Agentic AI in Telecom: From Pilot to Production
- 10Pearls Editorial Team
- 7 min read
Summary
In this blog, we discuss agentic AI in telecom, how it differs from classic RPA and more recent generative AI, why telcos are leaning towards AI agents, and the most viable use cases. We also go over the challenges AI pilots face in reaching production and how agentic AI actually fits within telecom operations.
Agentic AI is expected to create roughly $150 billion in value for the telecom sector across a wide range of operational and consumer-facing domains, with the latter taking the lead. Agentic AI adoption in customer engagement is already around 60% and expected to reach 90% by 2027. But adoption also entails experimentation, and like many other industries, the production-stage adoption of agentic AI in telecom is still lagging. Most AI agents in telecom may never touch a live customer conversation or a production order. The models themselves are capable enough, but integration, AI security, governance, and a wide range of operational challenges prevent agentic AI pilots from evolving to production-grade deployments.
In this blog, we will cover what agentic AI actually is in telecom, important use cases, and reasons why many agentic AI pilots fail to reach production.
What is agentic AI in telecom?
Agentic AI in telecom refers to the use of AI systems that can interpret goals and
intentions, plan their own sequence of steps, and act directly within OSS, BSS, and
network systems to complete tasks like diagnosing network faults or mitigating fraud,
with limited human intervention.
The three defining characteristics of AI agents in telecom are:
- Goal planning: Agentic systems interpret goals that are assigned to them, within the operational context and constraints. They decide what actions to take and the sequence of these actions to complete complex, multi-step tasks.
- Tool access: The tool and API access allow agents to interact with real, operational systems. How this access is scoped with the right level of permissions is critical to agentic governance and success.
- Escalations: AI agents must determine when and how they must stop acting autonomously and escalate an issue to the relevant individuals for human review and actions.
Agentic AI in the telecom industry faces unique challenges and constraints, including the complexity and sensitivity of operational systems like OSS, BSS, and CRM. Latency is another limitation. Decisions like network slicing are usually taken faster than agents can deliberate (for now), and deterministic automation might still be the best approach in such scenarios.
RPA vs. GenAI vs. Agentic AI
| Rule-based automation | Generative AI | Agentic AI | |
|---|---|---|---|
| How it decides | It doesn't. It simply follows instructions defined in advance. | Interprets and follows the instructions given in a prompt. | Interpret intentions and plan its own steps to reach a goal. |
| What it does well | Repeats a known, well-understood process at scale. | Summarizes, drafts, explains, classifies, and communicates. | Handles situations where there isn't a clear path to success. |
| What it cannot do | Handle unexpected situations that weren't anticipated and addressed in its instructions. | Complete a multi-step task on its own, without human direction at each step. | Guarantee the same steps and paths every time, because at each run, the agent(s) may interpret and act differently. |
| Who is accountable | The individuals who own the process and wrote instructions. | The individuals responsible for validating its output. | People who design and deploy the agents and who have oversight on their actions. |
| Telecom examples | Automated billing and threshold-based alerts. | Conversational chatbots. Writing and summarizing incident logs. | Optimizing network operations, fault detection, and selective escalations. |
| When to use | The process is stable and well understood. | Tasks related to content analysis and generation. | For processes where the path to success varies and humans currently improvise. |
Why telcos are moving on agentic AI now
The simplest answer to why agentic AI in the telecom industry is gaining traction now is because the foundation models (with agentic capabilities), and agentic tooling and frameworks have evolved and matured significantly. While generative AI telecom solutions resulted in significant transformation, it was limited to certain domains and operational efficiency avenues, both internal and customer-facing. Agentic AI’s ability to act across different telecom systems enables AI-driven automations that were previously impossible through RPA and even by adopting generative AI.
Other factors include flat Average Revenue Per User (ARPU) in a saturated market. The profits from new service categories and even potential new markets like IoT are often flattened against the cost of rolling out relevant technology infrastructure like 5G. Then there is the massive operational cost of manual network assurance. Because of the unpredictable and highly dynamic nature of the task, traditional RPA and even generative AI (relying on human directions) were unable to automate it. However, agentic AI has a realistic chance of handling this major, previously unsolved aspect of telecom network automation.
This is not limited to network assurance. Several telecom network operational domains or even granular tasks that were unfeasible for rule-based automation or even with generative AI may benefit from agentic AI.
How agentic AI can help telcos move up the TM Forum Autonomous Networks levels
The TMForum autonomous networks framework gauges how much of an operational loop is running without human intervention. While several last generation IT telecom solutions have allowed operators to achieve partial autonomy in specific, well-understood scenarios, agentic
AI may help them advance their capabilities further.
Manual
Manual executions throughout, with tooling only helping in limited capacity.
Assisted
Repetitive executions handled with scripts and tools, with humans taking care of sensing, analyzing, and deciding.
Partial autonomy
Closed-loop automation of specific scenarios within a specific domain, through statically configured rules.
Conditional autonomy
System understands the environment, handles routine decisions, and dynamically adjusts rules for automation optimization within a domain. Humans handle exceptions.
High autonomy
Automation jumps across domains, with predictive analysis enabling autonomous decision-making. Humans provide policy and intent.
Full autonomy
Everything is fully automated, with systems adapting to changing needs autonomously; no human intervention required. This is an aspirational stage.
