How to clean CRM data at scale

You cannot clean what you have not measured. Before the first merge, export every account, contact, lead, and opportunity to CSV and run basic profiling queries. For each object, measure: percentage of records with the required fields populated, duplicate rate by exact-match email and by fuzzy-match company name, percentage with no activity in the last ninety days, and percentage missing a related record (contacts with no account, deals with no primary contact, accounts with zero contacts). Write these numbers down. They are your baseline. Without them, you will finish the cleanup and have no way to prove it worked, which means leadership will quietly stop funding the work and the problem will come back in six months. Profile across the full dataset, not a sample. Fuzzy duplicates cluster in the long tail and sampling hides them.

• Export accounts, contacts, leads, and opportunities to CSV in full, not sampled
• Measure required-field completeness per object as a percentage, not a count
• Measure exact-email duplicate rate and fuzzy-company-name duplicate rate separately
• Measure the percentage of each object with zero activity in the last ninety days
• Measure orphan rates: contacts without accounts, deals without primary contacts, accounts without contacts
• Save the baseline numbers in a shared doc and date them; you will compare against these after each phase

The word duplicate hides more disagreement than any other term in CRM work. Before you build matching rules, decide and document what constitutes a duplicate in each object. For accounts, the usual rule is same legal entity, which means same domain for software companies, same legal name plus state for services firms, and same parent company plus billing country for multinationals. For contacts, the rule is usually same person, which means matching work email, or matching personal email plus phone, or matching first name plus last name plus current employer. For leads, you have an extra decision: do leads with the same email as an existing contact count as duplicates, or do they get merged into the contact. Pick one answer and stick to it. Write the rules down in plain English before translating them into match criteria. Rules that live only in your head produce merges nobody can explain six months later.

• Write the account duplicate rule in one sentence, then enumerate the exceptions
• Write the contact duplicate rule the same way; name whether personal email counts
• Decide the lead-to-contact collision policy: merge into contact, or hold as separate lead until converted
• Document the rules in your admin wiki and reference them in the merge job name

Order matters. Dedupe accounts first because every contact and deal points at one, and merging accounts automatically repoints the children. If you dedupe contacts before accounts, you will merge two contacts that each point at a different account, then merge the accounts, and the second merge will produce new conflicts you have to resolve by hand. Run the account merge as a two-phase job. Phase one: exact-domain matches. These are safe, high-confidence, and typically account for sixty to eighty percent of the duplicate volume. Automate the merge, pick the record with the most recent activity as the winner, and preserve all notes, activities, and related records from the loser. Phase two: fuzzy matches on company name, same billing country, created more than thirty days apart. Review these in a merge queue with a human in the loop. Never bulk-merge fuzzy matches without review; the false-positive rate is high enough to corrupt real customer relationships. After accounts, dedupe contacts using the rules you wrote in step two, then do leads last.

• Phase one accounts: exact-domain merge, automated, winner is most-recent-activity record
• Phase two accounts: fuzzy-name merge, human-reviewed in a queue, batched at fifty per sitting
• Dedupe contacts after accounts settle; email-exact is safe to automate, personal-email-plus-phone needs review
• Dedupe leads last; if your policy is to merge leads into existing contacts, run that sweep now
• Log every merge in a cleanup audit table so you can roll back a specific merge if a rep flags it

Now that the record count is smaller and cleaner, run enrichment. The economics here matter. Most enrichment providers charge per record, so enriching before dedupe wastes thirty percent or more of your spend on records that will be merged away. For accounts, enrich firmographics: industry, employee count, revenue band, HQ country, and tech stack if that signal matters to your motion. For contacts, enrich title normalization, seniority, department, and verified work email. Pick a single source of truth per field; if ZoomInfo and Clearbit disagree on employee count, decide which one wins before the sync runs, do not let the last write win. Set enrichment to run on create and then on a quarterly refresh for stable fields and monthly for volatile fields like job titles. Do not enrich every field every day; it costs money, hits rate limits, and makes change tracking impossible.

