autopilot.
- It never sends what it is unsure of, anything sensitive, anything with a blank to fill, or anything that needs your input. It says why.
- It never sends a link, email address or account number you have not given it yourself, so an inbound email cannot steer it.
- Automatic emails wait out a hold you can cancel, go only in your sending hours, and stop at a daily limit. About 1 in 5 still comes to you as a spot check.
- It never answers newsletters, auto-replies or no-reply addresses, and before sending it re-reads the live thread: if you already replied from your phone, its draft is withdrawn.
- Campaigns skip EU and Canadian recipients unless you confirm consent, and nobody gets more than one first cold email through YouBank a month.
- Every email goes from your own mailbox. Regulated mode turns autopilot off entirely, and the audit log exports every draft, who sent it and how much it was edited.
- Your data trains no model. What the engine learns is statistics and plain-language lessons in your account, which you can read and retire.
Live demos
Real data, real output. Nothing here is a screenshot.
Built from a frozen snapshot of the production database, 2026-09-21. Signed in, these run live.
| Company | Quarterly revenue | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| PLTRPalantir Technologies▲1.23% | $407.9B | $398.5B | 64.7x | 149.7x | 79% | 85% | 55% | 133 | |
| NETCloudflare▼2.87% | $115.0B | $114.1B | 45.4x | NM | 34% | 73% | 14% | 47 | |
| CRWDCrowdStrike Holdings▼3.73% | $242.0B | $237.7B | 44.1x | NM | 24% | 75% | 28% | 53 | |
| SNOWSnowflake▼2.19% | $115.2B | $115.2B | 21.2x | NM | 32% | 67% | 22% | 54 | |
| DDOGDatadog▼2.94% | $81.8B | $77.8B | 19.6x | 994.2x | 32% | 80% | 27% | 59 | |
| MDBMongoDB▼2.09% | $30.9B | $28.7B | 10.3x | 3411.4x | 25% | 73% | 24% | 49 | |
| NOWServiceNow▼2.06% | $140.1B | $140.8B | 9.6x | 51.2x | 22% | 75% | 31% | 53 | |
| GTLBGitLab▼2.36% | $8.3B | $8.1B | 7.7x | NM | 23% | 86% | 20% | 43 | |
| ESTCElastic N.V.▼2.66% | $9.0B | $8.1B | 4.5x | NM | 16% | 76% | 19% | 36 | |
| Peer median | 19.6x | 25% | 76% | 27% | 53 |
Built for the long tail of dealmakers
Enterprise-grade analysis and follow-through, priced for the firms that cannot justify an enterprise seat.
A general AI assistant will build a DCF for $20 a month, but not from verified data, and not the follow-through: on September 2026's Excel modelling benchmark, frontier models passed only 64% of numerical checks.1 The platforms that bring the data sell annual contracts in the thousands to tens of thousands of dollars.2 YouBank is cited, filing-backed work plus the agent that acts on it, for the teams in between.
Boutique advisors and placement agents
Comps, buyer lists and teasers without a Capital IQ budget, and an outreach desk that keeps every mandate moving.
Emerging VC and PE managers
Form D and funding signals on the companies you track, cited diligence, and a pipeline that files itself from your inbox.
Founders raising or selling
Cold outreach, investor and customer follow-up, and replies to your own team, with autopilot where you have earned it.
Students recruiting into finance
Real comps and DCFs on real filings, deal walk-throughs from 8-Ks, and the same desk you will use on the job.
Eight careers, each with its own researched toolkit
230 AI workflows that research and draft, and 98 calculators that compute exactly.
Pricing
Start free. Pay for the AI you need.
Every plan has the terminal and the public data. Paid plans add a larger monthly AI allowance and the premium features; the allowance is what keeps prices honest, because model time is most of what YouBank costs to run. Prices are in US dollars, before tax. Beta accounts will be told well before anything is billed.
