Using Self-Learning AI in Production: Our First Model

In September 2025, we deployed our first in house self learning AI model into production. Developed by the Decision Science team, the model is now live within our outbound dialler environment.

This marks an important step in how we use data and automation to support operational decision making. Rather than being a one off technical delivery, it reflects a shift towards more adaptive models that can respond to changing behaviours over time.

As we move through 2026, we plan to build on this work, with five additional self learning models scheduled for delivery. These models are intended to support decision making, improve operational efficiency and strengthen customer engagement across different areas of the business.

Improving Customer Contact Decisions

The first production model focuses on dialler optimisation. Drawing on multiple data sources and newly engineered behavioural variables, it is designed to identify customers with a higher likelihood of engagement, enabling more targeted and efficient prioritisation of outbound contact.

This represents a change in how we manage dialler capacity, moving away from static prioritisation towards a more responsive approach that adjusts as behaviour changes.

Key characteristics and benefits

Adaptability: responding to changing behaviour

The model operates through a self learning pipeline that regularly incorporates recent dialler outcomes. As customer behaviour changes, the model recalibrates automatically, helping to keep decisions aligned with current patterns.

This reduces reliance on manual recalibration, supports more stable performance over time and lowers ongoing maintenance effort. It also improves transparency and predictability by ensuring decisions are based on up to date data.

Granular ranking to support operational decisions

The model produces a more granular ranking of accounts, allowing dialler strategies to flex more effectively in response to capacity constraints. By focusing contact attempts on customers with a higher likelihood of engagement, it supports improvements in contact effectiveness and cost efficiency.

This gives operational teams better information to support day to day decisions, particularly when workloads fluctuate.

Deployment under controlled governance

All self learning models are deployed within a defined governance framework. While the models are capable of adapting over time, any progression into production is subject to structured evaluation and approval.

A monitoring environment allows new model versions to be tested alongside the existing live model, enabling direct comparison of performance and behaviour. This provides clear, data led insight into how models are learning and how their decisions differ from the current baseline before any full deployment is approved.

This approach ensures that models are introduced in a controlled, transparent way and remain aligned with business and regulatory expectations.

Looking ahead

The deployment of this first self learning model is the starting point rather than the end goal. With further models planned for 2026, we are gradually building a broader AI capability that can be applied across multiple operational use cases.

The early results show that we can develop and deploy these models in house and integrate them into live operational environments. As this capability expands, it has the potential to support more consistent, data driven decision making across the organisation.