Article

Generative AI and sensitive data pseudonymisation in asset and wealth management

Sensitive data in asset and wealth management can be pseudonymised for testing, integration and analysis by combining Generative AI for data model interpretation and rule definition with deterministic, verifiable transformations that protect sensitive information and preserve data consistency.

Preparing data for test environments

For banks, asset management companies and other asset and wealth management firms, data is a core resource. It underpins operational processes, enables applications to be tested and supports increasingly advanced analysis. However, using data in non-production environments requires specific safeguards to ensure that real and sensitive information is never exposed in clear text during development, system integration testing (SIT) or user acceptance testing (UAT).

Preparing protected datasets can be a lengthy and complex activity. Standard approaches take time, must be repeated when the data model changes and do not always preserve relationships between tables or the meaning of individual values. The result may be a formally anonymised but unrepresentative dataset that limits test coverage, slows integration activities and makes realistic scenarios more difficult to reproduce.

The challenge becomes greater when data must be shared with counterparties or suppliers. Each additional copy requires clear rules, documented transformations and controls that demonstrate how the information has been protected before it is shared.

An effective approach to sensitive data pseudonymisation

By combining domain knowledge, Generative AI and deterministic transformations, sensitive data in asset and wealth management can be pseudonymised while preserving its analytical value. Generative AI supports the interpretation of the data model, the identification of personal information and the definition of the rules to be applied. The execution of pseudonymisation remains separate and follows defined, version-controlled and reproducible logic.

This separation makes it possible to use Generative AI to analyse large schemas with inconsistent documentation without allowing the model to make autonomous decisions about data transformation. The classifications, rules and code produced during preparation can be reviewed and validated by the client before execution.

Two complementary pipelines for data preparation and pseudonymisation

The process is organised into two pipelines with different responsibilities. The first prepares the data model and the pseudonymisation ruleset; the second applies the transformations and generates the controls needed to verify the outcome. This separation makes the process easier to manage, audit and update over time.

Data model preparation

The first pipeline analyses the schema and available documentation to identify and classify personally identifiable information (PII). Generative AI can interpret ambiguously named columns, notes, descriptions and free-text comments, proposing classifications consistent with the application context. It also supports ruleset preparation, code generation and the definition of controls.

Generative AI reduces the manual effort required to review extensive data models and accelerates the preparation of test datasets. When the schema changes, analysis can focus on the differences introduced rather than repeating the entire process.

Pseudonymization execution

The second pipeline applies the approved rules to structured data and relevant sections of free text. The transformation is deterministic: given the same ruleset and input dataset, the result is defined and reproducible. Generative AI inference is not used at this stage, and the model does not decide autonomously which data to transform or how to transform it.

Execution produces checkpoints, controls and a protected mapping of the relationships required for re-identification. The client can therefore reconstruct the transformations performed and verify that they have been applied correctly across the entire agreed scope.

Semantic consistency and referential integrity

For a dataset to remain useful for testing, the same entity must be transformed consistently wherever it appears, so that joins, relationships and end-to-end scenarios continue to work after pseudonymisation.

Semantic consistency must also be preserved. Replacement values must retain the required format, domain, validity constraints and consistency with other fields in the record. An account code, a transaction or a position cannot be treated as an isolated column: its meaning depends on the process and on the relationships within the data model.

Understanding these constraints within the asset and wealth management context enables their translation into rules that can be applied and controlled. The result is a representative, non-real dataset suitable for more complete and reliable testing.

Data protection and model selection

The approach is designed primarily for on-premises execution. Production data, transformation rules and pseudonymisation keys remain within the client's environment. The client defines access rights and retains control of the information required for re-identification. Activities are performed on a database clone or representative copy and do not modify production systems.

A hybrid configuration can also be considered, depending on the organisation's security requirements and policies. Cloud models may support selected activities involving only authorised metadata or information that has already been protected, while real data and keys remain within the client-controlled perimeter. The choice of model does not alter the deterministic transformation pipeline.

How pseudonymisation improves data preparation, testing and compliance

  • Faster and more efficient data preparation
    Automating data model analysis, PII classification and rule generation reduces repetitive manual work and the time required to prepare usable datasets for non-production environments. It also makes datasets easier to update when the underlying data model changes.

  • More reliable and higher-quality testing
    Preserving referential integrity and semantic consistency enables teams to work with realistic and coherent scenarios. Relationships between entities remain valid, integrations can be tested correctly and the risk of incomplete transformations concealing anomalies or producing unreliable results is reduced.

  • Safer collaboration with counterparties
    For SIT and UAT activities involving external parties, the approach creates a protected and documented perimeter before data is shared, reducing the risk of exposing sensitive production information.

  • Broader use of protected data
    Development and quality assurance teams gain faster access to consistent datasets without using production information in clear text. Where permitted by the client's policies, pseudonymised data can also support advanced analytics or cloud-based Generative AI tools, while the information required for re-identification remains separate and under the client's control.

Addressing data preparation challenges in SIT and UAT

While supporting SIT and UAT activities with banking counterparties for a leading asset management company, Technology Reply identified a recurring challenge in the preparation of test data flows. These flows included personal and portfolio information and required specific protection measures before they could be used or shared. Preparing them involved several operational activities, with potential implications for preparation times and data consistency. This highlighted the need for a more efficient, standardised and repeatable process.

In response to this need, Technology Reply combined Generative AI-assisted analysis of the data model with deterministic pseudonymisation rules. The aim was to streamline data preparation while preserving semantic consistency and referential integrity. Data, transformation rules and the information required for re-identification remain within the client's environment. Beyond SIT and UAT, the approach will eventually provide a protected and consistent data foundation for advanced AI analysis. Where permitted by the client's policies, pseudonymised datasets will also support the generation of additional synthetic data for testing or analytical purposes without exposing the original sensitive information.

Sensitive data pseudonymisation for asset and wealth management: from pilot to production

A recommended starting point is an initial implementation within a defined scope, designed to gather requirements, agree measurable success criteria, and configure and validate the approach against the client's data model and policies. The results provide a clear basis for progressive adoption across additional environments, data flows and use cases.

Technology Reply supports the streamlining of data preparation while maintaining data protection. Starting with a clearly defined scope makes it possible to measure improvements in preparation time, test effectiveness and data quality, and to establish an adoption roadmap aligned with the existing architecture, policies and priorities.

Frequently Asked Questions

Technology Reply is the company of the Reply Group specialized in designing and implementing innovative solutions based on Oracle technologies, supporting companies in data-driven transformation journeys and artificial intelligence. With over 25 years of experience, it accelerates innovation through modern data platforms, cloud-native architectures, and AI solutions. Its multidisciplinary teams follow the entire project lifecycle, from defining strategy and architecture to development, release, and operational management, creating scalable solutions designed to evolve over time. A key partner for Oracle Cloud Infrastructure and Oracle technologies, Technology Reply operates in the areas of Data Platforms, Analytics, Integration, Digital Applications, and Enterprise Architecture. Thanks to its expertise in Generative AI, Machine Learning, and Agentic AI, it develops autonomous agents capable of orchestrating data, applications, and workflows, enabling adaptive, autonomous, and data-driven business processes across various sectors.