A startup enters the pipeline. Its website tells one story. Its pitch deck tells a more detailed one. Public posts add context and, sometimes, contradictions. Before anybody can decide whether the opportunity deserves deeper attention, somebody has to find, read, reconcile and structure all of it.
For this workflow, built around international technology startups, the goal is not to replace the investor.
It answers a narrower and more useful question:
Is this startup worth human diligence time now?
IZZY automated that first pass. A record in the CRM triggers public-source research. A pitch deck received later is extracted and mapped into a predefined structure. Models from different providers perform different jobs. A confidential question framework constrains the analysis. The output is an initial yes, no or possible classification.
It is not an investment decision. It is not automated due diligence. It is a work filter before the work that can commit the fund begins.
The client, investment strategy, companies, model providers and exact criteria remain anonymous.
The answer in 60 seconds
- A minimal startup record is created in the CRM.
- The workflow gathers relevant public information, with X and the startup’s own website among the source categories used.
- Materials supplied by the startup, usually a pitch deck, are extracted and stored in a predefined structure.
- Models from different providers perform complementary functions such as source-specific analysis, document extraction, structuring and evaluation.
- A predefined list of questions receives one of three bounded answers:
yes,noorpossible. - The resulting classification decides whether deeper human review should begin; it does not decide whether to invest.
- The project team reports that the full first-pass journey moved from work measured in days to work measured in minutes. The underlying measurement record is not public.
The lesson is not “use more models”. It is give every component a defined job and never let the classification lose its sources.
In this article
- The hidden cost is the work before conviction
- Build the dossier before producing the rating
- Why the models specialise instead of voting
- Turn private criteria into executable questions
- The classification is a work threshold, not an investment decision
- What “from days to minutes” really means
- When this system helps, and when it becomes dangerous
1. The hidden cost is the work before conviction
The initial cost of a deal is not simply reading the deck.
The team has to locate dispersed information, separate assertions from evidence, identify missing data, place the company in context and apply a consistent set of questions across different formats.
This work is necessary. It is not always the best use of a senior investor’s time.
A weak opportunity can require nearly as much preparation as a promising one. Before anybody reaches a useful view, the team may already have spent hours on:
- opening several public sources;
- reading presentations built in different ways;
- recovering information from PDFs;
- translating it into the fund’s internal structure;
- checking the same recurring criteria;
- preparing a dossier that can be compared with other opportunities.
The workflow automates that preparation. It does not automate the elements that make an investment judgement credible: market understanding, people, timing, risk, relationship and conviction built through direct investigation.
2. Build the dossier before producing the rating
A score produced too early is a compact opinion about badly organised evidence.
The system therefore starts with the dossier.
Stage 1: create the minimum record
The first available information about the startup enters the CRM. It is enough to identify the company and trigger the expected research, but not enough to produce a classification.
Stage 2: enrich it with public sources
The workflow gathers the relevant public material. In this case, X is an important source category, alongside the startup’s website and other accessible pages selected for the process.
Collection is not verification. A social post may be promotional, old, ironic or incomplete. The system needs to preserve source, date and context rather than collapse every item into a conclusion.
Stage 3: process founder-supplied material
When the startup provides more information, usually a pitch deck, a second layer begins. The document may contain text, tables, charts and pages whose visual structure carries part of the meaning.
The workflow extracts the material, maps it to predefined sections and adds it to the existing dossier. The job is not to turn the deck into a ten-line summary. It is to attach each piece of information to the question it can genuinely support.
Stage 4: evaluate the structured dossier
Once the material is organised, the evaluation layer works through the fund’s question list. Each answer should remain connected to the underlying evidence and retain an uncertain state when the material cannot support a binary conclusion.
minimum CRM record
│
├─> public sources ─> structured facts and signals
│
└─> supplied documents ─> extraction / OCR ─> dossier sections
│
v
predefined question list
│
yes / no / possible
│
deeper human review threshold3. Why the models specialise instead of voting
“Multiple models” often suggests a panel: ask every model the same question, count the votes and call the majority more reliable.
That is not the architecture used here.
Different providers are present because the tasks differ. One model is suited to analysing a particular public environment. Another handles other source categories. A specialised component recovers information from documents when OCR is required. Another contributes to structuring or the final evaluation.
They do not compete to manufacture consensus. They pass structured work from one stage to the next.
| Function | Input | Required output | Failure to avoid |
|---|---|---|---|
| Source-specific collection | One public source category | Dated, attributable evidence | Treating visibility as proof |
| Document extraction | Pitch decks and supplied files | Recoverable text and data | Inventing content from an unreadable page |
| Structuring | Heterogeneous evidence | A dossier organised by question | Erasing contradictions |
| Evaluation | Dossier and defined questions | Bounded answers with rationale | Converting uncertainty into confidence |
The number of models is not the product. The asset is the interface between stages: what each one receives, what it must return, which evidence must survive and which condition must produce possible instead of a forced answer.
Recent research also explores multi-stage startup evaluation and due-diligence systems. One published framework explicitly flags missing data instead of generating unverified figures (arXiv). That supports the relevance of the engineering problem; it does not independently validate this client workflow.
4. Turn private criteria into executable questions
A sophisticated score cannot rescue a vague question.
