Home/Case studies/Company intelligence

Case study 03 / Client research platform

5,000+ briefs yearly.
65,000 hours saved.

A custom multi-agent system decomposes company research, verifies every claim against a source, reconciles the findings, and produces a client-ready document in minutes.

Custom orchestrationEvery claim groundedStructured client output
01

5,000+

Company researches completed per year

02

65,000+

Human hours saved at client volume

03

$600K+

Equivalent salary capacity avoided

04

7h → <10m

End-to-end research turnaround

The research bottleneck

Cut seven hours
from every brief.

Client teams needed deep company research before they could advise, sell, or prepare an engagement. A single useful brief required hours of navigating websites, checking leadership, products, markets, hiring, customers, and competitive signals.

At more than 5,000 researches per year, the volume implied capacity equivalent to roughly fourteen full-time research assistants. Scaling the manual process would have added more handoffs, more inconsistency, and more opportunity for unsupported claims.

The output also had to be usable. A faster wall of generated prose was not the goal. The system needed facts tied to sources, consistent sections, explicit gaps, and formatting that could move directly to a client.

The system

Run nine agents.
Produce one brief.

Each agent owns a narrow research question. A paired verification step checks its evidence. Only then does a normalization layer merge, cite, structure, and format the output.

Research run / live architecture

One request, nine parallel investigations

Output in <5 min
REQUESTResearch Acmecompany + market briefPARALLEL SPECIALIST AGENTSCompanyRESEARCHVERIFY SOURCESMarketRESEARCHVERIFY SOURCESProductRESEARCHVERIFY SOURCESLeadershipRESEARCHVERIFY SOURCESHiringRESEARCHVERIFY SOURCESCustomersRESEARCHVERIFY SOURCESCompetitorsRESEARCHVERIFY SOURCESFinancialRESEARCHVERIFY SOURCESSignalsRESEARCHVERIFY SOURCESNORMALIZATION LAYERMerge duplicatesResolve conflictsCheck citationsApply structureFormat outputEXECUTIVE BRIEFCLIENT READYEVERY CLAIM MUST RESOLVE TO A SOURCENOT FOUND IS BETTER THAN MADE UP

The grounding contract

Source every claim.
Reject unsupported facts.

The system did not allow plausible language to stand in for evidence. Every material claim had to resolve to something the company actually published. If the source was missing, the output said so.

01ResearchA specialist agent searches only for its assigned slice of the company.
02VerifyA second step checks the evidence, rejects unsupported claims, and preserves source URLs.
03NormalizeConflicts and duplicates are resolved across all nine research streams.
04FormatThe final layer applies the client schema, citations, headings, and clean document structure.

The time equation

Turn seven hours
into ten minutes.

01 / Before

~7 hours

Manual browsing, synthesis, fact checking, and formatting for every company.

02 / System

~5 min

Nine agents research in parallel, verify evidence, normalize findings, and format the document.

03 / Human review

~5 min

Inspect sources, resolve flagged gaps, and approve the client-ready brief.

5,000+

researches completed every year

65,000+

human hours saved at client volume

$600K+

in equivalent salary capacity avoided

Beyond research

Monitor at scale.
Escalate only ambiguity.

The broader product also monitored hiring signals across hundreds of thousands of companies by connecting directly to underlying job-portal APIs. Clear qualification cases were handled by semantic logic. Only ambiguous, low-confidence records escalated to AI. The same principle governed the entire product: deterministic systems for volume, agents where judgment earned their cost.

Large-scale collection

Daily monitoring across hundreds of thousands of companies.

Confidence-gated AI

Simple cases stay cheap and fast. Ambiguity gets intelligence.

Product ownership

Customer discovery, UX, roadmap, Python, AI, sales, and support.

The outcome

Seven hours became
ten minutes.

More throughput, consistent structure, every claim grounded, and human attention reserved for judgment rather than collection.

Map a research workflow