Case Study · Data & AI Systems

Privacy-first CRM data standardisation with AI

A controlled workflow for improving a large CRM dataset while keeping personal and commercially sensitive customer data inside the customer’s AWS environment.

The challenge

Improve CRM quality without moving raw data outside the customer’s controlled environment.

The work began with Nearly 300,000 contact records. The objective was to identify the records relevant to a target market, standardise them, and enrich associated company information without treating AI as an unchecked replacement for data governance.

All production-scale processing ran through AWS Glue scripts within the customer’s AWS VPC. Raw customer records were not downloaded for analysis. Instead, the workflow produced aggregate metrics, metadata, confidence distributions, and exception reports for review.

Outcome

A narrower, higher-confidence population ready for controlled standardisation.

Nearly 300,000 contacts scanned
roughly 45,000 initial candidate contacts
around 41,000 final relevant contacts
roughly 10,000 associated companies
about 99% of the final population standardised

Method

AI assisted the method; it did not operate without controls.

  1. Define the data boundary Sensitive data remained in the customer-controlled AWS environment.
  2. Translate business knowledge into rules AI helped develop and refine explicit, testable signals for identifying relevant contacts.
  3. Run the rules close to the data Scripts processed the full dataset consistently at scale and produced reviewable metrics and exceptions.
  4. Enrich company-level gaps selectively Authoritative matching came first; AI then assisted with public company-level research and classification for unresolved cases.
  5. Keep uncertainty visible High-confidence results could proceed; lower-confidence cases were left blank or routed to human review.
  6. Apply changes safely Every change was first represented in a dry run, reviewed before application, audited, and retained rollback support.

What this demonstrates

Useful enterprise AI is controlled, selective, and traceable.

The value came from combining AI, deterministic software, trusted reference data, and human review—not from treating AI as a magic button. This approach allowed the customer to improve data quality while preserving privacy, operational control, and a clear audit trail.

Discuss a data project