- Telcos are moving fast on agentic AI, but orchestration needs to be part of the equation
- Governance and traceability are emerging as must-haves
- Agentic AI will force telcos to rethink automation, shifting from pre-programmed workflows to real-time governance, monitoring and accountability
Telecom operators hustling to deploy agentic AI may soon find that the hard part isn’t building and deploying agents, but ensuring they don’t work at cross-purposes across network, customer care and IT environments. The key to doing this is shifting focus from mere automation to orchestration, Amdocs Group President of Technology and Head of Strategy Anthony Goonetilleke told Fierce.
Goonetilleke explained that AI agents are like football or soccer players – each is highly skilled in its own right, but they must work together to win the game. Getting them to do that, though, is easier said than done. That’s because AI is introducing a slew of reasoning-based systems in an industry that has spent decades engineering deterministic control planes for networks, he said.
“Everyone is talking about building agents. Not enough people are talking about controlling them,” Goonetilleke said. Indeed, Gartner recently found that AI governance is lagging AI adoption, with many organizations treating governance as something to tackle after deployment rather than before. A separate study by EY found that even those who have a governance strategy aren't sure they can design, implement, evolve it effectively. But that’s a risky recipe.
Gartner predicted that by 2027, 40% of enterprises will pull autonomous AI agents offline due to governance gaps identified after production incidents occur.
“You could have one agent trying to optimize network performance, another trying to reduce costs and another focused on improving customer experience in their respective domains,” Goonetilleke said. “Each may do its job correctly, but they may not harmonize and achieve as much for the business as they should, and the objectives could vary depending on the network service being provisioned or assured.”
To get all the agents pulling the cart in the same direction, telcos must build a new orchestration layer capable of steering them. And to steer, of course, you need a harness.
“More and more, the AI harness has turned into not only a key differentiator in great AI implementations, but also the work horse for managing model gardens, component orchestration and adhering to key guardrails. Once you get into Loop and Graph engineering, a mission-critical harness starts to become the framework form your AI blueprint,” Goonetilleke said. “If agents become a new workforce inside the operator, the AI harness is effectively the combination of their security perimeter, operating policy and management system.”
In addition to a cart and harness, operators need to build the road AI agents are allowed to travel on. That’s where a robust governance framework comes into play. Goonetilleke said it’s critical that operators explicitly define what agents are authorized to do, the same way they have protocols that outline which humans can query, change, add or delete key network elements.
Highways also come with a way to police whether the rules of the road are being followed. For AI, that’s where traceability comes in. Operators should be able to trace the agent identity, assigned task, relevant model and tooling, authorization and result for every consequential action an AI agent takes, Goonetilleke said.
Much of this work will likely be outside of operators’ comfort zones, though. Traditional orchestration and automation were deterministic and rules based. But with agentic AI, “things are more dynamic - they can reason, adapt and interact with other agents. The challenge becomes coordinating intelligence on the fly rather than tasks in sequence.”
As Gartner noted, different agents also require different levels of governance, complicating matters even further.
“The old mindset was 'define every scenario in advance.' That's not realistic in an agentic environment. Instead, operators need to focus on governance, monitoring and accountability rather than trying to pre-program every outcome and let their models do their work,” Goonetilleke concluded.