Agentic AI in telecom operations: wide adoption, narrow value so far

Telecom operators are not short of AI. They are short of AI that changes how work gets done. The 2026 evidence from PwC, NVIDIA, TM Forum and industry surveys shows high experimentation, rising budgets and genuine function-level returns, alongside thin enterprise-level impact and very limited use of agents that actually execute work.

The adoption–value gap

PwC’s 2026 AI performance study, which surveyed 1,217 senior executives across 25 sectors, finds telecoms better prepared than the average company. 93% of telecom respondents have run AI pilots, above the 88% reported by AI leaders, and they report a 20% median return on AI spending within functions, close to the leaders’ 25%¹. Scale is the problem. Only about a quarter of telecom companies say AI is scaled or embedded across major functions, against upwards of 40% for AI leaders, and telecoms attribute 14% of enterprise revenues to AI, against 40% for leaders¹.

Investors have noticed. BCG’s 2026 telecom value-creation analysis concludes that AI ‘has not triggered an investor rerating for telcos’ and that it sees ‘limited future benefits from isolated, cost-cutting AI deployments’⁷.

Agents are still the exception

The agentic end of the spectrum is where telecoms trail most. Only 18% say their most sophisticated AI use case can execute multiple tasks within guard rails, compared with 31% of AI leaders, and just 7% describe it as autonomous and self-optimising, against 15%¹. Practitioners are cautious for good reasons: in a TelecomTV survey of around 260 industry respondents, 57.5% said the complexities of telecom networks have not yet been properly built into agentic AI developments, and only 30% saw emerging agent protocols as a game changer today³. Looking across industries, Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls, and warns of ‘agent washing’ by vendors⁴.

Budgets are moving the other way. In NVIDIA’s 2026 survey, 89% of respondents expected their AI budget to increase in the next 12 months, up from 65% a year earlier, and 35% expected increases of more than 10%². The use cases most often cited for return on investment were autonomous networks (50%), customer service (41%) and internal process optimisation (33%)².

Where agents are starting to work

The most credible early patterns share three features: a bounded workflow, a clear outcome metric and a person who owns the result.

• Customer care. A TM Forum Moonshot Catalyst uses an orchestration layer that interprets customer intent and routes requests to specialist agents across care, sales and operations. It targets 30% improvements in first-contact resolution and average handling time and automation of more than 85% of standard queries⁵. These are targets rather than results, but they define what a care-agent programme should be measured on.
• Network operations. A multi-agent NOC Catalyst reports, on champion operator data, up to 60% reductions in mean time to detect and to repair and around 40% fewer incidents escalated to people⁶.
• Field service and revenue assurance. Both are rules-heavy and cross-domain: a truck roll or a billing discrepancy touches network, inventory, billing and workforce systems. This is where PwC argues agents add value, by coordinating work that siloed teams manage inconsistently¹.

Why scale is hard

PwC’s data locate the constraint. Only 46% of telecom executives say employees can quickly find and use high-quality data for AI, against 57% of AI leaders; 45% keep a single trusted record of critical data, against 59%; and 38% have reusable AI components centrally catalogued, against 51%¹. Agents amplify these gaps. An agent that repeatedly hands work back to a person is often exposing an inconsistent product catalogue, an incomplete customer record or an integration gap rather than a weak model¹.

Telecoms do have an asset most industries lack. 71% of telecom respondents say their organisation uses proprietary data for AI, compared with 60% of AI leaders and 40% of other companies¹. The problem is not the volume of data but its fragmentation across network, billing, care, field and product systems built for different eras.

Governing a digital workforce

The management gap is as wide as the data gap. Only 38% of telecom respondents conduct frequent portfolio reviews to prioritise AI initiatives, against 56% of AI leaders, and only 29% say AI has helped them significantly transform their business model, against 59%, the largest gap PwC measured¹. Gartner expects agentic AI to make at least 15% of day-to-day work decisions autonomously by 2028, up from none in 2024, and estimates that only about 130 of the thousands of vendors claiming agentic capabilities offer the real thing⁴. Both points argue for discipline: fewer, better-governed agents, bought or built against specific workflows.

Governance for agents should borrow from what operators already do for people and for network changes. Each agent needs an identity and entitlements that can be revoked; a documented remit and escalation path; approval thresholds that widen only on evidence; and logs that let a supervisor, auditor or regulator reconstruct what it did and why. TM Forum’s agentic Catalysts already reference security guidance for agentic AI in autonomous networks⁵, and operators will need equivalent internal standards before agents act on customer accounts or live network configuration.