AI that proves it's right.
Kepler is verifiable AI for finance. Every number is traced to its filing, page, and line item, and you get the same answer every time you ask.
26M+ SEC filings14,000+ companies120+ countries30+ languages
Revenue grew 15.1% to $402.8B. Operating margin was 32.0%, versus 32.1% a year earlier.
Built for the people who
have to defend the number
“I can't put a number in front of a client if I can't show where it came from.”
“Kepler finds where we've done this before and points to the exact document, so I can check it before it goes to a client.”
“I can see exactly where every number and conclusion comes from, which means I can defend the work.”
Why investment teams
choose Kepler
Alphabet's revenue was $402.8B in FY2025.
metric: operating_margin company: Alphabet Inc. (GOOGL) period: FY2025
Yes. Microsoft was about 67% of CoreWeave's revenue in Q3 2025.
“We recognized an aggregate of approximately 67% and 65% of our revenue for the three months ended September 30, 2025 and 2024, from our largest customer, Microsoft.”
=C3/C2- 1
Every number, traced to its source
Click any figure to see the filing, page and line item it came from.
- 2
The model writes the words. Code writes the numbers.
Code pulls the data and runs the math, so no figure is ever guessed.
- 3
Every document, read in full
Kepler reads and indexes everything before you ask, so facts buried deep in a filing or in your own files get found.
- 4
Your definitions, built in
Correct a treatment once and it sticks for the analyst, the desk and the firm.
- 5
Works where you already work
Use Kepler as an add-in in Excel and PowerPoint, as a verified connector inside Claude, or through the Kepler MCP and API from your own agents and systems.
From question to
defensible deliverable
Ask for the model and the workbook comes back with every cell traced to the filing it came from.
Building Google's 3-statement historical model now.
- Identify Alphabet / Google
- Build 3-statement model (last 3 fiscal years)
Alphabet's 3-statement model is ready: Income Statement, Balance Sheet, and Cash Flow Statement across FY2023, FY2024, and FY2025, each cell cited to its source 10-K filing.
There is a trust gap
in enterprise AI
Why not just use a frontier model? Because three things still stop a firm from relying on one when a wrong number can cost millions, however smart it gets.
It can't prove a number.
Models are probabilistic. A wrong figure looks exactly like a right one, so your team checks every number by hand.
It doesn't know your firm.
Not your data, your definitions or your relationships. It answers in its own terms.
It doesn't read everything.
It answers from the few passages it retrieved. The rest of your documents never enter the answer.
We close all three
with one architecture
From the filing to your screen, every layer is built to read everything, know your firm and prove every number. Not one of them trusts a model with a number.
Primary sources
We read the filing. Not a summary of it.
Everyone else retrieves chunks of whatever text they scraped.
Kepler captures every source whole: each SEC filing with its full text and XBRL, earnings call transcripts, investor presentations, market data, and your own documents. Nothing downstream works from a paraphrase.
How it's built- SEC filings stored complete: primary document plus XBRL
- Earnings call transcripts, investor presentations, and market data
- Your own documents from your firm's document stores, or by upload
Ontology
We know who every filing is about.
Everyone else matches strings and hopes it's the right company.
Companies, subsidiaries, funds, people, and filers are resolved into one graph. Every relationship in it is a claim: one filer's statement, with the verbatim sentence attached.
How it's built- Names resolved to SEC company records: NVDA, NVIDIA, and NVIDIA CORP are one company
- Every relationship carries its filer and source sentence, checked before it counts
- Relationship types follow FIBO, the financial industry's standard ontology
Deterministic engine
Code does the math. Every time.
Everyone else lets a language model do arithmetic.
Financial statements are assembled from the filings' own XBRL with no model in the path. Every calculation runs as code over cited inputs and produces a derived citation that carries its formula.
How it's built- Every displayed number is a value a cited filing reported, or a sum of them
- As-filed or as-restated, with point-in-time history for any period
- Ratios, growth, and multiples computed in code, never by the model
Citations & verification
No receipt, no number.
Everyone else attaches a link and calls it a citation.
Every number is written with its source attached and checked against that source while the answer streams. A citation that doesn't match is stripped before you ever see it.
How it's built- Fourteen source types: filings, exhibits, transcripts, slides, workbooks, your documents
- Anchored to the exact XBRL fact, cell, or verbatim sentence
- Quotes must match the source word for word, or they are dropped
Agent runtime
A plan, not a prompt loop.
Everyone else lets an agent improvise until it stops.
Every request starts with a plan. Agents run as state machines that give each step only the tools it needs, and specialist agents read filings, transcripts, exhibits, and decks in parallel.
How it's built- Plan, research, execute, present: tools gated at every step
- Specialists for 10-Ks, transcripts, exhibits, investor decks, workbooks
- Search over each firm's own index; regression-tested against baselines
Skills & rules
Your definitions, enforced in code.
Everyone else treats your instructions as a suggestion.
Built-in skills encode how finance work is done. Your desk writes its own, and Kepler compiles, reviews, and guards each one, so it can change a definition but never the rules of evidence.
How it's built- Built-in skills for adjusted EBITDA, enterprise value, fully diluted shares
- Write a skill from a short brief; it is compiled, reviewed, then applied
- Every skill records who authored it: you, or your firm
Surfaces
One engine, wherever you work.
Everyone else ships a chat box and calls it a platform.
The Kepler app for research, workbooks, and company pages, a verified connector that brings the same engine, with the same receipts, into Claude, and the Kepler MCP and API for your own agents and systems.
How it's built- Chat, workbooks, and a citation sidebar in the Kepler app
- A verified connector for Claude, with secure sign-in
- Workbooks export to Excel with formulas and sources intact
Security & tenancy
Isolation by architecture, not by policy.
Everyone else promises it in a PDF.
Each firm's work is scoped to its organization, market data and customer data live in separate stores, and the systems that run agents are never exposed to the internet.
How it's built- Org-scoped projects and per-firm search indexes
- Customer data and market data kept in separate databases
- Execution runs on a private network behind the API
An ontology built
from your firm's data
Every company, fund, person and document in your data, resolved to one record with a source behind every connection. The first public equities ontology, below, is built the same way.
Explore the ontologyBuilt for regulated finance,
secure by design
- No training on your data
- Strict data isolation and boundary enforcement
- Least-privilege access and controlled permissions
- Full auditability of system behavior
- Deterministic and reproducible outputs
- Designed to withstand scrutiny