---

Data cleanup is where AI promises the most and where the most irreversible mistakes happen. A tool that can find duplicates, standardize job titles and fill missing fields can transform a neglected CRM in an afternoon. It can also merge two different customers into one record, overwrite correct information with wrong guesses, and make the problem harder to untangle than before. The difference lies in how you run it.

Understand the three kinds of cleanup

Not all cleanup carries the same risk, so treat them separately.

Standardizing changes how values are written without changing what they mean: country names, phone formats, capitalization, job title variants. Low risk, high payoff, and easy to verify.

Deduplicating finds records that refer to the same person or company and merges them. Medium to high risk, since a wrong merge combines history, ownership and activity, and can be hard to reverse.

Enriching fills missing fields from other sources or inferences: industry, company size, a missing role. Risk depends on source quality. AI-inferred values can look authoritative while being guesses.

Handle them in that order, starting with the safest, and use each stage to learn how your tool behaves.

Before you run anything

Take a backup you can restore. Export the affected objects, or use the platform’s backup and recovery option if it has one. Confirm the export actually opens and contains what you expect. Many teams discover during a crisis that the export lacked the fields they needed.

Define what a duplicate is. For contacts, perhaps the same email address, or the same name and company. For companies, the same domain, or the same legal name plus location. Write the rules down. A fuzzy match on names alone will merge unrelated people who share a common name.

Decide the winner. When two records merge, which values survive? Typically the most recently verified, the most complete, or the one owned by the active account owner. Specify it per field. Ownership, consent status and opt-out flags deserve special care, since losing an opt-out is a compliance problem, not just a data problem.

Check your dependencies. Automations, integrations, reports and routing rules may reference record IDs or fields. A merge can break them. Review before running, and pause automations that could react to a flood of changes.

Run in stages

1. Sample. Run the logic on a small slice, such as a few hundred records, and review every proposed change by hand.
2. Review the edge cases. Look especially at near-matches the tool was unsure about. These show where your rules are loose.
3. Run a larger batch with logging. Make sure each change is recorded with before and after values.
4. Spot check. Pull a random set from the results and verify against reality.
5. Full run. Only then process the remainder, ideally in batches, not one shot.

Keep a human approval step for anything above a defined risk, such as merges between records with different owners or large accounts.

Confidence thresholds and review queues

Most tools produce a match confidence. Use it. High-confidence matches, such as identical email addresses, can be handled automatically or with light review. Medium-confidence matches go to a queue for a person. Low-confidence ones are left alone. Tune the thresholds from what you see in the sample, not from the default.

For enrichment, treat inferred values as suggestions with a source label, not as facts. If the field matters, such as company size for routing or segmentation, record where it came from and when, so you can distrust and refresh it later.

What to put on the plan

Step	Purpose	Done when
Backup and test restore	Allows rollback	Restore tested on a copy
Match rules written	Prevents false merges	Rules reviewed by sales ops
Field survivorship decided	Controls what is kept	Documented per field
Automations paused or reviewed	Avoids side effects	Dependencies listed
Sample run reviewed	Catches rule errors	Every change inspected
Logs enabled	Enables audit and rollback	Before and after values saved
Full run in batches	Limits blast radius	Spot checks pass after each batch
Ongoing prevention	Stops new mess	Duplicate rules at entry in place

Prevent the mess from returning

Cleaning once without addressing the source simply resets the clock. Add duplicate detection at the point of entry, require key fields on forms and imports, and standardize picklists so free-text variants stop appearing. Assign someone to review data quality on a schedule. AI cleanup works best as a periodic tool inside a process, not a rescue mission every two years.

Privacy and compliance

Merging and enriching records changes the personal data you hold. Requirements around consent, accuracy, retention and individual rights vary by jurisdiction. Be careful that enrichment does not add data you have no basis to hold, and that merges preserve consent and opt-out status. Consult a qualified professional if you handle regulated or sensitive data.

A hypothetical lesson

A team ran an AI deduplication across their contact list using name and company similarity. About a thousand merges looked fine, but several accounts for a regional franchise group were collapsed into one because the businesses shared a parent name. Orders, notes and ownership merged together, and two reps both believed they owned the account.

Because the team had a pre-run export and before-and-after logs, they restored the affected records within a day and tightened the rules to require a matching domain or tax identifier for company merges. The cost was an annoying week. Without the logs, it would have been a long investigation.

Frequently Asked Questions

Is it safe to let AI merge records automatically? Only for very high-confidence matches with clear rules, and with logging. Anything involving different owners, large accounts or consent data should be reviewed by a person.

How often should we run deduplication? Prevention at entry reduces the need. Many teams do a periodic review, such as quarterly, sized to how fast their data grows.

Can we trust AI-enriched fields? Treat them as unverified unless you know the source. Label them and avoid using them for decisions where a wrong value is costly.

What if a merge was wrong and we have no backup? Check whether your CRM keeps merge history or a recycle area, and contact your vendor. This is why a backup beforehand is non-negotiable.

Should reps be involved? Yes, for the ambiguous cases. They know which “Acme” is which and can spot a wrong merge quickly.

Next Step

Export your contacts and companies today, open the file to confirm it is complete, and then write your duplicate rule for one object in a single sentence. Run only the standardizing pass first, and review a sample before touching any merges.