Agentic AI in telecoms: from answering to acting

The industry’s AI conversation has changed its verb. For two years the question was what AI could say: summarise this ticket, draft that answer. The question now is what AI can do. The GSMA calls this agentic AI — systems that understand intent, make autonomous decisions and execute complex tasks — and has made it a central theme of its Mobile AI Community, describing an industry moving from “AI for networks” to genuinely AI-native infrastructure.

Generative AI answers. Agentic AI acts. For an industry built on operations, that is the difference that matters.

What agentic AI means for telecoms

It helps to separate three things the industry keeps blurring. Traditional AI follows rules it was given: route this ticket, score that fraud pattern. Generative AI produces content: a draft answer, a summary, some code. Agentic AI holds a goal, watches the world change, and acts — and keeps acting until the job is done. Telecoms has had the first for decades and adopted the second in a rush; the third is the one shaped like the industry itself.

The GSMA’s own example is a good one: a network detects a drop in service quality, reroutes traffic proactively, notifies the affected user in real time and dispatches a virtual agent to resolve the issue — with no human in the chain. Nothing in that sequence is a chatbot. It is perception, decision and action, joined up. The same shape repeats across the industry’s early use cases: self-healing networks, maintenance scheduled before the fault, SIM-swap fraud caught mid-attempt rather than post-mortem.

And the pattern generalises far beyond the radio network. Everything a mobile business does — orders, activations, payments, ports, usage, support — is a stream of events that something has to notice, decide about and act on. Today that something is mostly a person with a queue. Agentic AI is the claim that it no longer has to be.

The industry is organising around it

This is not a vendor slogan; the industry body is building scaffolding for it. The GSMA’s Mobile AI Community runs three member-led workstreams: pilot projects and deployment blueprints that move agentic AI from experiment to production; an end-to-end architecture for telco agent-as-a-service (TAaaS), so agents from different vendors and operators can interoperate; and a technical security and governance track building a threat-modelling playbook for risks with new names, like boundary collapse and adversarial agent injection.

Read the workstreams together and the message is clear: the industry expects agents to become infrastructure, and it is preparing for the day they need the same interoperability and security discipline as the network itself. The GSMA’s whitepaper, Agentic AI for Telecom: Charting the Course for an Intelligent Future, is the best single summary of that ambition.

The next customer is an agent

There is a quieter shift underneath the use cases, and Ericsson’s researchers put it well: agents are becoming the primary users of telecom capabilities. A decade of investment went into developer experience — documentation, sandboxes, portals. The next consumer of your APIs will not read any of it. It will be an autonomous system that discovers a capability through a registry, reasons about cost, fit and permission, and calls it at machine speed. The design question stops being “can a developer integrate this?” and becomes “can an agent be safely delegated this?”

This is why we built Mindszi API-first and put an MCP server on top of the platform: the same scoped, audited tools our customers’ developers use are the tools their agents — and AI clients like Claude — call. When the primary user is an agent, being machine-callable with governance attached is not a feature. It is the product.

What it looks like when it runs a business

Most of the industry conversation starts at the network layer. We started at the other end: the business layer, where an MVNO or eSIM brand lives. The pattern is identical. A platform event fires — a payment fails, usage crosses a threshold, a port sticks. An agent picks it up, reads the real state of the customer and the service, acts through scoped tools, and reports back where the team already works.

• A failed payment is retried and the customer messaged before anyone opens a ticket.
• A customer near their data limit gets a one-tap top-up in their own language.
• A stuck port is chased with the losing carrier until it completes.
• The morning report lands in Slack: what ran, what changed, what needs a human.

That is agentic AI in telecoms at the commercial layer, and it is what Mindszi Agents productise: describe the job in plain words, and the builder drafts the agent with its tools, knowledge, triggers and approval points. No code, never YAML.

Governance decides who gets to scale

The GSMA is right to put security and governance in the founding workstreams rather than the appendix. Autonomy without control is a liability in a regulated industry. Our version of that discipline is boring on purpose: every agent runs with scoped tools and nothing more, anything sensitive pauses for a human yes, and every run — each step, tool call and decision — lands in the audit trail.

The operators who win with agentic AI will not be the ones with the most autonomous demos. They will be the ones whose agents are governed well enough to be trusted with production. That is the course the industry is charting, and the platform we have already built.