Ansa Scores 16,000 Companies a Month on Software It Built Instead of Bought
AI workflow automation at a lean venture firm: Ansa built its own deal-sourcing platform, scoring 16,000 companies a month and saving 100+ team hours, with 75% of the firm using it weekly.
16,000
Companies scored every month by the firm's own platform
100+
Hours of manual work saved each month
75%
Of the firm uses the platform every week
- Founded
- New York, 2021
- Website
- ansa.co
75% of our team uses the app on a weekly basis and we estimate we're saving over ~100 team hours per month.
A lean team competing with larger rivals faces a simple arithmetic problem: the other side has more people. The answer is leverage. Ansa, a venture capital firm in New York, competes for investments against funds many times its size, and it closes that gap with software the firm built and owns.
Ansa invests in companies from Series A to C and has backed Defense Unicorns, Bland, Gradient, and Selector; its partners previously supported some of the industry's largest outcomes, including Crowdstrike, Coinbase, and SurveyMonkey. Ryan Sullivan, an investor who oversees engineering and data science, is the architect of the firm's data-driven sourcing strategy, working with Managing Partner Marco Demeireles.
Tens of thousands of companies, a handful of investors
Ansa's investable universe runs to tens of thousands of companies, and its sourcing is thesis driven: the team needs to find every company in a theme of interest, track opportunities across the whole market, and reach the right ones at the right time. A large firm solves that with headcount. Ansa chose to solve it with software the firm owns.
We have a lean investment team and there's a lot of opportunities out there, so we're trying to automate as much as we can on the workflow side to help our team be efficient and research, review, and reach out to as many companies as possible.
There was no product to buy for this. Deal-sourcing tools exist, but Ansa's edge is proprietary: its own scoring model, its own data sources, its own view of what makes an investment. Software that embodies a firm's judgment cannot come off a shelf.
The platform, in three workflows
The team built a single application that now carries eight core workflows. Three of them show what the platform does.
Ask for companies in plain language
An investor can type "carbon accounting software companies with a CEO in NYC that score over 60" and receive a curated list. Behind the question, the platform has already read and organized what every company says about itself, so the whole market is searchable in plain language.
Score 16,000 companies a month, automatically
Private companies publish little, so Ansa relies on alternative data and a scoring model trained on a dataset the firm has curated for years. The platform automatically scores companies at a rate of 16,000 a month, and a second workflow scores any ad-hoc list an investor uploads, in any format, returning the results in the app and by email.
Reach out in two clicks instead of thirty
When the team has a shortlist, a third workflow prepares the outreach: checking ownership records, filling in missing details, finding the right contact, and assembling the context for a personal note.
Before we would do this all manually, now with this new workflow in Reflex, we've taken what was once 30+ clicks across 5 different apps and made it 5x faster with 2 clicks across 2 apps.
What changed for the team, side by side:
| Before the platform | With the platform Ansa built | |
|---|---|---|
| Finding companies in a theme | Manual research across databases | One plain-language question |
| Scoring a list | Analyst hours per batch | 16,000 companies scored automatically each month |
| Outreach preparation | 30+ clicks across 5 apps | 2 clicks across 2 apps |
| Time cost | 3 to 5 minutes per company | 100+ team hours saved per month |
For how to run the build-or-buy decision on your own tools, see our build vs. buy framework for internal software.
Why Ansa built for the long term
The firm had tried the shorter path first: a low-code Python framework with a drag-and-drop editor. It worked until it did not. The components looked dated, performance disappointed, and the ceiling was structural: a firm planning to build its edge into software cannot risk outgrowing the tool that holds it.
We don't want to run into a situation where this year or next year, we want to add more functionality that this low code framework doesn't have and we're not able to integrate it. Additionally, the rate of improvement and development velocity from the Reflex team gave us confidence that their offering would continue to improve over time. We're building this for the long term and we want to make sure we both have the flexibility to not outgrow it and are working with the best out there.
The concern was specific: the team knew AI would keep moving, and the platform had to move with it.
I started to feel that with this framework I didn't know if they could keep up with the pace of new developments with LLMs. They abstract a lot of the backend, so it's difficult to install third party libraries and you don't have full control over the database.
That is the pattern behind most low-code limitations: the tool is fastest on day one and most expensive the day your needs pass its edges. Ansa moved to Reflex so the platform could take on every new AI capability without a rebuild, and so the firm, not a framework vendor, decides what it can do next.
How it was built
This section is written for your technical team. In short: the platform is plain Python, so every new AI capability drops straight in.
View the technical details
The application is a single Reflex app carrying eight workflow automations. The natural-language search combines OpenAI, LangChain, and Chroma for embeddings-based retrieval over scraped and normalized company data; scoring runs through the firm's custom ML model hosted in Databricks, reached over an API; outreach preparation runs as Reflex background events with LLM steps for fuzzy matching and summarization.
The LLMs help us normalize across different companies, even if pages are named differently, so we can easily search through all of them and figure out what the company does.
Supporting pieces: a Supabase database, the Google Auth login component for employee sign-in, AG Grid tables for data display, and CSV export. The team is mostly data engineers and analysts working in Python; no JavaScript was written for the platform.
We don't have a full engineering team, so to build a full web app from scratch seemed like a lot to manage. In addition, our team is mostly data engineers / analysts so we are far more comfortable with Python than JavaScript.
Because the platform is plain Python, new tools slot in without waiting on a vendor.
Given that it's just python code, I'm always comfortable that we'll be able use different tools and to figure out how to make it work with Reflex versus being stuck with the integrations that our old solution had.
From tool to dependency
The platform is now a central dependency of how Ansa sources and analyzes investments: 75% of the firm uses it every week, it scores 16,000 companies a month, its workflows have run more than 3,000 times, and it has returned more than 100 team hours monthly to a team whose scarcest resource is attention. For a lean firm competing with larger funds, that is the point of owning the software: the edge compounds, and nobody else can buy it.
For a similar decision at a fintech startup, see how Bayesline shipped its analytics product without hiring a frontend engineer.
Frequently asked questions about AI deal sourcing
Can a small investment team automate deal sourcing with AI?
Yes. Ansa, a lean venture firm in New York, built its own AI-powered sourcing platform: it scores 16,000 companies a month, answers plain-language questions about the market, and prepares outreach automatically. 75% of the firm uses it weekly, and the team estimates it saves more than 100 hours of manual work per month.
Why do teams outgrow low-code and no-code tools?
Low-code tools are fastest on day one and most limiting once your needs pass their edges. Ansa's earlier low-code platform made it hard to install third-party libraries, offered no full control over the database, and could not keep pace with new AI capabilities such as LLMs and vector search. The firm rebuilt on Reflex in plain Python so nothing new would be out of reach.
Is there off-the-shelf software for venture capital deal sourcing?
Sourcing databases exist, but a firm's real edge is proprietary: its own scoring model, its own data, and its own view of what makes an investment. Software that embodies that judgment has to be built. Ansa built its platform on Reflex and owns the model, the workflows, and the compounding advantage.
How much time does workflow automation actually save?
At Ansa, outreach preparation went from more than 30 clicks across 5 applications to 2 clicks across 2, and scoring that once cost 3 to 5 analyst minutes per company now runs automatically across 16,000 companies a month. In total the firm estimates over 100 team hours saved monthly.
See what other teams built
See what your team could build.
Tell us about your workflow. See how Reflex can help you turn it into an application.