Private company research is still mostly manual. An analyst opens a dozen tabs, reads a company's site, skims the news, checks a funding announcement against a secondhand summary, and pastes fragments into a document. The document is out of date soon after it is saved. Research tools have improved, but a person still assembles the notes by hand.
Agent-Powered Research: How We Built a Research System Instead of a Database
How aVenture's AI agents research private companies and how we check what they write.
When we started building aVenture, we chose to have AI agents do the research. Agents read the web and add structured facts to a company's profile, and rules check every addition before it is saved.
Research runs
Many private-market databases are filled in bulk, from licensed data feeds, submitted forms, and periodic manual updates.
aVenture works in research runs. An agent gets a question about a company and researches the answer. It starts with what the company sells, adding each product or service to the profile on its own, apart from the industry the company is listed under. For each product or service, it finds competing products from other companies, and the market maps on aVenture are built from those matches.
The agent also maps the company's website, page by page. How a company describes itself, and how it organizes its site, shows how it positions itself. News and blog feeds are watched as well, and each article is matched to the companies and people it mentions, so a profile collects its press coverage over time.
Each run adds structured facts to a knowledge graph of companies, people, investors, funds, accelerators, funding rounds, products, and the links between them. Each fact has its own place.

Delegate the first pass on enterprise AI research and start your review with an editable draft.

Check the web results behind the enterprise AI research before deciding which findings to use.
Rules for agent writes
AI systems can be confidently wrong. The checks took us longer to build than the agents themselves.
Writes are checked before they are saved
Each kind of fact has its own rules, checked the moment the agent tries to save it. The rules cover which values are acceptable, which numbers make sense, and how the text must be shaped. If a result breaks a rule, it is rejected and the profile is unchanged.
Writes record the agent and model
Every fact notes who wrote it. When an agent wrote it, the note also names the model and what the agent read. If a fact is ever questioned, we can look up which agent and model produced it, what they read, and when.
Sweep jobs find duplicates
Two runs can find the same product under slightly different names, or the same article from two feeds. Automated sweeps run around the clock to find these duplicates and clean them up.
Pipelines have concurrency limits
Research runs, website mapping, and news gathering all happen in the background, with a cap on how many run at once. The servers take on as much of this work at once as they can handle and queue the rest, so researching many companies at the same time does not bring the site down.
People and agents use the same interfaces
Agents have no private way in, and people get no looser treatment. Human researchers and AI agents add facts through the same tools, with the same checks and the same note of who wrote what.
The same goes for outside tools. Besides the website, aVenture offers a command-line tool for repeatable research tasks and an MCP server. The MCP server lets AI agents search, read, enrich, and organize aVenture's company profiles. If you build with agents, your agents can work with the same profiles our agents keep up to date.
We expect research teams to mix people and agents on the same profiles. Separate paths for each would let the two drift apart.
Detail first, then more companies
From the start, our goal has been to research each company in detail first and to add more companies second.
A detailed profile covers a company's products, competitors, website, press coverage, people, and funding history. Building one takes more work than listing basic facts across a large directory of companies. We built the checks and cleanup sweeps for detailed profiles first, so covering more companies means running the same research more often.
A new competitor match adds a fact to two companies at once. A newly matched article adds to the profile of every company and person it mentions. A newly mapped website adds to what we know about how companies in a market describe themselves. That is the design goal, though we have not measured it.
We do not say that our agents match human analysts. Their work can be checked afterward.
What's live and what's next
All of this is live today. Profile pages with a full map of a company's website and a collected list of its articles are still to come.
Create a free account to research private companies on aVenture, or to connect your own agents to it. The aVenture apps are linked from the aVenture app download page. The Free plan includes funding history and news alerts and lets you view 15 company and person profiles a month; compare plans on the pricing page.
Similar content

Eight Recipes for aVenture Data: From a Website to a Brand Kit, a Target List, or a Portfolio
How to pull brand details, funding history, and investor portfolios out of aVenture from your own software.

aVenture in Claude, ChatGPT, and Your Terminal: The MCP Server and CLI
Research private company data from Claude, ChatGPT, or your terminal with aVenture's MCP server and CLI.

Identify a List of Names in One Job, and Stop Waiting on Slow Requests
Give aVenture a list of names and get back the company or person each one refers to, all in one background job.
