Artificial intelligence in telecom used to be a thing carriers talked about. In 2026, it's a thing they're spending billions on, restructuring around, and in some cases bundling with consumer mobile plans. The substance has shifted sharply since the pre-ChatGPT era. Here's what's actually happening, where the real value is showing up, and what to make of the hype.
What changed since 2023
The earliest wave of AI in telecom — roughly 2018 to early 2023 — was about machine learning models doing specific narrow tasks: predicting network failures, segmenting customers for marketing, classifying support tickets. Useful but unspectacular. Most carriers had a few production deployments, scattered across departments.
Three things changed:
- Generative AI arrived in late 2022 and made customer service interactions actually useful. Rule-based chatbots that frustrated customers were replaced by LLM-based assistants that could resolve real queries
- The economic case sharpened. Telcos run on tight margins with massive workforces. AI offered a way to compress operational costs that pure 5G capex couldn't justify on its own
- Network architecture itself started to absorb AI. What used to be applied to the network is now being embedded into it — the AI-RAN movement, agentic operations, intent-based networking
By 2026 the story isn't "should we use AI?" It's "where does it actually pay off, and where is it just hype?"
Four layers where AI is applied
It helps to separate the conversation into the layers where AI is doing meaningful work. Each has different maturity, different economics, and different vendors:
- Customer experience — chatbots, virtual assistants, call summarisation, agent assist, churn prediction. Most mature; clearest near-term ROI
- Network operations (AIOps) — anomaly detection, root cause analysis, predictive maintenance, capacity planning, automated provisioning
- Network layer — traffic steering, energy optimisation, security and fraud detection, intent-based configuration
- RAN layer (AI-RAN) — AI embedded into the radio access network itself: spectrum optimisation, beamforming, baseband processing. Newest and most contested
The further down the stack you go, the longer the development cycles, the bigger the capex, and the more vendor-dependent the choices become.
Customer experience
This is where most retail customers are encountering telecom AI. The progression in the last 18 months has been clear:
- From rule-based to LLM-based chatbots. The old chatbots that asked "did that resolve your query?" after sending a help-page link have been replaced by generative AI that actually engages with the question. Resolution rates without human escalation are meaningfully higher
- Agent assist tools for human contact-centre staff. The AI listens to the call (or reads the chat), surfaces relevant policies and customer history, drafts response options, and writes the call summary. Average handling time drops; agent satisfaction usually improves alongside it
- Multimodal "digital humans" like Optus's Sally AI — video-based AI agents with full-face animation. Still novel; uncertain whether customers actually prefer them to text-based interfaces
- GenAI bundled with consumer plans. Optus became the first Australian telco to include a 12-month Perplexity Pro subscription with consumer and business mobile plans (June 2025) — turning AI tools into a differentiator at the plan level
Network operations and AIOps
AIOps is the unglamorous part of telecom AI that probably saves the most money. The use cases:
- Anomaly detection across millions of network elements, identifying issues before they cause customer-visible degradation
- Root cause analysis across multi-vendor environments — historically a manual investigation that took hours; AI tools now narrow it down to minutes
- Predictive maintenance for cell sites, transport equipment, and core network components. Site visits cost real money; predicting which sites need attention this week vs. this quarter changes the maintenance economics
- Capacity planning and traffic forecasting at the cell, region, and national level. Carriers have rich historical data for this; AI just makes the forecasting tighter
- Automated provisioning of customer services — enterprise circuits, mobile plan changes, network slices — reducing the time from order to live service
- Energy optimisation — cellular networks are large electricity consumers. AI-driven sleep modes for underutilised cells, dynamic power control, and smart cooling for data centres add up to material savings
For an Australian operator running tens of thousands of cell sites, even single-digit-percentage efficiency gains in operations translate to tens of millions of dollars annually.
AI-RAN and the radio layer
The most active and contested area in 2026 is AI-RAN — embedding AI directly into the radio access network rather than bolting it on top. The basic idea: instead of fixed algorithms doing spectrum allocation, beamforming, and Layer 1 baseband processing, learning systems that adapt to actual conditions.
Two things are pushing this forward:
- Vendor activity. The AI-RAN Alliance has grown to 132 members and showcased 33 AI-driven demonstrations at MWC 2026. The market is projected to grow from around USD 3 billion in 2025 to USD 36 billion by 2035 (28% CAGR)
- The Nokia–NVIDIA partnership — a USD 1 billion deal aligning Nokia's 5G-Advanced and 6G baseband software with NVIDIA's GPU platforms. Early field trials in 2026, with operators including T-Mobile and Indosat involved
Ericsson is taking a notably different approach. Where Nokia is treating the RAN as a distributed compute layer running on general-purpose GPU silicon, Ericsson is doubling down on custom ASICs and energy-efficient Layer 1 acceleration on specialised hardware. Both vendors can't be right; how this plays out will shape the 5G-Advanced and 6G era.
Open RAN is benefiting alongside this. Vodafone's "Spring 6" programme is rolling out one of Europe's largest Open RAN deployments — Wismar in Germany is set to become the first city fully equipped with Open RAN by Spring 2026, with Ericsson's Intelligent Automation Platform handling multi-vendor RAN management through AI-driven rApps for traffic steering, interference mitigation, and energy savings.
