Clean Your CRM Data Before AI Touches It

Your CRM has a contact named "John Smith" and another named "J. Smith" at the same company, same email domain, entered six months apart by two different people. Your lead scoring model sees two contacts with moderate engagement. The real John Smith visited your pricing page four times and opened seven emails. The model never sees that. It scores a fragmented version of a person who should be at the top of your pipeline.

The model is not the problem

A reasonable objection to raising data quality before model quality goes like this: modern AI features in platforms like HubSpot are engineered to tolerate noise, and a well-configured scoring model with sensible feature weights will produce directionally correct rankings even with some contamination. This is not wrong in every case. At low duplicate rates, a model prioritising high-signal behaviours like demo requests over single email opens will still surface your best leads roughly in the right order.

The objection fails on the specific mechanics of how duplicates corrupt signals. When a contact record splits across two rows, the engagement history splits with it. The model does not receive a noisy version of the complete signal. It receives half a signal, twice. No weight adjustment fixes a frequency count that was halved before the model ran. ISO/IEC 25012, the data quality standard most directly applied to CRM analytics, classifies uniqueness as a property of the data itself, upstream of any model configuration.

What ten to thirty percent looks like in practice

Industry analyses of CRM databases put duplicate rates between ten and thirty percent. On a 5,000-contact database, that is up to 1,500 records the model treats as distinct people. Churn prediction depends on reading behavioural patterns across time for the same customer. If that customer appears under two slightly different company names or email formats entered by the same sales rep over two years, the longitudinal record the model needs to detect departure signals does not exist as a single coherent object. The research connecting decentralised data storage and weak system integration to quality decay is explicit: inconsistency accumulates even in single-rep, single-channel environments, because naming conventions and lifecycle stage labels drift without enforced standardisation.

The checklist, in order of impact

Export your full contact list from your CRM as a CSV. Open it in Google Sheets. Start with uniqueness, because duplicates cause the most direct damage to frequency-based AI features.

Use the Remove Duplicates function under Data to flag exact-match duplicates on email address. Then write a COUNTIF formula against the email column to surface partial duplicates where the same address appears with a trailing space or capitalisation difference. Conditional formatting set to highlight any COUNTIF result above one makes these visible in seconds. Merge the records manually, keeping the most recently updated version and copying any engagement history the other record holds.

Move to consistency next. Create a separate column using a formula to extract the domain from each email address, then sort by domain. Contacts at the same company entered under "Acme Corp," "Acme," and "ACME Inc." will cluster together. Standardise the company name field across all of them before your next model run.

For completeness, use a COUNTIF or ISBLANK formula to count how many records are missing values in the fields your AI feature uses as inputs. Lead scoring models that weight job title or industry will silently downgrade or ignore contacts where those fields are empty. Fill what you know. For the rest, flag those records as low-confidence inputs so you read their scores accordingly.

Set data validation rules on the fields you just cleaned. Under Data > Data Validation in Google Sheets, restrict company name entries to a dropdown list of your known accounts. This does not fix historical drift, but it stops new entries from reintroducing the same inconsistencies the moment you re-import.

What you will see after

Re-import the cleaned file and re-run your lead scoring. The contacts whose records were split will consolidate their engagement history. Scores that were suppressed by fragmented signals will move. Your churn prediction model will have continuous records to work with instead of broken timelines. The AI did not get smarter. The inputs it was always meant to receive finally arrived intact.