Agentic AI for telecoms: the agent is just the beginning

Agentic AI promises to transform the telecoms industry by automating actions across customer, commercial and network operations. By acting on both sides of the transaction, agents could fundamentally reshape digital commerce and how telecom services are bought and sold. What do CSPs need beyond the agent itself to turn that potential into reliable, commercially valuable services?

With early applications such as email summaries and chatbots now familiar, attention is turning to AI agents that can act autonomously.

These agents can pursue an objective, reason about how to achieve it, interact with other systems and take actions without requiring explicit human instruction at every step.

For telecom operators, an agent could identify why a customer’s broadband service has deteriorated, correlate the problem with a network fault, arrange remediation, communicate with the customer and potentially apply compensation.

And the opportunity extends far beyond customer service. Agents could optimise networks, manage orders and negotiate with other agents acting on behalf of customers or other suppliers. In time, they could even become network users and customers themselves.

But an agent’s intelligence alone will not determine its value. Just as important is everything surrounding it: the context it can access, the systems it can control, the other agents it can communicate with, the permissions governing its actions and the cost of all that reasoning.

Building the agent is one thing. Creating the environment in which thousands of agents can operate safely, economically and reliably could be the much bigger challenge.

Telecom operations involves many situations that don’t fit neatly into traditional automation.

A billing query may originate with an incorrectly configured product. A customer complaint may be the first indication of a network problem. Resolving an enterprise service issue might require information from CRM, billing, order management, service assurance and network systems.

An agent might need to reason across these boundaries and take action rather than simply following a predefined workflow.

The models themselves, however, are unlikely to provide CSPs with much lasting differentiation. Foundation models are widely available and advancing rapidly. Instead, operators need to consider the assets they possess that others don’t.

A generic model doesn’t inherently understand that a customer has three services, recently changed tariff, has an outstanding order and is experiencing degraded network performance, but the operator does.

That context sits within the BSS/OSS applications, in customer records, product catalogues, orders, network inventory and service assurance, and this is where operators can differentiate.

Consider a service outage. A network agent could identify anomalous behaviour, a service assurance agent determine which services are affected, and a customer service agent identify impacted subscribers. A billing agent could then determine whether compensation is appropriate.

As TM Forum’s Andy Tiller argues in The Fast Mode, autonomous telecom operations cannot be built from isolated AI agents: “siloed AI point-solutions duplicate logic, compromise human control, limit scalability and build technical debt.”

Agents therefore need to work together across customer, service and network domains, using shared context and consistent definitions of customers, products and services. Otherwise, operators risk recreating the same silos they have spent decades trying to dismantle.

Imagine every department deploying its own agents. Before long, hundreds of agents could be independently retrieving information, reasoning about problems and invoking business processes. So which has the authority? How do they discover one another? What happens when two reach contradictory conclusions?

The industry needs an architecture through which intelligence can operate and interoperate. This is where work already underway on TM Forum’s Open Digital Architecture (ODA) and AI-native operations becomes increasingly significant.

There’s a significant difference between allowing an AI agent to retrieve a customer’s bill and allowing it to issue a £500 credit. Likewise, recommending a network configuration change is very different from making one automatically.

Recent events involving OpenAI and Hugging Face demonstrate why that distinction matters. An incident involving AI agents being used for security testing reportedly resulted in actions extending beyond their intended scope and compromising Hugging Face’s systems, subsequently raising questions about who is responsible when increasingly autonomous AI systems take unintended actions.

The problem isn’t that AI agents are inherently uncontrollable. It’s that connecting probabilistic AI systems to live operational platforms increases the consequences of an incorrect action.

A hallucinating chatbot might give a customer the wrong answer, but an agent with sufficient privileges could incorrectly act on this.

• What information can an agent access?
• Which tools can it invoke?
• How much money can it spend?
• Which decisions require human approval?

For Cisco President and Chief Product Officer Jeetu Patel, the emergence of agents means “we have to move from just plain access control to action control”. Authenticating an agent is no longer sufficient; operators also need to determine which actions that particular agent is authorised to perform.

The practical future may therefore be less about unlimited autonomy and more about bounded autonomy.

A customer service agent might automatically issue a £5 goodwill credit while a £500 adjustment requires secondary approval. A network agent might diagnose a fault but require validation before making consequential configuration changes.

Human involvement doesn’t necessarily mean autonomy has failed. Review, audit, rollback and approval can be deliberate features of a reliable agentic architecture.