What Australian operators are doing
Both major Australian carriers have made AI explicit in their recent strategies, with notably different framings:
Telstra announced its "Connected Future 30" strategy in May 2025, positioning AI as central. CEO Vicki Brady told investors the company expects a smaller workforce by 2030 with AI driving "significant unlock" in efficiency. The numbers attached: over $2 billion annually in customer engagement savings, over $1 billion in software and IT development. Telstra has partnered with Accenture for delivery, and is taking a "build, then sell" approach — develop AI tools internally for telco operations, then offer them as products to other industries. Specific deployments include generative AI for call summarisation in customer service.
Optus has been more publicly visible with consumer-facing GenAI. Its AI Concierge chatbot, built on Google generative AI, reduced human-assisted enquiries by around 15%. Optus is using Google Cloud as a foundation-model provider alongside Anthropic for its broader AI strategy. The Sally AI multimodal digital human, developed with Tavus, is in production for customer interactions. And in June 2025, Optus became the first Australian telco to bundle a GenAI service — a 12-month Perplexity Pro subscription — with all consumer and business mobile plans.
The ACMA reported that operators including Telstra, Optus, TPG Telecom and Superloop have shifted from isolated AI use cases to integrating AI into existing systems and workflows — a meaningful maturity step. The Australian market is no longer experimenting; it's deploying.
International context
For perspective on where Australian operators sit, the international landscape:
- Deutsche Telekom + Google Cloud have moved from RAN Guardian agents to broader MINDR (machine intelligence for network decisioning) systems for service-level intelligence — widely cited as one of the more advanced agentic AI deployments in telecom
- NTT DOCOMO, BT, Elisa, Vodafone are deeply involved in Nokia's AI-RAN initiative
- Rakuten Mobile partnered with Intel to embed AI directly into the virtualised RAN stack — one of the more aggressive plays
- Indosat completed Southeast Asia's first AI-powered 5G call with Nokia and NVIDIA at MWC 2026
- SK Telecom + Samsung partnering on AI-RAN for 6G
Earlier-generation case studies that defined the conversation — AT&T's "Mia" virtual assistant for customer support, BT's "Assure Cyber" for network security — still operate, but the action has moved on. The frontier is now generative and agentic AI, not the rule-based ML systems of five years ago.
Agentic AI: 2026's theme
If 2024 was about generative AI for content and 2025 about generative AI for service interactions, 2026 is shaping up to be about agentic AI in telecom: AI systems that take actions, not just answer questions.
What that looks like in practice:
- An AI agent monitors a regional network, detects an emerging congestion pattern, decides to reroute traffic, executes the change, and logs the action — all without human approval at each step
- A customer service agent that takes the customer's intent, checks plan eligibility, applies the change to billing, and confirms back — in one interaction
- A field operations agent that schedules a technician, books the parts, and updates the customer — driven by a single trouble ticket
Rakuten Symphony has predicted that 2026 will be the breakout year for agentic AI in telcos, moving from pilots to scaled deployments. Ericsson has highlighted "agentic experience" — machine-to-machine APIs as a primary interface — as a fundamental ecosystem shift.
The risk is the obvious one: an autonomous system that takes wrong actions at scale costs more than the human decisions it replaced. The carriers furthest along on agentic deployments are also the ones investing heaviest in observability and rollback infrastructure.
Challenges that remain
The hype is loud; the obstacles are real:
- Data quality and access. Telecom data lives in dozens of legacy systems with inconsistent schemas. Useful AI needs that data unified and labelled, which is mostly unglamorous integration work that vendors don't talk about
- ROI scepticism. Two years into the GenAI era, many enterprises are still struggling to demonstrate returns on AI investments. Some analysts caution against premature workforce reductions based on projected efficiencies that haven't yet materialised
- Privacy and regulation. Telco data is sensitive. Customer call records, location, network usage — all regulated under the Privacy Act 1988 and sector-specific obligations. AI deployments need to be explicit about data handling
- Energy and infrastructure constraints. AI-RAN means AI inference at every cell site. The power and cooling requirements for that are not trivial; the industry is grappling with whether this scales economically
- Vendor lock-in. The Nokia–NVIDIA versus Ericsson custom-silicon split forces operator decisions that lock in specific paths for years. Open RAN is meant to mitigate this, but the multi-vendor reality is still messier than the marketing suggests
- Workforce transition. If AI is genuinely going to compress workforces by 10–20% over five years, the people-and-culture problem of getting there is real
What this means for the rest of us
For Australian businesses outside the telco sector, AI in telecom matters in a few specific ways:
- The plans you're being sold are changing. Bundled AI tools (Optus + Perplexity is the first; others will follow) are becoming a real differentiator at the consumer and small business level
- Your network experience is being optimised continuously. If you operate a fleet, an IoT deployment, or a fixed wireless service, the AI-driven traffic management and energy optimisation that carriers are deploying affects how reliably your equipment performs
- API access to network functions is opening up. Telstra is investing in network APIs through the Aduna joint venture (with Ericsson and other carriers); business customers can increasingly programme network behaviour rather than just consume connectivity
- The vendors you buy from are shifting. If you procure cellular routers, antennas, IoT devices, or coverage-improvement equipment, the products coming through in 2026 are being designed for AI-native networks. Specs that weren't meaningful five years ago — GPU acceleration, programmable interfaces, observability hooks — increasingly are
AI in telecom isn't a futurology question anymore. It's an operating reality that's reshaping the industry's economics, workforce, and product roadmap simultaneously. What we don't know yet is which approaches will prove durable and which are this cycle's NFV — a vendor narrative that quietly ages out.
If you'd like a hand thinking through how AI-driven networks might affect your specific deployment, or you're looking to source equipment built for the next generation of network operations, see Techwave Store or get in touch.
