BT weighs AI’s impact on sustainability goals

When it comes to meeting emission reduction targets, AI is both a great help and a potential hinderance for communications service providers (CSPs), as Mark Briers, Director of Data Science and AI at BT, discussed in an interview with TM Forum Inform. Given this apparent conundrum for the telecom sector, the operator has developed internal tools that will help manage the energy intensity of its increasing use of AI and Generative AI (GenAI).

BT aims to be net zero for direct carbon emissions (Scope 1 and 2) by the end of March 2031, and has implemented various AI tools to improve efficiency and reduce energy consumption in networks. In the nine months to 31 December 2024, the telco reported energy usage in its networks decreased 3%.

For example, the operator leveraged data intelligence and analytics to support the process of shutting down the legacy 3G network, which was completed in February 2024 and is expected to save 17.44 million KWh per year. BT has also rolled out “cell sleep” technology that automatically turns off 4G capacity when it is not needed, potentially delivering energy savings of 4.5 million KWh per year.

In such cases, AI is a “massive net positive contributor”, said Briers. The operator will continue to pursue network energy-saving initiatives “because it’s the right thing to do for the planet” as well as for efficiency gains for BT, he explained.

However, BT acknowledged it is “challenging” to calculate the overall impact of AI and GenAI on emissions because most of its usage is accounted for in the indirect emissions reported under Scope 3. For most companies, Scope 3 is the biggest source of carbon emissions and notoriously difficult to measure and manage.

The operator can make well-structured assumptions based on reporting from hyperscalers, as required, but it would like to have more granular insight into areas such as LLM training, which is the most energy intensive part of GenAI.

Low emissions impact for telcos 

BT does not train LLMs, and the energy intensity of using GenAI is “relatively low” for telcos because they are not involved in all the work of creating models and apps from scratch, said Briers.

And while applying the LLMs in so-called inference also demands a lot of energy, the impact on an organization’s emissions is meaningful only when there are “millions of concurrent users”, he explained.

“Offloading the [GenAI] work to public cloud and into Scope 3 territory is advantageous for a number of reasons, including energy intensity,” he said but emphasized that this does not make it any less of a problem.

“That would be like saying, I drive an electric car but I shouldn’t worry about the energy that I took to build the electric car. I’m not saying that it’s not a problem at all,” he said.

A challenge for telcos, or any enterprise users of GenAI, is that the technology is evolving so rapidly that it is difficult to keep up with changes further up the supply chain in hyperscaler data centers.

“We want the most expediently trained LLMs, but in that area, there is so much churn in the research literature that it’s hard to pinpoint which is the right one. And if we pinpoint it today, I can guarantee in a month’s time it will be different,” he said.

Gateway to sustainable AI

Sustainability has become a bigger factor in BT’s technology buying decisions. The operator’s procurement arm BT Sourced assigns at least a 15% weighting to purpose, which includes responsible and sustainable criteria in initial supplier assessments.

With the proliferation of GenAI, the operator has also given more thought to sustainability in how it develops and applies the technology to new use cases as it seeks wider efficiency gains and cost savings.

“Cost efficiency is highly correlated, in terms of computational units, with energy efficiency. Where we strive for one, we get the other, and so we strive for both,” said Briers.

For example, BT’s internal GenAI Gateway platform, built in collaboration with Amazon Web Services, provides access to multiple LLMs, such as from Amazon, Anthropic, Meta and Cohere. It enables BT to manage costs, security and privacy when developing its own GenAI applications.

By decoupling the applications from the underlying hardware and foundation models, the platform gives BT flexibility to choose the most cost-efficient options as advances are made among the cloud providers, Briers explained. The platform is currently linked to AWS but will eventually work with other hyperscalers.

“The world is moving so fast…We developed this platform so that we could switch models and not have to do constant re-engineering. Now, as the AI ecosystem evolves, we can continually evaluate these models across multiple dimensions, like performance, accuracy, hallucination potential, as well as energy efficiency and cost,” he explained.

The gateway applies to scenarios where BT needs to build its own tool for specific processes that out-of-the-box GenAI cannot address. In cases where it buys tools like Microsoft’s Co-Pilot, AWS’s Q Developer, or ServiceNow’s embedded AI for service agents, the operator is developing an evaluation framework that will include requirements for performance and sustainability, among others.

BT can then compare supplier scores against external benchmarks and, eventually, use the results as leverage to encourage vendors to improve energy efficiency or other functional requirements.

Briers said this “engineering good practice” will help BT to “optimize in the energy space.”

Although difficult to have full visibility of GenAI’s impact on emissions at this stage, BT is taking steps to curb energy intensity where it can and factoring sustainability considerations into AI business cases.

“If anything was put past my desk that had a significant negative impact to the environment, then I'm sure I will not be the only one shouting from the rooftops for us not to allow that forward,” said Briers.