CRM Enrichment Playbook

Claude Code reads the CRM through its API and writes the field map before any data is bought.

• Claude Code pulls every company and contact field from HubSpot or Salesforce:
  • Its type and allowed values
  • How often it is filled
  • When it was last updated, and by whom
• It flags duplicates, fields nobody fills, and fields that contradict each other.
• It compares your won and lost deals, so the map starts from the fields that separate them.
• Each field in the map gets four rules:
  • Allowed values: the exact options the CRM accepts
  • Source of truth: where the value should come from
  • Refresh cadence: how fast it goes stale
  • Overwrite rule: who may replace it
• The field map and overwrite rules follow Jacob Tuwiner's HubSpot enrichment series, rebuilt for coding agents.

• Claude Code
• HubSpot
• Salesforce

A value a person put in the CRM beats anything a vendor finds, so the overwrite rules are set before the first run.

• A blank field is filled automatically.
• A value a rep typed in, or one your product collected, stays. A mobile number a customer gave your rep beats any vendor's.
• A rep's value can only be replaced once it is stale, for example after nine months with no update, and only after review.
• Customers and open deals get the strictest rules. The rest of the database can be filled more freely.
• HubSpot records where every value came from, so the rules can tell a rep's entry from an import. In Salesforce, turn on field history tracking for the fields you protect.
• Back up the CRM before the first write.

• HubSpot
• Salesforce

The CRM loads into a Freckle workbook on a schedule, and only records with a blank or stale field go through.

• Freckle imports a HubSpot list or filter, or a Salesforce view or query, on a schedule.
• Later imports update the same records, so nothing is duplicated.
• A value the CRM already has is reused, never bought again.

• Freckle
• HubSpot
• Salesforce

Company fields fill through a waterfall. Providers run in order of precision, and each one only sees what the one before it missed.

• Apollo fills headcount, industry and location. TheirStack fills the tech stack.
• Freckle's research agent is the last step for anything the providers miss.
• The company's LinkedIn page is found two ways, from the URL already in the CRM and from the domain:
  • They match: done
  • They differ: the research agent decides by your rule, such as the brand's own page over its parent company's
• The tiebreak only runs when the two disagree, so the check stays cheap.

• Freckle
• Apollo
• TheirStack

Some fields only exist in your own definitions. Freckle's research agent reads each company's site and answers them.

• Typical questions:
  • Does it sell to businesses, consumers, or both?
  • Which of your own categories and subcategories fits it?
  • Is it a franchise brand or the parent company?
• The agent gathers its own evidence first, like who the company sells to and what it offers, then answers.
• Every answer is checked against the CRM field's allowed values. Anything else is left blank.

• Freckle

Jev sorts each account into your tiers from the data the row now holds.

• Jev runs inside Freckle as its decision agent. It classifies and scores only, and never looks anything up.
• It flags a row it can't decide, so nothing is tiered on a guess.
• The tier sets how deep the contact search goes:
  • Deep: customers, open deals and tier one get every field, mobiles included
  • Light: everyone else gets the few fields that route them

• Jev

Emails and mobiles come from a waterfall, and every email is checked twice before it is written.

• Work emails come from LeadMagic, then Findymail for the ones it misses.
• Mobiles, for the deep tier only, come from Freckle's own finder first, then BetterContact.
• A research agent can't reliably find emails or phone numbers on the open web, so the contact waterfall ends at its last provider.
• MillionVerifier checks every address first, then BounceBan checks every address that passes. Freckle calls both through its HTTP step.
• An address that fails is never written. The contact keeps a flag instead, so nobody is silently dropped.

• LeadMagic
• Findymail
• BetterContact
• MillionVerifier
• BounceBan

Nothing runs on the whole database until it matches what you already know.

• Pick 20 to 40 records you know well, like current customers, and fill in their fields by hand.
• Freckle runs the workflow on a sample first and forecasts the credits for the full run.
• Claude Code compares every field with the hand-checked answers. Each field has to be right at least 90% of the time.
• A field below that bar gets a better prompt or another provider, then it is tested again.
• The hand-checked answers always come from a person, so no model grades its own work.

• Freckle
• Claude Code

Every change is compared with the live record, and only safe changes go straight in.

• Claude Code builds the writer on the CRM's API. Each record is read again right before the write, so an edit a rep made mid-run is never lost.
• Each change is sorted:
  • Blank field: filled automatically
  • Different value: reviewed field by field in Slack before anything is written
  • Set by a rep: left as it is
• Every write is logged with the old value, the new one and the rule that allowed it, so any field can be put back.

• Claude Code
• HubSpot
• Salesforce
• Slack

Records are re-enriched when something changes, and each field refreshes on its own schedule.

• New records are enriched as they arrive.
• Freckle watches your contacts for job changes and your accounts for new hires.
• A job change re-enriches that person and alerts the account owner in Slack.
• Fields past their refresh date go back through the workflow. Slow fields, like business model, are filled once.

• Freckle
• Slack