- STL Partners’ count of agentic AI announcements at MWC 2026 found only two live deployments acting on sub-second data, both customer-facing fraud detection, while no live network operations deployment yet depends on sub-second data.
- Closing that gap takes live event data thorough enough to catch a genuine anomaly and trusted context to interpret it, which is why a hybrid architecture combining central platforms with distributed, edge-local processing is emerging.
- An AI recommendation is an input, not a decision. Something still has to evaluate it against current network state, apply policy, commit the change, and record why, before it becomes an action.
- In a joint Volt and Ocient demo, Volt evaluates every network slice against its SLA once a second and acts on its own, throttling traffic, steering around congestion, reallocating bandwidth, or refusing admissions, with each decision recorded as a single ACID transaction.
- Ocient owns the model, retraining continuously against full-fidelity telemetry; Volt runs that model live against real-time state and enforces the outcome. Neither system is built to do the other’s job.
TL;DR
In June, in the middle of the World Cup, Jai Rajaraman of Ocient and I presented a webinar with STL Partners. We built it around a fan who commutes to work and has paid for a streaming subscription to watch a match. They don’t care that a few thousand people nearby have had the same idea at the same time. They’ve paid for a quality of service and expect to get it. For the network to protect that on its own, it has to spot the problem, work out what it means, and act on it, while the match is still being played.
STL has now published the research behind that session: eleven interviews across Tier 1 and Tier 2 operators, NEPs, systems integrators, and BSS/OSS providers. The finding I keep coming back to is their count of agentic AI announcements at MWC 2026. Two live deployments acted on sub-second data, and both were customer-facing fraud detection. Not one live network operations deployment depends on sub-second data. The trials mostly do. That gap between what is being trialed and what is actually running is the thing worth explaining.
Most of the explanation is about data. An agent needs live event data that is thorough and timely, and it needs trusted context to interpret it. Telco data has historically been sampled and aggregated to keep transport and storage affordable. A five-minute average or a 1-in-100 sample is fine for a dashboard and useless for telling a genuine anomaly from noise. Shipping every event to a central platform before anything is reasoned over it then adds delay of its own. STL concludes that a hybrid architecture is emerging, keeping central platforms for analytics, governance, and long-term training while processing high-volume event streams close to where they are generated.
I would add one thing from what I see in live networks. Reasoning is not deciding. An AI recommendation is an input. Before it becomes an action, something has to evaluate it against the current state of the network, apply policy, commit the change, and record what was decided and why. A decision that arrives after the moment has passed could be worse than no decision at all. That layer is usually the one missing from the architecture diagram.
It is easier to show than to argue about, which is why Volt and Ocient built a joint solution demo. A sample set of cell sites carries five slices, one of which is video. Volt evaluates every slice against its SLA once a second and takes corrective action on its own: throttling best-effort traffic, steering traffic off a congested cell, reallocating bandwidth, or refusing new admissions. Part of that same evaluation is Volt executing an ML model against the live event, which flags a cell that is degrading before the SLA actually breaks, so traffic can be moved early. Whether rules or a model, each decision is a single ACID transaction with an audit trail, so there is a record of what was decided and why.
Ocient owns the model. It holds full-fidelity telemetry at petabyte scale, retrains continuously against that record, and promotes updated models into Volt, where the live scorer picks up a new version on the fly and keeps running. The pattern a live event gets judged against comes from everything the network has seen, not a sample of it, and it stays current as the network changes. Volt handles the current, permitted state. Ocient handles what it means. Neither is built to do the other’s job.
Jai and I will keep making this argument in public until it is boring. The report is worth your time in the meantime, particularly the worked RAN example near the end.
Download the Report: Real-time Data for Autonomous Networking (STL Partners, August 2026)
Book a joint Volt+Ocient demo.
This report was prepared by STL Partners under commission from Volt Active Data and Ocient. STL Partners maintains strict editorial independence.
What is the gap between agentic AI trials and live telecom network deployments?
It’s the difference between what operators are testing and what’s actually running in production. STL Partners’ research on agentic AI announcements at MWC 2026 found two live deployments acting on sub-second data, both customer-facing fraud detection, while no live network operations deployment yet depends on sub-second data. Most trials do use sub-second data, which means the gap sits specifically at the point of going live.
Why does live network event data need to be thorough, not just fast?
Network data is traditionally sampled or aggregated to manage volume and cost, and a five-minute average or a 1-in-100 sample is enough to spot a trend on a dashboard. It hides the individual spikes, retries, and correlated failures an agent needs to distinguish a genuine anomaly from noise, so a partial data picture can cause an agent to miss the moment a response is actually needed.
What's the difference between an AI agent reasoning about a network problem and deciding to act on it?
Reasoning produces a recommendation. Deciding requires evaluating that recommendation against the current state of the network, applying policy, committing the change, and recording what was decided and why, all before the moment it addresses has passed. An agent’s output is an input to that process, not a substitute for it.
What architecture do telecom operators need to support real-time agentic decisions?
A hybrid architecture that keeps central platforms for analytics, governance, and long-term model training, while processing high-volume event streams close to where they’re generated for low-latency action. Shipping every event to a central platform before anything reasons over it adds delay that a live decision, like protecting a video stream’s SLA, can’t absorb.
How do Volt and Ocient work together in real-time network decisioning?
Ocient holds full-fidelity telemetry at petabyte scale and continuously retrains the model that flags degrading network conditions, then promotes updated versions to Volt. Volt runs that model live against real-time state, evaluating every network slice against its SLA once a second and enforcing the outcome, whether that’s throttling traffic, steering around congestion, reallocating bandwidth, or refusing new admissions.
Why does a network decision need to be recorded as an ACID transaction?
An ACID transaction guarantees that the decision, whether it came from a rule or a model, is applied completely and recorded atomically, with no partial updates. That gives operators a single audit trail of what was decided and why, which matters both for troubleshooting and for building trust in decisions an AI agent influenced.
Can an AI agent be trusted to change live network configuration without human review?
Not on the strength of its reasoning alone. What makes that trustworthy is the layer underneath the agent: one that evaluates the recommendation against live, current state, applies policy, and enforces the outcome as an auditable transaction. Without that layer, a fast recommendation is still just a recommendation.
About Author
Featured Resources
