How to Build a Private-Company Research Agent When Financial Data Barely Exists

We spoke to a growth-equity fund who ran a structured accuracy test on 58 companies across their primary data vendor. The result was 69% accuracy on foundational fields like headquarters location, employee count, and total funding raised, meaning nearly a third of the basic facts were wrong.

First, this shows the poor quality of data vendors out there, even in 2026, which is concerning.

Second, this means that when you build a research agent on top of that foundation, scoring layers, screening thresholds, and routing decisions all inherit the error. Standard approaches to building company research agents skip this problem entirely because they use public companies with clean SEC filings as their examples. Private company research is a different problem. The data inputs are incomplete by default, and the agent has to be designed around that absence rather than assuming it away.

This article covers how to build a private-company research agent that treats missing financial data as a design constraint from the start. The architecture uses hiring patterns, web signals, and structured API data as primary inputs, with confidence bands and human review built into the scoring layer.

Why private-company research breaks when you assume financial data will be complete and up to date

The default assumption in most company research workflows is that you can query a database, get a funding total, an employee count, a revenue estimate, and a headquarters location, and then build scoring and filtering logic on top of those fields. For public companies, that assumption holds. For private companies, it fails in ways that are quiet and compounding.

What the major databases actually contain for private companies

The coverage problem is worse than most teams realize before they test it. Crunchbase and PitchBook records for private companies frequently have missing investors, outdated funding totals, and deals where key participants do not appear in the database at all. Funding figures that do appear are often out of date or sourced from a single self-reported disclosure.

"Coverage" in practice means a record exists. It does not mean the fields that matter for investment decisions, such as accurate funding totals, up-to-date headcount, or estimated revenue, are populated or correct.

How one bad input cascades through an automated system

A growth-equity fund discovered during their accuracy test that a specific funding figure, $625.8 million, appeared repeatedly across multiple company records. The number was a data anomaly that had no connection to actual fundraising. When a scoring layer trusts that number, companies get disqualified or prioritized based on data that has no relationship to reality.

As one team put it after running their own validation: "If we get the basic things wrong and disqualify a company based on information that's incorrect, that is a problem." The consequence is lost deal flow from companies that should have made it through the screen, and wasted analyst time on companies that should not have.

This is a design constraint that should shape the agent architecture. The architecture of a private-company research agent has to treat financial data as one signal among many, weighted by how verifiable it is, rather than as the foundation the rest of the system depends on.

What to use instead: hiring patterns, web signals, people movement, news, and public-web traces

When financial data is absent or unreliable, the agent needs a different signal taxonomy. The goal is to build a composite picture of a company's trajectory using signals that are observable, structured, and independently verifiable.

Headcount growth

Headcount growth over 6-month and 12-month windows is the most widely available proxy for company trajectory. A company that grew from 50 to 120 employees in a year is telling you something about its revenue trajectory and capital position even if neither number is public.

Department-level headcount adds a layer that total headcount misses. A company that doubled its engineering team while sales stayed flat is in a different position than one that tripled sales headcount while engineering shrank. The Company Enrichment API returns historical headcount series, growth rates, and function-level breakdowns by department, making it possible to track not just whether a company is growing but where it is investing that growth.

Job postings by function

Job posting volume and function distribution reveal lifecycle stage in a way that raw headcount cannot. A company hiring three machine-learning engineers, two sales reps, and a VP of Marketing is in a different phase than one hiring exclusively for customer support. The Job Listing API provides historical and live job postings with role classifications, making it possible to track these patterns programmatically.

Company news

One team that built an internal aggregation tool across four structured data providers identified company news as their single biggest remaining data gap once headcount and jobs were covered. As that team's lead described it: "The most specific gap we're experiencing today is frankly what I'll call broadly company news. A company comes out with a press release, companies in an article, companies mentioned somewhere."

The reason news matters more than most teams initially expect is that it creates timed windows for action. A VP of Sales hire, a new product launch, or a press mention is both a discovery signal (this company is doing something worth investigating) and an outreach trigger (there is now a specific reason to contact the CEO). One growth-equity fund described using news of a product release as the basis for a CEO touchpoint: mentioning the new product, asking about it, and using it as a way to start a conversation rather than sending a cold outreach with no context.

News is also the hardest signal to get from structured APIs. Aggregators lag behind the actual event by days or weeks, and many press releases, partnership announcements, and regulatory filings never make it into any structured database at all. This is the signal type where web search has to supplement API data most heavily.

