Agentic AI has swept across the telecom and business sectors this year after largely being a curiosity in 2024, a remarkable rate of acceleration which highlights its pivotal role.
Andy Markus, SVP, chief data and AI officer at AT&T told Mobile World Live (MWL) agentic AI divides complex elements into smaller, manageable tasks.
“We’re designing the entire workflow to be agentic,” he said. “We’re taking the bigger complex problems, breaking them down into smaller pieces and then we’re creating agents to be great at each piece, bringing those together in an agentic workflow to solve the overall solution in a much more accurate way.”
While the concept of chaining agents together to solve problems across networks or customer service operations seems simple enough, Markus explained agentic workflows are not autonomous. AT&T is using checkpoints in the process with human oversight.
Without this human element, chaining agents together could lead to AI hallucinations or inaccuracy rates of a few percentage points within each step. Markus said accuracy is a high priority for the deployment, along with low latency and return on investment.
AT&T’s agent army
AT&T currently has more than 410 generative AI agents in production. It is using a framework called Ask AT&T Workflows to develop those agents in software development cycles.
Ask Operations is being used by employees on the network side to automate tasks using agentic AI with human oversight.
“We also have this concept called Ask Data, which is the ability to take generative AI and have the subject matter expert interact with their data with human language,” Markus said.
Its framework allows AT&T employees to build their own agents and learn which are available from third parties. Markus compares the agents to Lego toys, which are interlocking plastic building blocks that allow users to create various structures and models.
“You could create a totally different workflow with these 410 agents that serve different use cases, and then people will create more as we go along,” he explained.
AT&T is also creating a registry of software-as-a-service (SaaS) provider agents it can use as needed, which could be useful for enterprise cases.
“These agents could be called upon to perform specialised tasks via agent-to-agent communication,” Markus said.
He noted AT&T uses open source small language models (SML), such as Meta Platforms’ Llama 3, for fine-tuned use cases.
While a chain of AI agents is more accurate than a single large language model, Markus noted the former can be more expensive.
“They chat with each other to do their job, and so there’s more token calls to the solutions,” he said of the AI agents. “The solutions are slightly more costly but much more accurate.”
Agentic AI in action
Meanwhile, Verizon is using agentic AI across its customer services by tapping into Google Cloud’s Gemini models. This has led to a reduction in average call times and a 96 per cent accuracy rate in agent assistance.
Praveen Atreya, VP technology development and planning at Verizon, told MWL while most of the operator’s agents are coming from vendors, it is also developing its own.
“The data that we collect goes into our unified orchestration and automation platforms that are fronted by these agents,” he said, noting frameworks were available featuring the ability to have its own agents with personalities specific for certain tasks.
“We are exploring custom agents we build ourselves and agents that come from our vendors.”
Lilac Ilan, global head of business development for telco operations for Nvidia, told MWL it is also important to make sure agents don’t talk to a competing vendor.
“Agentic AI is a system of reasoning where models are not just spitting out the answers. They’re thinking. They’re reasoning. They’re acting,” Ilan added.
With vendors such as Accenture, Amdocs, Salesforce and ServiceNow creating their own agents, Markus noted there’s often no need for operators to create their own, but it is essential operators use their own data for specific tasks.
Big tech moves
Recent examples of agentic AI being in the news include Google Cloud’s launch of a new marketplace and an agent-to-agent (A2A) protocol which allows communication across various telecom and enterprise environments.
The cloud giant is also partnering with Nvidia to bring agentic AI capabilities to enterprise customers using its Gemini models for on-premises deployments.
Google Cloud’s agentic AI initiatives are helping the industry create a common ecosystem where third-party agents work with each other or operator-developed agents, but there’s more to be done.
Markus stated he is in discussions with TM Forum to create agentic frameworks and agents which could be common for the telecom industry.
“There are some agents that are very custom and that are our IP, but there are other things, like the OSS space, that are best practices that we could share. We’re going to get together with the TM Forum folks and talk to them about that.”