There is also a temptation to assume that rapidly falling AI token prices will make the economics of agentic AI increasingly trivial.

Gartner thinks the opposite could happen.

It predicts that AI inference costs per agentic workflow will increase more than fivefold by 2028 – an illustration of the “inference paradox”, in which cheaper models and tokens encourage developers to build more sophisticated applications that consume much more inference overall.

A chatbot might interpret one request and produce one response. An agent may reason through multiple steps, call tools, evaluate results, communicate with other agents and repeat the process several times before completing its task.

In fact, Gartner says routing a task to an agentic reasoning model already increases provider inference costs by at least five times compared with a basic chatbot interaction.

And inference isn’t the only resource agents consume.

In one Cisco test, an Open Deep Research agent explored 44 web sources and generated 26.8 MB of traffic – 450% more than performing the task manually. Around 70% of the incremental traffic was associated with AI inference. This is one workload rather than a universal multiplier, but it shows why operators need to assess the full resource cost of an agentic workflow, including compute and connectivity. As Cisco’s Jeetu Patel puts it, “Humans click, but agents swarm.”

This makes efficient orchestration essential. Deterministic tasks can remain with conventional software; straightforward AI requests can use smaller models; and expensive reasoning can be reserved for problems that genuinely require it.

So far, we’ve discussed agents working for the operator. But what happens when they work for the customer?

Instead of visiting several websites to compare tariffs, a consumer might ask an agent to find and purchase a rolling mobile plan with at least 50GB of data, EU roaming and a maximum price of £20 per month.

But agents may become more than intermediaries buying conventional telecom products – they could become network users in their own right.

Ericsson characterises this as a move from developer experience to agentic experience: rather than a developer manually selecting and integrating a network API, an agent could discover appropriate capabilities itself and evaluate factors including latency, reliability, price and compliance.

A video agent, for example, might determine that it needs a particular quality of service and request the corresponding network capability.

Ericsson describes this evolution as moving from machine-readable capabilities to “machine-delegable” ones. The agent needs to understand not only what a capability does, but whether it is authorised to invoke it, under what constraints and who bears the economic responsibility.

CSPs may increasingly need to serve both human customers and agents that discover, invoke and pay for network capabilities on behalf of people or businesses.

Taking that another step, the transaction itself could in the future occur between machines.

An enterprise’s procurement agent might request connectivity meeting particular requirements. The CSP’s sales agent would identify eligible products and provide a quotation. The procurement agent might then ask for better terms, prompting the sales agent to check its commercial rules before responding.

But their objectives may differ. The enterprise’s agent could be minimising cost while the operator’s agent seeks to maximise conversion, margin or lifetime value.

Digital commerce could consequently become less about persuading a person navigating a user interface and more about making products, prices, eligibility rules and ordering capabilities discoverable and understandable to machines.

But however the agents negotiate, the resulting transaction must still satisfy deterministic business rules. A product must have defined characteristics. An order must be valid. Usage must be charged correctly. Charges must be billed accurately and revenue recognised correctly.

Agentic AI could simultaneously reduce the role of the telco’s online portals while making its underlying digital capabilities more critical.

An agent doesn’t eliminate the need for domain knowledge embedded in the systems behind it. Products, eligibility rules, charging models, order dependencies and fulfilment processes don’t disappear because the customer no longer interacts directly with the application where they are implemented.

This is why letting AI agents loose doesn’t make BSS/OSS applications obsolete. If anything, these systems of record become even more important.

Agents still need authoritative customer and product data, as well as deterministic business rules and trusted systems through which their decisions can be executed. Accurate product catalogues, common information models, Open APIs and controlled access to operational capabilities consequently become part of the customer experience – even if the customer never sees them.

Operators may also have to rethink what constitutes a digital channel; today, the customer comes to the operator’s website or app, but tomorrow, the operator’s products and capabilities may need to go wherever the customer’s agent happens to be.

For CSPs, the practical starting point is a workflow with a clear business outcome, trusted data, explicit action limits and a measurable cost per completed task. The same foundations will matter when the agent belongs to the customer. The agent is just the beginning; the real test is whether the surrounding systems can turn its decisions into reliable business outcomes.

Discover how Cerillion AI Agents are bringing agentic AI into the heart of BSS/OSS, helping CSPs automate complex processes, access trusted operational context and take action across customer and business workflows.