Prioritize
A strong, verified team or market, traction at the level of the stage, no weak criterion.
METHOD
A written VC method,
applied to every deal.
Casian pairs AI models, which read the documents and run the research, with a method written in code. Sources are required, citations are checked and the screening opinion is computed by a fixed grid.
From deck
to decision.
Every deal goes through the same stages, in one workspace. The team stays in charge of the thesis, the memo and the decision.
Thesis
Your team writes the fund’s rules, preferences and exclusions. Only the published version applies.
Screening
Within minutes: an opinion, sourced key figures, gaps between the deck and public sources, and questions for the first call.
Due diligence
An agent reads the whole data room, investigates six themes, writes a sourced note and lists the open questions for the founders.
Memo
The due diligence note becomes the investment memo. The team edits it and exports it to Word or PDF.
Decision
Each member votes, then an authorized member records the official decision, with its rationale and conditions.
ChatAt every stage, the chat answers questions about the deal and cites the documents. It can also move a deal forward or re-file a document, with undo.
Screening
in seven steps.
Three steps use an AI model. The other four are code, which applies the same rules to every deal.
- 01
Public brief
AIThe model reads the documents, with no web access, and lists what can be checked in public. Code removes anything that looks like a confidential figure from the brief.
- 02
Research
AIUp to seven parallel searches: each founder, the market and competitors, the company and its traction, public figures. LinkedIn, paid databases and pages behind a login are excluded.
- 03
Assembly
CodeFacts, figures, competitors and fundraisings without a source are dropped. A fact about a founder is kept only when their identity is confirmed. Sensitive personal data is removed.
- 04
Write-up
AIThe model writes the analysis from the documents and the research, with no web access. It rates six criteria, each with a level and an evidence status: verified, declared or unknown.
- 05
Citations
CodeEvery document citation carries the exact excerpt, compared with the text at the cited location. A citation that does not match is removed, and a section with no source is dropped.
- 06
Opinion
CodeThe grid turns the six criteria into an opinion, the same way every time.
- 07
Checks
CodeTen automated checks review the result. Failures are stored with the analysis.
Four possible opinions,
one fixed grid.
The model rates the criteria, the code applies the grid. Missing information never counts as a strength.
- 1Team and founder-market fit
- 2Market
- 3VC potential
- 4Product and differentiation
- 5Traction for the stage
- 6Thesis fit
Follow up
A solid startup that is not ready yet. The opinion sets a review date, in 3, 6 or 12 months, and the condition to move to Prioritize.
Pass
A mandatory thesis rule breached, with evidence, a verified blocking red flag, not the shape of a VC investment, a weak team or market with no strength to offset it, or no verified strength.
Incomplete analysis
A criterion could not be rated. A deal is never passed for that reason.
Due diligence
run by an agent.
The agent works in a sandbox, with the data room mounted read-only. Each step starts a fresh thread and writes its results to files that the next step reads.
New documents
When new documents arrive, the agent rereads only those, restores its saved state and updates the analysis.
The team’s text
A section rewritten by an analyst is never overwritten by a later run.
Extraction
Text, pages and tables are extracted with their location, with no AI model.
Reading
6 in parallelSix readers write one note per document. The run stops if more than a quarter of the documents cannot be read.
Cross-checks
Gaps between documents, then an investigation plan.
Investigations
6 in parallelSix themes. Five can search the web; finance works from the data room only.
Note and opinion
The investment note and Casian’s opinion: invest, invest with conditions, defer or pass.
Control
Document coverage, the log of checks and corrections to the note.
Questions
Answers to open questions, from the data room and the web, then up to ten new questions for the founders.
Your thesis,
applied as written.
The fund’s criteria are versioned rules. Only the published version applies, drafts are ignored, and a fund without a thesis is never blocked.
- Each rule is assessed one by one: met, breached or unknown, with the document or source behind the judgment.
- Missing information never counts as meeting a rule. It becomes a question.
- A rule that describes how to run the analysis applies to Casian’s work, never to the company.
| Type | Example | Effect on the opinion |
|---|---|---|
| Rule | Seed and Series A only | Mandatory. A sourced breach gives Pass, out of thesis. |
| Exclusion | No gambling companies | Mandatory, like a rule. |
| Preference | Recurring revenue | Counts toward thesis fit, without passing the deal. |
Ten checks
on every screening.
After each screening, code reviews the result. A failed check is stored with the analysis and used to improve the method.
- 01Key figures dated and ranked by weight
- 02Data age consistent
- 03Every contradiction listed
- 04Every founder researched
- 05Sectors consistent with who pays
- 06No investor shown as a competitor
- 07At most five call questions
- 08No sensitive personal data
- 09A two-to-four-sentence overview
- 10An opinion equal to a fresh computation
A tested method
Every rule of the screening method has an automated test on a fictional deal. A test fails if any prompt contains the name or a figure from a real deal, which prevents tuning the method to a single case.
Stable opinions
Before each release, a test recomputes opinions already issued and requires the same results.
A few
details.
What is the analysis built on?
On AI models from a specialized provider, called through its API, and on a VC method written in code. The models read, search and write. Code requires sources, checks screening citations and computes the opinion.
Is the AI trained on our data?
No. The analysis adapts to your fund through the published thesis and the deal’s documents, with no training on your data. Under its API terms, the model provider does not use the data it receives to train its models.
How can an analysis be checked?
A citation opens the document at the cited page. A web source carries its address, its date and its status: verified, contradicted or uncertain.
What if information is missing?
It never counts as a strength. Screening adds a question for the call, and due diligence lists the open questions for the founders. The analysis can be wrong, and your team checks findings and sources before adopting a conclusion.
Where do our documents go?
The Security page covers hosting in Paris, isolation between funds and the AI model gateway.
How do we get started?
We agree on the scope, access, processing terms, timeline and onboarding with your team. Documents are imported into Casian; automatic collection from external tools is not included in the current pilot.
Your next deal.
A shared place to think.
We open access in small batches.
Each fund starts with a deck it has received.