AI Agent System for Automated B2B Research and Sales Pipeline Generation

unify · Tech · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.2 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: B2B sales research

Unify wanted to automate unstructured go-to-market research across thousands of companies with repeatability, observability, and cost control.

Constraints:

• web search/scraping
• token volume
• customer-defined questions
• model cost

Architecture

Unify’s B2B research system turns go-to-market targeting into a large-scale agentic search problem. Customers define questions, output types, and free-form research guidance; agents then run across thousands of companies in the Plays engine, using internet search, website search, scraping, later deep internet research, browser access, HTML search, and dataset access. The architecture evolved from ReAct agents with GPT-4/o1 planning to cost-optimized planning with GPT-4.1 and richer tools that pull actual page content, decompose browser trajectories, and reduce misinterpretation.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  customer-research-specification[Customer research specification] --> plays-research-engine[Plays research engine]
  plays-research-engine --> research-planning-agent[Research planning agent]
  research-planning-agent --> core-web-tools[Core web tools]
  research-planning-agent --> enhanced-research-tools[Enhanced research tools]
  enhanced-research-tools --> browser-access-sub-agent[Browser access sub-agent]
  model-selection-layer[Model selection layer] --> research-planning-agent
  plays-research-engine --> evaluation-datasets[Evaluation datasets]

Components

Data flow

A customer supplies questions about a company/person plus guidance on how to research them. The Plays engine runs research agents over thousands or tens of thousands of companies. The agent plans, searches the internet and target sites, scrapes or browses when needed, ingests page content into context, answers typed outputs such as text/enum/boolean, and feeds the resulting facts into targeted sales outreach. Model and tool layers are iterated using hand-labeled evaluation sets and cost/performance measurements.

Why this architecture

Unify chose agents because go-to-market research is a broad search problem over semantically rich, unstructured web data. The architecture keeps evolving around model cost and tool quality: early search/scrape tools produced poor internet behavior, so the team added deeper search and browser tools that more closely mimic how humans open links, filter sources, and inspect page content.

Models used: GPT-4o · o1-preview · o1 · o3 · GPT-4.1 · o4-mini

Scale

• Agents run over thousands or tens of thousands of companies for each customer research program.
  “This agent then runs on thousands or tens of thousands of companies, um, answers those questions in an engine called plays”transcript, stated
• Unify processed 36 billion OpenAI tokens in April and token volume was increasing quickly.
  “In April, we pushed uh 36 billion tokens and that has increased pretty quickly every month since.”transcript, stated
• A planning model swap reduced an initial agent run from around 35 cents to around 10 cents with similar performance.
  “an initial agent run that used to cost around 35 cents now costs around 10 cents with 41 with similar performance.”transcript, stated
• The first evaluation hand-labeled about 500 examples across 100 companies for each of five core datasets.
  “I actually handlabeled probably like 500 examples across, you know, 100 companies for each of these five core data sets.”transcript, stated

Alternatives considered (roads not taken)

Manual human research for go-to-market targeting

Rejected because: Agents provide repeatability, observability, and scalability over huge unstructured research space.

Naive search-preview-based internet research

Rejected because: Agents picked answers from unreliable previews, so Unify pulled URL and page content in one tool call.

Scraping-only web data access

Rejected because: Interactive data sources, Google Maps, images, and query-driven sites required browser access.

Patterns

tool-use-orchestration (research)

Unify’s agents combine search and scraping tools to answer structured sales-research questions at scale.

cost-aware-routing (operations)

The system reduced per-query costs through model swaps and architecture changes under very high token volume.

Tradeoffs

How to replace human research labor.

Chose parallel agents over thousands of companies over manual per-company research.

Rationale: Agents make the process repeatable and observable, but require cost controls and quality evaluation.

Planning model economics.

Chose GPT-4.1 for production agentic planning over slower and more expensive o1/o3/o1-preview planning.

Rationale: The swap preserved similar performance while cutting per-run cost from about 35 cents to about 10 cents.

Research tool design.

Chose deep search and browser access that ingest content over search previews and basic scraping only.

Rationale: Richer tools reduced misinterpretation and unlocked use cases unavailable to naive search/scrape agents.

Gotchas (what bit them)

Scaled web research becomes a token and cost problem very quickly.

Root cause: Running agents across thousands of companies pushes enormous monthly token volume.

Resolution: Instrument cost and route/model choices deliberately instead of using the strongest model for every step.

Evaluation

The source discusses model evaluation and cost metrics rather than a public accuracy benchmark.

Metrics: 36B monthly tokens · $0.35 to $0.10 per query

Outcomes

• Unify operationalized agentic B2B research for sales-pipeline generation.

Vendor/product talk; metrics are self-reported.