Google just put an AI agent directly on the phone. At Google I/O in May, it launched Gemini Intelligence — an on-device layer that doesn’t just answer questions, it acts: reading the screen, moving between apps, and finishing tasks for the user.
For telecom operators, this is bigger than another OTT feature. It’s the customer relationship moving one more layer away from the network. Here’s what agentic AI on devices actually changes, and where operators still hold real leverage.
The short version
- Gemini Intelligence is an execution layer, not a chatbot. It completes multi-step tasks across apps using a new Android mechanism called AppFunctions.
- It runs hybrid: light inference on-device (Gemini Nano v3), heavy reasoning in the cloud. It needs a flagship chipset and ~12 GB RAM, so most of the installed base is excluded at launch.
- The risk: the agent captures commercial intent and pushes the network further toward bit-pipe status.
- The opportunity: agents need things only the network has — SIM identity, exposed APIs (CAMARA/NEF), guaranteed latency, fraud signals.
What Gemini Intelligence actually is
It’s an agent that does things, not one that talks. It reads what’s on your screen, navigates installed apps, and runs multi-step tasks in the background — pulling a reference from your email, building a cart, holding a booking, then waiting for you to approve.
The piece operators should note is AppFunctions: a new Android mechanism that lets apps expose their capabilities to the agent, so it can call them directly instead of tapping through a UI. That turns the agent into the front door for shopping, search and booking.
It’s hybrid by design. Latency-sensitive inference runs on the device on Gemini Nano v3, while heavier reasoning goes to the cloud. That’s why the hardware bar is high, with the rollout starting on the newest Galaxy and Pixel devices before reaching watches, cars and glasses.
Why operators should care
Two things change for a connectivity business.
The customer relationship. Once an agent books, shops and arranges on the user’s behalf, it sits between the customer and every service they reach — including the operator’s app, portal and upsell. The moment of decision moves to the agent layer. This is the same disintermediation we saw with OTT players, one level deeper.
The data. An agent that reads your inbox, calendar and purchase history builds the richest behavioural profile on the device. That value flows to the OS vendor, not the network carrying the traffic.
Where the network still wins
The high hardware bar is also an opening. A service that needs a flagship handset won’t reach the prepaid base or most emerging markets for a while, and operators control how that base upgrades. Four levers stay in our hands:
- Network APIs. Identity, location, quality-on-demand and fraud signals exposed through CAMARA and the NEF are agent inputs that live only on the network side. Build them as products.
- Trust and security. An agent that can act across your apps and your bank is a new risk class. The operator brand and the SIM as a hardware root of identity are positions the OS vendor can’t occupy as cleanly.
- Device distribution. If the premium experience is gated to flagship hardware, the operators financing those devices have a negotiation — visibility, preload terms, revenue share — not a done deal.
- Packaging. Network-grade AI services, secure agent connectivity and identity-as-a-service are billable and differentiated because they sit at the network layer.
The bigger shift: agentic traffic on the network
An always-on agent running background tasks creates a new traffic profile: lots of small, frequent, signalling-heavy, uplink-leaning sessions. That lands on the control plane as much as on throughput, and it’s exactly the AI-driven uplink demand the industry has been debating.
The strategic question isn’t how to beat Gemini on the handset. It’s how to make the network the place agents must come for what only the network provides — and how to plan Smart CAPEX for a traffic mix that looks less like video streaming and more like a swarm of small autonomous clients. That’s the same direction 6G is already taking in 3GPP and the O-RAN Alliance.
Frequently asked questions
What is agentic AI on devices?
It’s AI that runs on the phone and takes actions on the user’s behalf — completing multi-step tasks across apps — rather than only answering questions. Gemini Intelligence is the first large-scale example.
Why is Gemini Intelligence a concern for telecom operators?
It captures commercial intent at the agent layer and deepens disintermediation, pushing the operator further toward bit-pipe status while the OS vendor owns the customer data and the point of decision.
How can operators respond to agentic AI?
By exposing network capabilities (identity, location, quality-on-demand) through CAMARA and the NEF, positioning as the trusted and secure layer, using device distribution as negotiating leverage, and planning CAPEX for agentic traffic.
Does agentic AI increase network traffic?
Yes, but the shape matters more than the volume. Background agent activity is signalling-heavy and uplink-leaning, which stresses the control plane and the uplink rather than raw downlink throughput.
Go deeper on network monetization. Network API monetization, CAMARA, NEF and AI-native network strategy explained for operators and enterprises.
Benefit from Massive discount on our 5G Training with 5WorldPro.com
Start your 5G journey and obtain 5G certification
contact us: contact@5GWorldPro.com