Leadership and people changes

C-suite and leadership changes serve two distinct purposes depending on where a team sits in the investment process.

For due diligence, one PE firm's AI/ML lead maps C-suite history at 10 or more competitor companies over 15-year windows when evaluating an investment, then reaches out to former executives who have since left those companies to serve as industry advisors. The signal is not just that a change happened, but who went where and what sector expertise they carried with them.

For deal sourcing, leadership changes are often the earliest detectable sign that someone is building a new company. One early-stage VC described how this plays out: "The last three founders that we've invested in, literally one of them wrapped up at Citibank. The only indication that they were building something was that they changed their LinkedIn bio." Another VC tracks vesting period timing at target companies as a predictor of when senior operators might be ready to leave and start something new. The teams that detect these changes in real time get priority access to founders before competing investors find them.

People changes are available through structured APIs (stealth founder tracking is one implementation pattern), though trade press coverage of senior hires often adds context that profile data alone does not capture.

Talent inflow and outflow

Tracking where a company's employees come from and where they leave to reveals patterns that no single data point can. One PE firm uses this for investment due diligence: "If we are looking to invest in company X, we want to understand what's their hiring pipeline from, let's say, the big four accountant firms. And where are they hiring from? Where are people leaving to? Understanding profiles of hires, understanding trending in hires."

The same team uses inflow/outflow analysis to detect competitor product development. When a portfolio company's competitor starts hiring machine-learning engineers from well-known research labs, that is an early signal of a new product initiative before any public announcement. As that team described it, they watch for "new product development happening at some alternative investments," meaning they track hiring patterns at competitor companies to understand what those companies are building next.

Talent flow data is queryable through people search APIs with filters for current and past employers, seniority levels, and job titles, making it possible to map flows across an entire sector programmatically rather than tracking individual departures manually.

Website activity and web traffic

Web traffic estimates and website change patterns function as health signals for companies already in a portfolio or under active evaluation. Monthly visitor estimates and quarter-over-quarter trends will not tell you a company's revenue, but they will tell you whether the company's public presence is growing, flat, or declining, which is useful for triangulation against other signals.

One fund-of-funds manager takes this further by tracking content activity alongside traffic: "If you invest in somebody and they're not updating anything on their marketing site for days, weeks, months, that gives you a different signal than if they are constantly blogging." That team runs weekly analysis of marketing site changes across their portfolio companies, tracking what content gets published, how the company's messaging evolves, and whether there are long periods of inactivity. A company with flat web traffic and no new content for three months is not necessarily failing, but it warrants a conversation that a company publishing weekly product updates does not.

Social posts

Company and founder social activity captures announcements that often precede structured data by weeks. A founder posting about opening a second office, a company page announcing a new product line, or a CEO sharing a hiring milestone are all public signals that do not require web scraping to access because they are available through structured social post retrieval APIs in real time.

Social posts also fill a gap that company news leaves open. Not every company gets press coverage, and smaller private companies rarely appear in trade publications at all. But founders and company pages post regularly about product updates, customer wins, conference appearances, and team growth. For companies where news aggregators return nothing, social activity is often the only public evidence that the company is active and building.

Signal comparison

How to turn weak signals into a scorecard with confidence bands and human review

The typical company research agent stops at data collection, querying the API, getting the result, and presenting it to the user. For private companies, the harder problem is what to do when three data sources give you three different answers for the same field, or when half the fields you need are empty. The scoring layer has to represent what you actually know, and how confident you are in it, rather than producing a single number that hides the uncertainty.

Designing scores that represent what you actually know

Each signal in the agent's output should carry three properties, including the value, the confidence level, and the source. A company with four corroborating signals at moderate confidence, such as growing headcount, increasing job postings, recent leadership hires, and rising web traffic, is more actionable than a company with a single high-confidence data point from a source you cannot independently verify.

One insight that emerged from teams building multi-source research systems is that cross-referencing across sources reveals discrepancies that are often more valuable than confirmations. When two providers disagree on a company's employee count by 40%, that disagreement is itself a signal worth surfacing.