Campus
Free
with a .edu address
- The full terminal and data
- AI workflows with a monthly allowance
- Recruiting pack: practice comps, deal walk-throughs
- AI allowance: $10 a month
Pro
$59
per month, or $49 billed yearly
- Everything in Campus, with a larger AI allowance
- Relationships agent and one mailbox
- Nurture, signals, compose and AI meeting briefs
- Edge's premium layers: deal re-reads, claim cross-checks, the daily Pulse, the Radar screener, five theses, live voice
- Top up with AI credit packs when you need more
- AI allowance: $25 a month
Deal Team
$149
per seat per month, or $125 billed yearly; three seats minimum
- Autopilot and campaigns
- The adaptive engine across the team
- Shared workspaces, team theses and live collaboration
- The meeting notetaker, Planet imagery in 3D, Dune queries and Eva-4B on calls
- Assign seats to people on your team
- AI allowance: $60 a month per seat
Enterprise
$299
per seat per month, billed yearly; five seats minimum
- Regulated mode and audit exports
- SSO, admin controls, data residency options
- Bring your own data licences (Nightfire flare volumes)
- GPU retraining of the deal model
- AI allowance: $130 a month per seat
AI credit packs from $10 top up any plan. Compare every plan, the packs and the FAQ →
Make it yours
Fourteen styles, from amber-on-black terminal to soft daylight.
Density, corner radius, glass and glow all change with the style, not just the colors. Try one now: this page will change with you.
Data and method
Auditable by design.
- Fundamentals
- SEC XBRL company facts. LTM = fiscal year plus year-to-date less prior year-to-date, with the concept the filer actually used, restatements deduped by accession.
- Filings
- EDGAR submissions and full-text search across every filing and exhibit since 2001: merger proxies, credit agreements, indentures, comment letters, 8-K item codes.
- Private markets
- 18,037 startups from Y Combinator, a16z, Thiel Fellows, Show HN, SEC Form D and AI web discovery, refreshed nightly.
- Prices
- Quotes, market cap and 52-week range from a market data API. Consensus estimates are not licensed, so NTM figures are entered by you and marked as manual.
- The adaptive engine
- Per-account statistics you can inspect: how often you send each kind of draft unchanged (certified on your own decisions with exact Beta bounds), Thompson sampling over outreach choices rewarded by human replies, and plain-language lessons from your edits. No model is trained on your data.
- Limits we state plainly
- Reported EBITDA is operating income plus D&A, not company-adjusted. Fiscal years are not calendarized. The agent's confidence is its own estimate, which is why autopilot also has to earn your trust.
Will it send emails without me?
Only if you switch autopilot on, only for the kinds of email you choose, and only when a draft passes every check: high confidence, nothing sensitive, nothing it needs to ask you, inside your hours and daily limit. Until then every email waits for you to press Send.
How does it learn?
From what you do. Each draft you send unchanged or edit, and each automatic send you stop, updates how far it can be trusted with that kind of email; autopilot is offered only once it is 90% confident at most 1 in 10 would need your edits. Edits become lessons, your own past emails set the tone, and replies to outreach tell it which openings and send times work. All of it is visible and reversible in the app.
Is it suitable for a FINRA-registered firm?
Regulated mode is built for it: autopilot cannot be switched on, every email is sent by a person from their own mailbox (so the firm's archive captures it), and every AI draft can be exported for supervision. Your compliance team decides; we give them the controls and the log.
Can it build my models and decks?
Yes, in Studio. Describe what you need, or start from a DCF, LBO, comps, merger or cap table template filled from SEC filings, and watch the agent build the workbook and the slides. Slides stay linked to the model, every run is audited and can be undone, and it all exports to Excel and PowerPoint. You can also upload your own workbook.
Does it work in my own Excel and PowerPoint?
Yes. Install the YouBank add-in (Excel and PowerPoint on the web, Windows or Mac, with Microsoft 365 or Office 2021 and later) and connect it with a code. Link a workbook and the agent writes into it cell by cell while you watch; your edits sync back to YouBank; decks in PowerPoint refresh from the model. An IT admin can deploy it to a whole team.
Where do the numbers come from?
SEC EDGAR and a price API, with the source link on every figure. Anything from the web is cited. If a number is derived, the method is stated next to it.
Is this a Bloomberg replacement?
No. Bloomberg's edge is licensed real-time market data, chat and fixed-income depth. YouBank's edge is the free regulatory corpus plus an agent that produces the deliverable and does the follow-through.
What does it cost?
There is a Free plan, Campus is free with a .edu address, and Pro is $59 a month ($49 billed yearly). Each plan includes a monthly AI allowance that resets on your billing date. When it runs out, an AI credit pack (from $10) keeps AI going; otherwise AI pauses until it resets and everything else keeps working. Beta accounts will be told well before anything is billed.
Tell it what you do. It learns the rest.
A two-minute survey sets your desk. Connect a mailbox when you are ready, and give the agent as much or as little autonomy as it has earned.