Amdocs says its GenAI is ready to move on from PoC stage

For all the hype around generative AI, its implementation in the telecom industry (and elsewhere) still seems to be at a relatively early stage. Yet many companies are making progress with deploying it in the telecoms space.

Speaking to Light Reading during the recent Digital Transformation World (DTW) event in Copenhagen, Amdocs' chief marketing officer, Gil Rosen, said his employer's amAIz platform is ready to be put in front of customers, claiming it performs better than human customer service agents.

Primarily known for developing business and operational support system (B/OSS) software, Amdocs is using telecom-specific data to adapt large language models (LLMs) for use in the telecom industry. While it doesn't build LLMs – it is LLM-agnostic – the company trains them further, using information that isn't publicly available and would not have been used in the system's original training. Moreover, Rosen said, "we know which [LLM] would most likely give the best answer."

He argues that generic GenAI is ill-suited to telecom-specific applications. Asked by a customer about the difference between two bills, a generic LLM might not only give the wrong answer but also be less efficient from a cost and performance perspective, according to Rosen.

"Nvidia, Microsoft and also AWS and Google basically chose Amdocs for their go-to-market in the telecom space," because they need a third party to connect the customer information systems to the LLM, said Rosen. "They need that middleware and, basically, knowledge in order to bring their capabilities to our industry."

The cloud-based physical stack is handled by partners, including Nvidia and Amazon, he added. What Amdocs has built is the telecom taxonomy running on top of the physical stack.

GenAI trained on telecom-specific information could also help address network issues that cause disruption for enterprise customers, according to Amdocs. Usually, said Rosen, "it takes a network operator a very long time to figure out where the problem is and which customers it affects."

With GenAI, however, it is possible to identify which customers are affected and which ones are the highest-paying. 

Amdocs is currently moving from the proof-of-concept (PoC) stage to production environments. On average, setting up a PoC takes about two to three months before it can be scaled to production "on a continuous line," said Rosen. This may, however, mean adding more computing resources or bolstering security.

The amAIz capability is pre-integrated in Amdocs' latest suite of products but could be included as an add-on for legacy products. Amdocs has also been working with customers that do not use any of its other products, according to Rosen.

In practice, he said, GenAI brings an improvement for telcos of anywhere between 30% and 80% – measured in terms of either performance or cost efficiency. This could mean, for example, reducing the handling time for customer calls, or cutting use of computing resources, according to Rosen.

Leap of faith

"And that's one of the concerns because the compute needed for generative AI is so big," he added, noting that this also brings added environmental benefits. On the technical side, an operator could, for instance, parse a task into "the right slices" and send only what's needed to the back end to get an answer, he explained.

On the issue of hallucinations, Rosen acknowledged LLMs can invent information. During initial training, he believes they are accurate only about 80% of the time, but Amdocs has since "minimized hallucinations almost to zero," he claimed, and it continues to focus on that effort.

"It's more of a trust journey on the customer side to move from a co-pilot for an agent to putting an LLM in front of the customer, trusting that it will not make these mistakes," said Rosen. "We're there from a capability perspective, but now it really depends on a leap of faith from the industry to actually implement it in the field."

He maintained, however, that the combined technologies of companies like Nvidia and Microsoft, together with the amAIz framework and Amdocs, is helping. "You could today put it in front of customers and I have to tell you, it's as if not more accurate than regular human beings," he said, pointing to the example of the company's GenAI-powered avatar that was on show during the event.

The regular human beings employed in telco call centers worldwide may not be thrilled to hear they are replaceable by a piece of software. But their employers, which constantly complain about revenues, can be expected to jump at the opportunity.