Undo Labs

Systems

Prototype

Facility-Contact Discovery

Designed to take a raw target list and find the person responsible for boilers or facilities at each site, using automated research first and an AI phone call only when research can't resolve an account. Today the import and cleanup stage is built and tested.

The problem it solves

Sales teams buy or inherit lists of facilities with a main phone number and nothing else. Reps spend hours calling receptionists to find out who owns the boiler room.

Clean a sample prospect list

How it fits together

  1. BuiltImportMessy CSV/Excel to clean, deduplicated accounts
  2. DesignedResearchWeb search, site crawl, AI extraction
  3. DesignedEnrichPeople search by domain and title, email check
  4. DesignedScoreResolved, or needs a call
  5. DesignedCallAI asks the front desk for the facilities contact
  6. DesignedReviewA person checks low-confidence results
Built and testedDesigned, not built yetPeople

The pipeline

Stage What it does Status
Import Reads messy CSV/Excel, maps columns, normalizes phones, sites, and zips, and removes duplicates across files Built
Pilot selection Picks a test set spread across industries, favoring rows with a phone and website Built
Research Web search and site crawl, with AI extraction of facilities staff and the main line Designed
Enrichment People search by domain and title, plus email verification Designed
Scoring Confidence score sorts each account into resolved or needs a call Designed
Calling An AI agent asks the receptionist for the facilities contact Designed
Review Low-confidence or conflicting results go to a person Designed

Every stage reads and writes one database, and the cost ledger (built) records every paid API call. The pilot will report cost per verified contact, not just contacts found.

Guardrails (designed in, enforced as each stage is built)

  • Research before calling. The phone is the last resort, not the first step.
  • Live calls off by default. The configuration already has a live-call switch set to off, and a test number that every call will be redirected to during testing.
  • Business landlines only. Numbers will be checked before dialing so mobiles are never called. This is designed around the FCC’s 2024 ruling that AI-generated voices count as “artificial” under the TCPA, and there will be a compliance review before any live call.
  • Disclosure and limits. The agent will say it’s automated and that the call is recorded. A do-not-call table already exists in the database. Calls are planned for business hours only, with no pitching.
  • Swappable vendors. Search, AI model, enrichment, and voice each sit behind a small interface with a zero-cost mock for testing.

What exists today

The database schema, the importer (column mapping, title rows in Excel, phone extensions, numbers Excel stored as numbers, zip+4, URL cleanup, cross-file dedupe), pilot selection, the cost ledger, the command-line tool, and vendor-neutral interfaces with mocks. 4 automated tests pass.

What comes next

Automated web research with AI extraction, then enrichment and confidence scoring, then test calls to our own phone only. No live call goes out without a compliance review.