The client uses a confidential list of questions. The individual criteria and their exact number are not public. We can disclose the shape of the contract:
- each question addresses an identifiable issue;
- each answer is
yes,noorpossible; - a question the dossier cannot answer is recorded as missing, not silently treated as a negative or a positive;
- the rationale remains connected to evidence in the dossier;
- the combined answers produce a bounded first-pass classification.
The third state matters.
Possible does not mean “the model feels slightly unsure”. It means the dossier does not yet justify a binary conclusion. The correct next step may be to request evidence, verify a claim, arrange a conversation or assign the question to a person.
Without that state, the workflow is rewarded for sounding decisive. False certainty accelerates the wrong part of the process.
5. The classification is a work threshold, not an investment decision
The output resembles a verdict: yes, no or possible.
But the decision is deliberately narrow:
Should the team invest more human time in this dossier?
It does not answer:
- whether the fund should invest;
- how much it should commit;
- which terms it should negotiate;
- whether these founders are the right partners;
- whether the underlying claims have survived deep verification;
- whether legal, financial or technical risk is acceptable.
A positive classification opens human investigation. It does not replace it.
That boundary protects the workflow from an impossible promise. Investment decisions can depend on private conversations, unavailable information, market changes and judgement that no submitted file contains.
The system’s value is to give a person a structured dossier sooner, with the recurring questions already traversed and the uncertain areas made visible. Keeping the output bounded and the human gate explicit is the same rule we apply in our AI pilot-to-production checklist.
6. What “from days to minutes” really means
The project team reports that the full path from initial collection to first-pass classification moved from work measured in days to work measured in minutes on average across a startup sample.
The sample size, period, logs and measurement protocol were not released for this article.
The defensible public statement is therefore:
A first pass previously measured in days is now measured in minutes.
This is:
- a first-party operating observation attributed to the project team;
- limited to the path from initial collection to first classification;
- not independently audited;
- not evidence of better investment quality or fund returns.
Saving preparation time does not prove that judgement improved. It changes when qualified human judgement can begin.
7. When this system helps, and when it becomes dangerous
AI-assisted pre-screening is useful when:
- dossier volume consumes substantial preparation time;
- the same question families need to be applied consistently;
- source provenance can survive collection and transformation;
- uncertainty is accepted as a valid output;
- a person takes over before any investment decision.
It becomes dangerous when:
- an opaque score replaces the dossier;
- missing information is converted into a convenient answer;
- a public post is treated as a verified fact;
- multiple models are used to stage an artificial consensus;
- the system automatically excludes a company without an appropriate review path;
- the team confuses pre-screening, due diligence and investment judgement.
Current investor discussion reinforces the boundary. If you only analyse a pitch deck with AI, you are almost certainly missing something, as investor Sarah Barber of Jenson Ventures put it in a UK Business Angels Association review of AI in fundraising. That is an argument for better provenance and human follow-through, not for asking a model to invent the missing context.
Conclusion: automate access to judgement, not judgement itself
The theatrical design would ask several models whether the fund should invest and present their consensus.
The useful design is less visible: collect the right sources, extract the documents, preserve contradictions, apply consistent questions and show what remains unknown.
In this project, the models do not sit around a virtual investment committee. They occupy different stations in a preparation line. The result is not artificial conviction. It is a dossier that reaches the person capable of building real conviction sooner.
The pattern is not specific to investing. Any recurring screening step, such as supplier vetting, grant applications, candidate pre-selection or inbound lead qualification, has the same shape: scattered sources, a document to extract, a fixed list of questions and a person who should only see the dossiers worth their time.
Bring us your first filter, not your final decision
Show us how a dossier enters today, which sources analysts open, which questions recur and when a senior investor decides to continue.
IZZY can map the collection, documents, rules, model roles and human gate, then tell you which part is genuinely worth automating.
This is exactly the scope of our n8n AI Automation service.
Frequently asked questions
No. The workflow prepares a first dossier and a classification used to decide whether deeper human investigation should begin.
Because the jobs differ: source-specific analysis, document extraction, structuring and evaluation. The models are complementary; they are not voting on the same answer.
They are confidential. Publicly, we can confirm only that a predefined list of questions produces yes, no or possible answers and a bounded initial classification.
No. They provide initial context. Startup-supplied documents and human review remain necessary, and some information can only be tested through direct conversations or deeper diligence.
It is an average reported by the project team across a startup sample for the full pre-screening path. The sample size, period and underlying records are not public, so it is not an independent benchmark.
Sources and limits
- The client workflow is described from IZZY’s own project knowledge, recorded in August 2026. The fund, startups, strategy, documents, model providers, exact criteria and outputs remain confidential.
- A Multi-Agent Orchestration Framework for Venture Capital Due Diligence and DIALECTIC: A Multi-Agent System for Startup Evaluation provide current research context for structured evaluation systems; neither independently validates this project.
- The UK Business Angels Association review of 13 August 2026, quoting investor Sarah Barber, supplies current context for the limits of deck-only analysis.
- No investment outcome, accuracy rate, risk reduction, fund return or comparative decision quality is claimed.
- The reported time change covers pre-screening through first classification, not full due diligence.