A simplified scoring record for a single company might look like this:

{
  "company": "Acme Corp",
  "signals": [
    {
      "type": "headcount_growth_12m",
      "value": 0.42,
      "confidence": "high",
      "source": "company_enrichment_api",
      "last_updated": "2026-05-10"
    },
    {
      "type": "funding_total_usd",
      "value": 15000000,
      "confidence": "low",
      "source": "crunchbase_aggregated",
      "last_updated": "2025-09-01",
      "flag": "last_updated > 6 months ago"
    },
    {
      "type": "job_postings_current",
      "value": 23,
      "confidence": "high",
      "source": "job_listing_api",
      "last_updated": "2026-05-14"
    },
    {
      "type": "revenue_estimate",
      "value": null,
      "confidence": "none",
      "source": null,
      "flag": "no revenue data available"
    }
  ],
  "composite_score": 72,
  "confidence_band": "moderate",
  "flags": [
    "funding_total outdated (9 months)",
    "no revenue data available",
    "headcount and job posting signals corroborate growth thesis"
  ]
}

{
  "company": "Acme Corp",
  "signals": [
    {
      "type": "headcount_growth_12m",
      "value": 0.42,
      "confidence": "high",
      "source": "company_enrichment_api",
      "last_updated": "2026-05-10"
    },
    {
      "type": "funding_total_usd",
      "value": 15000000,
      "confidence": "low",
      "source": "crunchbase_aggregated",
      "last_updated": "2025-09-01",
      "flag": "last_updated > 6 months ago"
    },
    {
      "type": "job_postings_current",
      "value": 23,
      "confidence": "high",
      "source": "job_listing_api",
      "last_updated": "2026-05-14"
    },
    {
      "type": "revenue_estimate",
      "value": null,
      "confidence": "none",
      "source": null,
      "flag": "no revenue data available"
    }
  ],
  "composite_score": 72,
  "confidence_band": "moderate",
  "flags": [
    "funding_total outdated (9 months)",
    "no revenue data available",
    "headcount and job posting signals corroborate growth thesis"
  ]
}

{
  "company": "Acme Corp",
  "signals": [
    {
      "type": "headcount_growth_12m",
      "value": 0.42,
      "confidence": "high",
      "source": "company_enrichment_api",
      "last_updated": "2026-05-10"
    },
    {
      "type": "funding_total_usd",
      "value": 15000000,
      "confidence": "low",
      "source": "crunchbase_aggregated",
      "last_updated": "2025-09-01",
      "flag": "last_updated > 6 months ago"
    },
    {
      "type": "job_postings_current",
      "value": 23,
      "confidence": "high",
      "source": "job_listing_api",
      "last_updated": "2026-05-14"
    },
    {
      "type": "revenue_estimate",
      "value": null,
      "confidence": "none",
      "source": null,
      "flag": "no revenue data available"
    }
  ],
  "composite_score": 72,
  "confidence_band": "moderate",
  "flags": [
    "funding_total outdated (9 months)",
    "no revenue data available",
    "headcount and job posting signals corroborate growth thesis"
  ]
}

The confidence_band on the composite score tells downstream consumers whether to act on this record directly or route it for additional verification.

Red flags the agent should surface explicitly

The most useful output from a private-company research agent is the red flags. Contradictory data across sources, such as a funding amount that differs by 3x between two providers, should be surfaced explicitly rather than averaged into a single number. Missing data that should exist, such as a Series B company with zero job postings, is itself a flag worth investigating.

A founder building a private-company research agent described the architecture they are constructing. The agent takes a company name as input, queries structured APIs for baseline data, supplements with web scraping and news retrieval, and produces a report that includes both scores and red flags. The red flags are a core part of the output.

Outdated data is a subtler problem. A record with a "last updated" timestamp from nine months ago may still appear in query results alongside fresh records. The agent should flag the gap between the timestamp and the current date, rather than treating old data as equivalent to new data.

Where human review enters the loop

The agent generates the report, and a human decides whether to act on it.

Confidence bands determine routing. High-confidence records with corroborating signals and no red flags go directly to the deal team's queue. Low-confidence records, where critical fields are missing or contradictory, go to an analyst for manual verification before the deal team sees them.

As one investment team described it: "Trust is everything." When the system produces a recommendation that turns out to be based on incorrect foundational data, it costs more than one deal. It undermines the team's willingness to use the system at all. Building human review into the routing, rather than treating it as a fallback for when the agent fails, is what makes the system trustworthy enough to use daily.

The Watcher API adds a time dimension to this loop. Once a company has been scored and reviewed, a watcher can monitor for changes in headcount, leadership, job postings, or funding events and re-trigger the scoring pipeline when something material changes. The analyst does not have to manually re-check every company on a quarterly cadence.

Where structured APIs end and agentic web research has to take over

A private-company research agent cannot run entirely on structured API data, and it should not run entirely on web scraping. The practical architecture is a two-layer system where structured APIs handle the baseline and targeted web research fills specific gaps the first layer identifies.

