Published: Dec 11, 2025
NCS pilots Responsible AI with Singtel
Enterprises are under increasing pressure to adopt responsible AI, yet most open-source toolkits are not built to support complex, large-scale enterprise datasets. Singtel needed a customised approach to evaluate its live AI solutions against its internal Responsible AI principles. Through a proof of concept together with NCS and the NUS-NCS Joint Lab for Cyber Security, Singtel enhanced 4 open-source toolkits to integrate seamlessly with Databricks, deliver detailed risk reporting, and scale easily for future AI models. This approach sets a new benchmark for the telcos. As a result, Singtel reduced AI risks, strengthened compliance, and built trust in its AI systems, while cutting test times from 9 hours to minutes.
Key takeaways
- Configurable frameworks that scale across AI models and principles
- Detailed risk reporting with explainability beyond binary outputs
- Seamless Databricks integration for enterprise workflows
- Test time cut from nine hours to minutes
The challenge
Rapid AI adoption demanded stronger governance to ensure ethical, fair, transparent, and secure systems. Singtel needed to validate its AI model against 10 Responsible AI principles defined internally. However, existing toolkits could not handle complex data from both predictive and Generative AI models. Reporting was shallow, runtimes were slow, and the tools did not align with Singtel’s Databricks workflows. Singtel needed a customised evaluation approach that could be tested within a proof-of-concept scope before wider rollout.
The solution
Singtel partnered with NCS and the NUS-NCS Joint Lab to design and implement a proof of concept that customised and optimised 4 open-source toolkits to assess and strengthen 2 live AI solutions, the Shirley chatbot, which leverages Generative AI to handle customer enquiries, and an internal customer engagement model which uses Predictive AI for customer contract-renewal insights. This Responsible AI Toolkit was engineered to handle enterprise-scale data securely and efficiently, setting a new operational standard for Responsible AI in the telco industry.
Snapshot of capabilities
- Detailed risk reporting: Pinpoints specific prompts, features or model components that contribute to risks rather than providing binary outputs
- Seamless integration with Databricks: Optimised toolkits to run efficiently on Singtel’s distributed Spark clusters
- Future-proof and configurable: Can be easily extended to cover more AI models and Responsible AI principles
The impact
The Responsible AI toolkit enabled Singtel to detect and mitigate bias, improve model explainability, and ensure consistent Responsible AI practices across their operations.
- Reduce risk: Mitigate bias, unsafe responses, and reputational risks across AI systems
- Strengthen compliance: Provides clear, evidence-based reports for audits and regulatory checks
- Build trust in AI: Increases confidence among business users, stakeholders and customers in the chatbot’s reliability
- Boost operational efficiency: Test time reduced from 9 hours to minutes, saving time and computing resources