Clean Your CRM

Clean your customer-tracking software before you add AI to it. Start with eight fields that decide what happens next: identity, owner, stage, value, last contact, next action, next-action date, and closed reason. AI can help repair the records, but it should not guess which customer, promise, or deal is real.

Problem: Your CRM says one company is three companies. The active deal belongs to someone who left. "Proposal sent" means four different things. Half the records have no next action. Then an AI assistant arrives and confidently builds a forecast from the mess.

Quick Win: Repair the fields that drive a decision before touching the archive. Salesforce's 2026 State of Sales survey found 51% of sales leaders using AI say disconnected systems slow their AI work, and 74% of sales professionals are focusing on data cleansing, including duplicate removal, error correction, and standard formats (Salesforce State of Sales 2026). The unglamorous work is the prerequisite.

AI makes the old mess move faster

Bad CRM data is annoying when a salesperson reads one record at a time. It becomes dangerous when software reads every record, ranks deals, drafts messages, and changes fields in bulk.

Salesforce surveyed 4,050 sales professionals across 22 countries for its 2026 report. The company found manual errors, duplicate records, disconnected systems, and security concerns are all getting in the way of sales AI. It also reports that 84% of data and analytics leaders say their data strategies need an overhaul to reach their AI goals (Salesforce State of Sales 2026).

That does not mean you need a company-wide data program before one useful automation can ship. It means you should clean the slice of data the automation will actually use.

If the first job is detecting forgotten follow-ups, fix active contacts, owners, last-touch dates, and next actions. Do not spend the week standardizing the fax-number field from 2014.

The eight fields that matter first

Start with the fields that change a decision.

The word "meaningful" matters. A marketing newsletter open should not reset the last-contact clock on a six-figure deal. Neither should an automated sequence email that received no reply.

Define what counts. A reply, completed call, attended meeting, signed document, or explicit scheduling exchange may qualify. Your definition can differ. It just cannot differ by salesperson.

The same goes for stages. "Proposal" might mean a document was drafted, sent, reviewed, or verbally accepted. Pick one meaning. Put the others in separate fields or stages if the difference affects what the team does.

First stop the mess from returning

Cleaning a database while new bad records pour in is miserable.

Before the bulk work, add simple controls to record creation:

1. Search existing companies by website domain before creating a new one.
2. Require an owner for every active deal.
3. Require a next action and date before a deal can remain in an active stage.
4. Limit stage and closed-reason fields to an approved list.
5. Keep free text for context, not for information the system must count.
6. Record important changes, including the previous value and who changed it.

Do not require every field. People fill mandatory fields with rubbish when the field does not help them do the job. Require the few fields that drive routing, follow-up, forecasting, or reporting.

The one-week cleanup

This is a repair sprint, not an archaeological expedition.

Day 1: Take a snapshot

Export or back up the current data before changing anything. Record the number of active companies, contacts, and deals, plus the share missing an owner, stage, next action, next-action date, or last meaningful contact.

Those counts are your baseline. Do not announce a cleanup win because the database "looks better."

Also list every system that writes to the CRM: website forms, email tools, calendar tools, billing software, enrichment providers, imports, and manual entry. A duplicate rule inside the CRM will not help if an integration creates a fresh record every night.

Day 2: Define identity

Choose the durable key for a company, usually a verified website domain plus a company identifier when subsidiaries matter. For people, email is useful but not permanent. People change jobs and addresses. Keep old identities as history rather than erasing them.

Group likely duplicates by confidence. Merge exact matches after review, queue strong matches for a person, and leave uncertain records alone until better evidence arrives.

Day 3: Fix active ownership

Every live relationship needs an owner or an explicit team queue. "Former employee" is not an owner. Neither is "admin."

Start with open deals, customers approaching renewal, and recent inbound inquiries. Assign the record to the person expected to act. If ownership is disputed, do not hide the dispute by picking a name at random. Put it in a review queue with a deadline.

Day 4: Make stages mean one thing

Write a one-sentence entry and exit rule for each active stage.

A proposal stage, for example, should say whether the document was drafted, sent, or reviewed by the buyer. Pick one meaning and require evidence before a deal enters it.

Then review open deals against the definitions. Do not move everything forward to flatter the forecast. Put uncertain records in a visible review state.

Day 5: Repair the clock

Follow-up automation needs two dates:

• The last meaningful contact.
• The next action due date.

Rebuild these for active deals from email, calendar, call, and meeting records where permitted. Do not treat every activity as a contact. A sent email with no reply is an attempt.

Gong says its pipeline research found sellers who prioritize timely follow-ups and complete open actions shorten deal duration by 11% on average. It also reports higher win rates when AI-recommended actions are completed. These are findings from Gong's own product data, so treat them as vendor evidence rather than a universal promise (Gong pipeline data research).

The useful lesson is narrower: an action cannot be late if nobody recorded when it was due.

