I recently sat down with Dr. Sudhir Mittal, Executive Vice President and Director of Architecture at Jio Platforms, for a discussion around the dichotomy of AI in telecom, why cross-domain trust is the next frontier and the role data, procurement and governance need to play in the journey.
A few things stood out: silo-level autonomy already works, the models aren't what's holding cross-domain trust back and knowing when to trust an agent really comes down to how catastrophic the failure would be if it's wrong
These are the top takeaways from our discussion.
Sudhir laid out the scale of Jio's operation: 530-plus million customers and more than 240 exabytes of data moving across the mobile network. At that volume, accuracy, repeatability and so on, are not optional. An agent can't be trusted with decisions unless it can deliver these consistently.
There have been studies in the past that pegged AI pilot failure rates at above 80 percent. For a while these numbers hung over every AI conversation in the industry. But breakthroughs are finally happening. Earlier this year, Rakuten Mobile was recognized for running the world's first level 4 autonomous network for energy management, with Rakuten Symphony as the software vendor behind it. This is in production, with measurable energy savings secured within twelve months. Jio would have comparable case studies of its own on its roadmap.
Silo-level autonomy is already proven. Leaders like Jio and Rakuten Mobile, along with a handful of European operators, are running trustworthy, controlled level 4 implementations in narrow domains today. But a network isn't a silo. It generates billions of events, at every layer, continuously.
Getting to genuine autonomy means orchestrating across domains, in near real time, using AI. But you will still need to keep a human in the loop while doing it. Skip that step and all you've built is automation that stops at the edge of the RAN, or the access layer, or the core, or OSS, or BSS. Each domain can become smarter on its own, without giving the network the same level of intelligence across domains.
There are three areas that are holding this back. The good news is that the underlying capability of the models has nothing to do with it.
The problem starts with data. Every network has its own architecture and specific operating patterns. If a high-quality dataset is not engaged, you won't hit the confidence levels required for autonomous decisions, no matter how good the model is.
Next is procurement. Operator procurement models were built around ticket-based, T&M pricing. This doesn't hold up against the outcomes CXOs now want at board level. Besides productivity, operators also want new revenue streams. So, the buying model needs to evolve.
The third is organizational. The orchestration layer that sits across domains is still weak. This cannot be attributed to any single external vendor. It comes from operators and vendors building it together. I’moptimistic that there will be convergence at some point, and some common standards will become a part of routine operations.
Led by a team of AI experts, Rakuten Mobile has built a risk-mitigated, taxonomy-based framework to overcome this problem. Here is how it works, in a nutshell. When an agent hits a critical decision, the framework rates and scores it. It only lets it act on its own above a certain confidence threshold. The workflow is tested diligently before it's allowed near production. If you can build a codebase around actual human reactions and decisions, you can automate it.
The harder case is when three experienced people would give three different answers to the same question. That's not ready for an autonomous agent yet. In this scenario, the agent's job will be to recommend next steps, not act autonomously. The human in the loop makes the call.
The other consideration here is the cost of failure. It's fine to let the system get it wrong occasionally, but with caution. If you get something wrong on the BGP network, the blast radius reaches everyone downstream. Your agent needs to earn these kinds of decisions.
To achieve autonomy, controlled, siloed domains are a sensible start. Boundaries are known and cause and effect are understood. Failure scenarios can be modelled against an error database.
What makes this harder than it sounds is that the network is dynamic. Synthetic data can only approximate real-world situation. Meanwhile, every correction changes the network's behavior. Your network can behave in one way today and differently tomorrow. This means your autonomous system needs continuous validation. This is why governance is turning into a permanent fixture for the industry.
If an agent recommended a change to your network today, would you know how to validate that decision, quantify the risk, and define the acceptable blast radius? If not, the challenge may not be the AI. It may be the governance around it.
Where the conversation needs to move next year is toward a unified metric for the economics of AI in telecom. Telecom needs a framework that measures the economics of moving from single-domain autonomy to cross-domain orchestration. Getting there will be a win for the whole industry.
How are you defining trust, accountability, and autonomy in your networks? Join me at the AI Native Telco Forum, where I'll be discussing AI leadership and the weight of accountability. As AI agents take on greater responsibility, I'd be interested in hearing how you're tackling this question today.