• List the firmographic and demographic fields that actually drive segmentation or routing; enrich only those
• Pick a single provider as source of truth per field; do not allow two providers to write the same field
• Set enrichment cadence: on create, quarterly refresh for stable fields, monthly for job titles and employee counts
• Flag records where enrichment returned low confidence; those go to a review queue, not into production
• Reconcile billing monthly to catch provider sync loops that burn credits on the same record daily

Not every record deserves to live forever. The hardest conversation in CRM cleanup is agreeing to archive records that reps emotionally refuse to delete. Build a written archival policy and get leadership signoff before you archive a single record. A reasonable baseline: accounts with no activity in twenty-four months, no open opportunities, and no active contacts move to an archived state. Contacts with bounced email, no activity in eighteen months, and no related open deal move to archived. Leads older than twelve months with no response move to a nurture pool outside the primary lead object or get archived outright. Archive does not mean delete. Keep the records readable, exclude them from default list views and reports, and preserve them for compliance and historical reporting. Deletion is only for records you are legally required to remove, such as GDPR subject access requests or hard-bounce suppressions. Everything else is archive, not delete.

• Write the archival rules per object, including every exception leadership wants to carve out
• Set a status field (active, archived, suppressed) rather than deleting; archive is reversible, delete is not
• Exclude archived records from default views, routing, and reports; keep them visible on direct lookup
• Review archival volumes monthly; if archive rates spike, something upstream changed and you need to find it
• Keep a six-month unarchive window during which a rep can reactivate with a documented reason

Standardization is the step most teams skip, and it is the reason most CRM dashboards lie. Every picklist, every country code, every industry label, every stage name needs one spelling and one value. Pick a controlled vocabulary and enforce it. Industry should match a published taxonomy like NAICS or SIC rather than free text; otherwise you will have Software, Software and Technology, SaaS, Software as a Service, and Technology Services all competing in the same report. Country codes should match ISO 3166, not United States, USA, US, and America fighting for the same field. Job titles need a seniority mapping (IC, Manager, Director, VP, C-level) that strips the eight-hundred unique strings reps type into a title field down to a reportable set. Standardization happens in three places: the picklist definition, a nightly normalization job that cleans the long tail of free-text drift, and a validation rule that blocks new drift at creation. If you skip any of the three, drift wins over a quarter.

• Convert industry, country, state, and seniority to picklists tied to published taxonomies
• Build a nightly job that normalizes legacy free-text values to the controlled vocabulary
• Add validation rules that block new records from using off-list values
• Set a quarterly review to add or retire picklist values based on actual sales patterns, not committee opinion
• Separate display label from stored value so you can rename labels without breaking historical reports

Clean records that point at nothing are almost as useless as duplicates. Walk every orphan class your profiling flagged in step one and resolve them. Contacts without an account: match to an account by email domain where possible; if no match, create a stub account named after the domain and flag it for enrichment. Opportunities without a primary contact: this one is a forecasting hazard because you cannot route communications or run close-plan reports; assign the deal creator as the temporary primary contact and surface the gap in a weekly rep view. Accounts without contacts: if the account has an open deal, flag it urgently; if not, archive it under the step-five policy. Opportunities with closed dates in the past but still-open status: these are the single most common forecast lie; close them with a loss reason of time-expired or push the close date with a required justification note. Fix orphans before you declare the cleanup done. They are the records most likely to break downstream reports once everything else looks tidy.

• Match orphan contacts to accounts by email domain; stub-create accounts for unmatched contacts
• Assign a temporary primary contact to every opportunity missing one and surface the gap for the rep
• Flag accounts without contacts; open-deal accounts are urgent, no-deal accounts follow the archival policy
• Close or justify every opportunity whose close date is in the past; this is the single most common forecast corruption
• Rebuild the parent-child relationships on any account merged in step three where a child was accidentally detached

A one-time cleanup is a waste of two weeks if you do not install the guardrails that keep the data clean. Four guardrails matter. First, dedupe at creation: your CRM should check for existing matches when a rep or import tries to create a new account or contact and either block the create or offer to merge. Second, required-field enforcement at stage transitions: a deal cannot move from stage two to stage three without the fields that stage three reports require. Third, decay flags: a daily job that flags any record that has drifted from the controlled vocabulary, any record that lost a required field, and any record that lost its parent relationship. Fourth, a weekly data health review: a thirty-minute meeting with revenue operations and a sales manager where you review the decay dashboard, approve the merges, and sign off on the archivals. If any of the four guardrails is missing, drift will return within a quarter. If all four are in place, you can run the full cleanup pass once a year instead of once a quarter and spend the saved time on actual revenue work.

• Enable dedupe-at-creation on accounts, contacts, and leads with the matching rules from step two
• Attach required-field validation to every stage transition in your pipeline
• Build a decay dashboard that counts drift, missing fields, and orphan creation per week
• Schedule a thirty-minute weekly data health review with revops and a sales manager; make it recurring
• Measure the baseline numbers from step one on a monthly recurring cadence; publish the trend line to leadership