Day 6: Add the exception queues

Some records will not fit the rules. Good.

Create queues for:

• Possible duplicates.
• Active deals with no accountable owner.
• Stages unsupported by evidence.
• Next actions with impossible or missing dates.
• Deals with recent customer activity but marked closed.
• Contacts attached to the wrong company.
• Records touched by an integration after manual correction.

An exception queue is more honest than forcing uncertain data into a clean-looking field.

Day 7: Test the first automation

Run the proposed AI or automation against a sample without letting it write changes.

For a follow-up detector, ask:

• Did it flag genuinely forgotten actions?
• Did it ignore automated noise?
• Did it send the alert to the right owner?
• Could the owner see why the record was flagged?
• Did it miss records because another system held the latest interaction?

Review the errors. Fix the rule or the data source. Run it again.

Only then allow limited writes, starting with suggestions or a review queue.

What AI can safely help clean

AI is useful for reducing a large pile to a smaller decision queue.

It can suggest that two company names refer to the same organization. It can normalize country and industry labels. It can extract a promised next step from a call summary. It can flag a stage that conflicts with the recorded activity. It can group free-text loss reasons into a controlled list.

Keep a person on irreversible or consequential changes:

• Merging company or contact histories.
• Reassigning a live deal.
• Changing deal value.
• Marking a relationship closed.
• Rewriting a customer's consent or communication preference.
• Sending a message based on the corrected record.

Ask the system to provide its evidence and confidence. "These look the same" is not enough. A suggested merge should show the matching domain, address, contacts, and conflicting fields.

The European Union's AI Act gives the same underlying lesson for covered high-risk systems: when deployers control input data, they must ensure it is relevant and sufficiently representative for the intended purpose, and they must assign competent human oversight (EU AI Act, Article 26). Your CRM workflow may not be a high-risk use under that law. Relevant inputs and real oversight are still sensible operating rules.

Measure whether the cleanup worked

Do not use "records cleaned" as the main result. It rewards bulk activity and says nothing about the sales process.

Track:

• Active deals with a valid owner.
• Active deals with a specific next action and date.
• Duplicate outreach incidents.
• Deals with contradictory stages.
• Follow-up alerts accepted, dismissed, or marked wrong.
• Time from an inquiry to an accountable owner.
• Forecast changes caused by corrected records.

The goal is not a beautiful database. The goal is a reliable next decision.

Where CRM cleanup goes wrong

The first failure is cleaning everything. A company spends six months fixing inactive contacts while live inquiries still have no owner. Active revenue comes first.

The second is merging on name alone. Similar names are common, and corporate groups are messy. Preserve uncertainty until the evidence is strong.

The third is letting activity replace meaning. A record can have 40 automated touches and no relationship. Define contact around two-way human or customer action.

The fourth is changing stage definitions without changing behavior. If the sales meeting still uses the old meaning, the new dropdown labels will drift within weeks.

The fifth is an AI cleanup with no audit trail. Bulk changes happen, the forecast jumps, and nobody can explain why. Keep the old value, proposed value, evidence, reviewer, and time.

The sixth is treating cleanup as a one-off project. Validation at entry, clear ownership, and exception queues are what keep the data usable after the sprint.

What comes after cleanup

Once the active data is dependable, start with detection.

Flag a new inquiry with no owner. Flag a live deal with no next action. Flag a promised follow-up whose date passed. Flag an account with meaningful activity after it was marked closed. Put each item in front of the person who can decide.

This is different from buying another database. Your CRM already records much of what happened. The missing layer notices when the expected next thing did not happen.

The CRM filing-cabinet problem explains why storage alone never watches the relationship. The real cost of slow follow-up shows how silence becomes a revenue leak. If old deals are already sitting untouched, use the dead-deal revival method after the active records are sound.

Clean data is not the finish line. It is what lets the company notice a slipped relationship while there is still time to recover it. That is the job of sales follow-up automation: one trustworthy view, clear ownership, and a running list of what needs attention now.

Frequently asked questions

Should we replace our CRM instead?

Usually not as the first move. A new CRM can improve the interface or data model, but migrating dirty definitions and unclear ownership moves the same problem. Define identity, stages, ownership, and next-action rules before deciding whether the current tool can support them.

Should we delete old CRM records?

Not by default. Old records can hold consent history, customer context, or evidence about past decisions. Set retention and deletion rules with the people responsible for legal, privacy, and security requirements. For the cleanup sprint, separate inactive records from the operational view instead of deleting them casually.

Can enrichment data fix missing fields?

It can fill public company details, but it cannot tell you who internally owns the relationship, whether the buyer agreed to a next step, or why a deal stalled. Enrichment helps identity and fit. It does not replace the interaction record.

What should the first AI workflow do?

Make it detect and recommend. Flag active records with no owner, no next action, or a passed due date. Show the evidence to a person and let that person correct the record or dismiss the alert. Earn the right to automate changes later.