What structured APIs cover well

Headcount, job postings, funding history (when available), firmographics, people movement, and social posts are all queryable through structured endpoints that return JSON, support filtering and batching, and carry timestamps. The Company Enrichment API returns 250+ datapoints per company, including historical headcount series, growth rates, leadership profiles, web traffic estimates, and funding history. The People Discovery API covers 60+ filters for finding people by title, seniority, company, geography, and career history.

The advantage of starting with structured data is that every company record comes back with the same structured format, timestamped fields, and confidence properties. You can filter 10,000 companies in a single query and get structured results back in seconds. That is the foundation layer of the agent, and it is where you should spend the least analyst time per company.

Full API documentation covers endpoint-level detail for each of these surfaces.

What requires web search and page-level extraction

Company news is the most common gap. Product launches, partnership announcements, regulatory filings, and leadership hires reported in trade press are often the highest-signal events for private companies, and no structured API indexes them in real time. The Web Search API provides structured search results across web, news, and social sources, which gives the agent a starting point for targeted extraction.

Patent filings, SEC exemption filings (like Form D for private fundraising), and state-level business registrations are publicly available but not aggregated into any single API. For companies operating in regulated industries, these filings often contain information about corporate structure, principal officers, and capital raises that does not appear anywhere else.

The practical implementation is a two-pass system. The first pass uses structured API queries to build a baseline profile with confidence scores for each field. The second pass looks at the gaps and red flags from the first pass and runs targeted web research for the specific missing data points. The second pass is expensive in both time and compute, so the first pass has to be good enough to tell the agent where to look and what to verify.

Why this belongs in an internal workflow, not a one-off prompt

A one-off prompt to a general-purpose AI model can produce a passable company research summary. It can pull together publicly available facts, synthesize them into a narrative, and answer direct questions. What it cannot do is maintain accuracy over time, route records by confidence level, track changes, integrate with a deal pipeline, or tell you which of its claims it is least sure about. This is the same reason investment teams increasingly build proprietary sourcing pipelines rather than relying on shared platforms.

Why repeatable workflows beat one-off research

One-off prompts have no memory, so every research session starts from zero. An analyst who researched a company three months ago has no structured record of what was found, what was flagged, and what has changed since. The research is a point-in-time assessment that begins going out of date the moment it finishes, with no mechanism to detect when it becomes materially outdated.

One-off prompts also have no audit trail. When a deal committee asks why a company was scored a certain way, or why it was excluded from a pipeline, the answer has to be traceable to specific data points, sources, and confidence assessments. A chat transcript does not serve as an audit trail.

The compounding value of a workflow over a one-off prompt comes from every subsequent run that does not require an analyst to start from scratch.

What the internal workflow looks like

The full pipeline, assembled from the components in the previous sections, follows this sequence: a company name or domain enters the system, the agent queries structured APIs for enrichment data, identifies gaps and runs targeted web research for missing signals, scores each field with confidence bands, routes the result based on confidence (high-confidence records go to the deal team, low-confidence to an analyst for review), and sets up real-time monitoring via webhooks for ongoing changes.

This is the same architecture described across the previous four sections, connected end to end. The difference between a research agent and a research workflow is that the workflow persists, tracking what has been researched, what has changed, and what needs human attention.

As one team described the shift after connecting their multi-source research tool to an AI orchestration layer: "the ceiling on this tool has just exploded." What was previously a shortcut to avoid checking multiple databases manually became a system that could evaluate companies at a scale no analyst team could match, while still preserving the human judgment layer where it matters most.

Teams building this architecture can review how a top-5 VC built an internal deal-sourcing tool with Crustdata and how AI investment platforms use real-time company data for deal sourcing for reference on what the production version of this workflow looks like.

Conclusion

The financial data gap for private companies is not a temporary problem that better vendors will eventually solve. Private companies will always have less structured financial data than public ones. The teams that design their research systems around this constraint, building agents that use multiple signal types, score with explicit confidence bands, surface red flags, route by certainty, and keep humans in the review loop, will have a compounding advantage in deal sourcing and diligence over teams waiting for a single database to get the data right.

The architecture described here is already being built by investment teams and founders who have tested the limits of existing data providers and decided to build past them. The structured data layer, from company enrichment to job listings to real-time monitoring, provides the foundation. The scoring, routing, and human review layers are what turn that foundation into a system worth trusting.

Book a demo to see how the data layer works, or explore the API documentation to start building.