Marketing and CommunicationsProspect IntelligenceGlobal

How a global marketing leader automated prospect intelligence

A global marketing and communications holding company built an AI platform that turns a new prospect into a scored, sourced recommendation, replacing repeated manual research and duplicate records.

Up to 95%Fewer duplicate
records
Up to 70%Less research time
100%Sourced prospect
summaries
A team member working through prospect data in a spreadsheet at a wooden desk
Highlights

The Highlights

  • About 20x less time spent chasing duplicate records
  • About 3x more prospects researched, in the same time
  • 100% of qualified prospects now get a sourced, structured summary
01Client

A global leader in marketing and communications.

A global marketing and communications holding company, spanning a network of creative, media and technology agencies worldwide.

Region
Global
Sector
Marketing and communications holding company
Annual revenue
> $10B/year
Employees
>50,000+
Function in scope
Prospect intake and qualification, Phase 1
02The problem

Prospect research bled time, was inconsistent and decentralized.

Prospects were duplicated, researched manually, and evaluated inconsistently, with no central way to bring it all together.

  1. Duplicate records, piling upThe same prospect company could exist multiple times in the system, with no reliable way to catch it before research began.
  2. Manual research, every prospectEvery qualified prospect required a person to research the company manually, a real time cost, every single time.
  3. Inconsistent summariesCompany research wasn’t captured the same way twice, making it hard to compare prospects side by side.
  4. No central way to qualifyWithout one place to apply consistent criteria, evaluation and decision-making slowed down.
03Native limits

More than just catching duplicates.

A basic deduplication script wasn’t going to fix this. Four things had to come together.

  1. Catch duplicates two ways, not oneMatch on exact identifiers such as ID, website and domain, and on fuzzy similarity, so near-duplicates don’t slip past either method alone.
  2. Research every prospect the same wayPull from the same public sources, such as the official site, news, press releases and leadership pages, for every company, not whatever a person happened to check.
  3. Apply qualification rules without hardcoding themLet qualification criteria change without needing new code, since the exact criteria hadn’t been finalized yet.
  4. Keep evaluation entirely internalRead prospect data from the system of record, but never write back, reach out, or take action without a person deciding first.
04What Lyzr built

Four agents, from duplicate check to a scored recommendation.

Lyzr built an agent pipeline that reads pending prospect accounts, finds likely duplicates, researches each company from the web, and returns a scored recommendation for review.

Core Capabilities

  • Prospect intakeSecure, read-only access to pending prospect accounts
  • Duplicate detectionCombines exact-match identifiers with AI-driven fuzzy matching on company name and domain
  • Parallel web researchA search agent and a site-crawler agent research each company at the same time, each citing its sources
  • Scored recommendationA fourth agent reconciles both research threads into one summary, a score from 0 to 100, and a recommended action
05Architecture

From a pending account to a scored recommendation.

A request moves through duplicate detection, parallel research and reconciliation before a reviewer ever sees it.

RESOURCE GROUP AND PRIVATE SUBNET Access MDM backend Lyzr Studio agents Lyzr services User VPN protected MDM dashboard SSL certificate Key Vault Salesforce client ID and secret MDM backend Orchestrates each account Cosmos DB Encrypted persistent storage Account dedup Search agent Scrape agent Consolidator AZURE-DEPLOYED OPENAI Lyzr Search agent SearXNG Lyzr Crawl Crawls HTTPS sites EXTERNAL Public web Salesforce Read-only query, search and suggestions APIs 1 2 8 4 3 5 6 7
1VPN-protected user opens the MDM dashboard
2Dashboard fetches accounts up for approval, duplicates and research
3Backend reads accounts from Salesforce, read-only, with secrets in Key Vault
4Backend triggers the account research workflow
5Search agent sends account name, billing country and website to web search
6Scrape agent sends the account website to Lyzr Crawl
7Search and crawl reach the public web
8Results are stored in Cosmos DB, encrypted
System architecture. From the MDM dashboard to a scored recommendation, inside a private subnet.Fig. 01
06Controls and governance

Every recommendation, checked before a person sees it.

The system researches and scores. A person always makes the final call.

Read-only access
The system only reads prospect data, it never writes back or modifies a record automatically.
Two-layer duplicate check
Exact-match identifiers and AI-driven fuzzy matching both have to agree before a duplicate is confirmed.
Every field, sourced
Every researched value carries the source it came from, not just a final answer.
Full audit trail
Every step of every run, started, completed, or failed, is logged for review.
Human decision, always
The system returns a score and a recommendation. A person makes the final call.
07Re-imagined workflow

From a pending prospect to a scored recommendation, automatically.

Before, a sales team manually researched every qualified prospect, one company at a time. After, pending accounts are checked for duplicates and researched automatically, with a scored recommendation ready for review.

The prospect research workflow, before and after Lyzr.Fig. 02
08Results

What changes once duplicate research stops happening.

Same evaluation the sales team always needed, and the target is to get there without the repeated manual work.

MeasureBeforeAfter target
Duplicate recordsNo systematic way to catch themChecked automatically, on exact and fuzzy matches
ResearchDone manually, per prospectTwo agents research every account in parallel
SummariesInconsistent, one analyst to the nextThe same structure, every account
SourcingOften untrackedEvery field traces back to where it came from
Decision-makingDelayed, without a central viewA scored recommendation, ready for review
Over 95%

reduction in duplicate processing

Over 70%

reduction in manual research time

100%

structured summary for every qualified prospect

09What is next

Built to extend, not rebuild.

The same pipeline is designed to take on more, configurable qualification rules, broader automation, and scheduled runs, all as extensions of what’s already in place, not a rebuild.

One pipeline, built to extend

Got a use case in mind

Platform, engineers and governance all in. We will map your workflow against the same four tests this one had to pass.

Talk to the team