An article by Rishi Shukla, Vice President & APAC Sales Head, Rakuten Symphony.

Most enterprises want to go autonomous these days — deploying software agents with a defined scope and the permission to act on their own within limits. This shift from automation to autonomy has quietly changed what a sales conversation with an operator looks like. Most vendors haven’t caught up to the details of that shift yet; many are still pitching autonomy at a high level, while few have gotten down to brass tacks. What’s interesting is that the questions from operators have gotten far more specific.

It helps to be precise about where the industry actually stands. The term “autonomous enterprise” gets thrown around loosely, so it’s worth anchoring to something concrete: TM Forum’s autonomous networks framework scores operators from Level 0 (fully manual) to Level 5 (fully self-directed). By that measure, most of the industry still sits between L1 and L3.

The jump that matters commercially is L3 to L4 — and it isn’t incremental. At L3, a system recognizes situations it’s been trained on and executes a pre-designed playbook. But what happens when the rules of the game change? At L4, the playbook stops being the point. Agents reason through new situations as they arise, build a plan, act on it, and adjust based on the outcome. That’s a fundamentally different kind of system.

Only a small number of operators have reached this level of autonomy, and only in a single domain so far. Increasingly, the question they’re asking isn’t whether an agent can follow the playbook — it’s how the agent performs when there isn’t one.

Agents are the reason L4 is suddenly on the table

What changed the ballgame is the arrival of agentic AI as an architecture. Even TM Forum has pointed to this: agentic AI is the key driver of the recent movement toward L4. There are now dozens of live agentic AI deployments across operators worldwide. With agentic AI, an orchestrator simply hands over cases to specialist agents. Optimization could be one case, while prediction or diagnosis could be another. There is a feedback loop for every job.

This is also why “autonomous enterprise” has stopped being solely a network engineering discussion. An agent doesn’t have to be pointed at the network per se; rather, it can be pointed at any workflow with rules and a meaningful business outcome.

That shift in scope is showing up in who’s asking the questions, too: these days, I am hearing commercial and business requests that sound a lot like the ones I used to only hear from network operations teams.

The APAC story

The enthusiasm for network autonomy is very real in APAC. In many ways, APAC is ahead of other regions when it comes to the usual technology-adoption curve. IDC research finds that 70% of APAC organizations expect agentic AI to disrupt their business models within the next eighteen months.

But is being open to AI the same as being ready for it? Recent surveys suggest otherwise. APAC organizations report the highest share of technical challenges in any region — 60% — and they’re prioritizing data-quality validation and internal efficiency above all else.

That resonates with what I have been hearing recently at various industry conferences and customer discussions. Conversations have moved past the back-and-forth over whether agentic AI is worth adopting. The real interest now is in knowing when an agent is about to get it wrong — before it acts. This is the governance conversation finally catching up — and it’s a healthier place for the discussion to be than a five-year executive roadmap.

Conclusion

Cost pressure is pushing some of the APAC accounts I work with towards a path that pursues L4 autonomy across RAN, core, and transport. Such an endeavor would typically take years and require AI maturity. Alternatively, what’s needed is a use-case-first approach that targets L4 autonomy in two or three high-value areas — say, energy optimization and root-cause analysis — within a year or so. (Aimed at operators who need a board win before going further).

As a seller, that changes a few things. It means I am rarely opening with a platform pitch anymore. I will be scoping one agent, doing a meaningful job, with some metrics. I am also prepared with a credible plan for what the second agent looks like with the first one in production.

Most operators, however, are still in the pilot stage of agentic AI deployment. Autonomous enterprise is very much a direction, and right now it’s mostly still the domain of pilots and single-domain wins. Enterprise-wide autonomy is on the distant horizon. Can an operator demonstrate an agent that is running in production? Can its security team sign off on it? For me, these are the more interesting tests over the next two years.

If you’re at FutureNet Asia in Singapore on September 22–23, come by Rakuten Symphony Booth 14 and join Sandeep Arora, Head of Global Sales, alongside leaders from True Corporation, StarHub, and XL Smart for “The Telco Reinvention: How AI Can Fuel Value Creation” on September 22 at 10:00 AM SGT. I’d be keen to continue the conversation there on the questions that matter most: what does it really take to move from AI pilots to agents running reliably in production?