Agentic AI use cases in telecom
Agentic AI use cases in telecom are an expanding set. As the technology matures, confidence in the technology improves, and early agentic deployments start producing measurable results, telcos will naturally start looking into more use cases. For now, the use cases offering strong evidence for agent-driven improvements include:
Network operations & autonomous networks
AI agents can triage the backlog of faults, often with more contextual awareness, speed, and accuracy than their human counterparts. This leads to fewer unnecessary escalations. They can also ensure that whatever alerts are routed to engineers are accompanied by the relevant supporting evidence, enabling easier resolution. They can also leverage pattern recognition to identify root causes that trigger multiple faults.
Specialized AI agents can be used to identify network issues and recommend RAN optimizations by going over performance data in real-time (or execute these recommendations if they have the permissions). Orchestrating planned work like maintenance shutdowns is typically a coordination- and communication-heavy task, and since agents can adapt to changes in run-time (like a shutdown stretching beyond planned hours), they can minimize disruption and keep impacted parties informed.
Customer care
Customer care is one of the most mature and promising agentic AI use cases in telecom, enabling productivity gains of around 45% to 60%. According to a Google report, customer service and experience is the third-largest use case for agentic AI adoption in telecom, with 45% of organizations reporting agentic deployments in this area.
AI agents can transform Tier 1 resolution in telecom by not just responding to customer complaints but actually reaching across systems to fetch relevant information and escalating issues to the relevant teams. Other areas where agentic AI shows promising results are handling billing disputes, making retention-related offers and incentives to high-risk customers, and voice agents, which benefit from both agentic capabilities and improved voice-based conversations, reducing load on human agents.
BSS, OSS & revenue assurance
Order fallout is one of the strongest fits for agentic deployments in telecom, since agents can treat each one on a case-by-case basis. They can diagnose failures across OSS and BSS and escalate ones they can’t resolve on their own, reducing the manual burden. The same capabilities apply upstream to provisioning exceptions.
AI agents can also help with revenue assurance by identifying leakages through continuous monitoring of network logs, billing systems, and usage records and identifying anomalies early. They can also run root-cause analysis and patch certain leaks without human intervention (within their defined limits) or escalate with recommendations for faster manual remediation. Number portability is another critical use case for AI agents, since it requires validating and coordinating against rules and handling exceptions.
Field service
While AI systems already handle dispatch optimization, AI agents can be more resilient
against changes in real-time and adjust the original dispatch plans to accommodate new variables like technician availability, parts unavailability, etc. Agents can also help field technicians retrieve site/issue-specific records to identify common fault patterns and
previous resolution approaches.
Both agent-driven field service benefits may contribute to increased first-time resolutions, which is the single largest controllable cost in field operations. Many repeat visits can be traced back to incomplete diagnosis before dispatch, something that AI agents can improve upon.
Fraud, security & compliance
Fraud detection is already a mature analytics problem, but there are still manual elements, including correlating different signals, verifying whether a fraud pattern is genuine, and containing the damage. Agents can take over many tasks between alert and action. One great example is SIM Swap fraud, which spans across CRM, authentication, and network systems. An agent that can check relevant signals across all three in the narrow window available when the fraud takes place has a much higher potential of success than humans. A similar pattern can be seen in bypass fraud that focuses on routing international traffic. Agents don’t just detect the fraud through pattern analysis; they can initiate the blocks as well.
B2B & wholesale sales
Developing quotes for enterprises is research-heavy work that may take days, delaying responses and possibly losing enterprise clients to competitors. Agents can check serviceability at each customer site, confirm available capacity, and implement pricing rules, all before a human reviews the quote.
Contract analysis is mainly a generative AI domain, but agents build upon it by acting on the extracted information, autonomously when feasible or with human oversight.
Another agentic use case is partner settlement, where agents can reconcile interconnect and roaming usage records between carriers and identify discrepancies, automating away several manual parts of this process. It’s worth noting that all three use cases carry consequences for business relationships, so agentic permissions and escalations should be carefully established.
Not sure how to identify the right use case for agentic deployments?
Our agentic AI teams and industry experts can analyze your operations, AI adoption maturity, and data ecosystems toidentify optimal use cases for agentic deployments.
How agentic AI actually runs on a telco stack
An agentic system that resolves a fault or fixes an order needs at least six core elements.
I. Orchestration and agent runtime
A runtime is where an agent's logic, state, and lifecycle control live, while orchestration ensures that multiple agents work in tandem to complete a task. This layer decides what the agents are working on, what (and how many) steps they would take, and when to stop.
II. Model layer
This is where the reasoning capability lives. It hosts or calls the foundation models behind each agent and routes requests to the appropriate model, ensuring that heavier reasoning goes to stronger, more capable models, while simpler requests are handled by local or more affordable models.
III. Tool and API layer
The tool and API layer sits between agent requests and systems like BSS, OSS, network management, and CRM, handling permissions, actions, and logging requests.
IV. Retrieval
Retrieval is responsible for contextual awareness of agentic AI systems, grounding it in network topology, customer records, and prior tickets.
V. Observability and evaluation
This layer records agentic behavior and actions and evaluates whether it was done well.
VI. Guardrails and human-in-the-loop (HITL)
Guardrails curate agent behavior, ensuring it is appropriate and within governance boundaries, while HITL takes care of manual checks and oversight.
Adding AI agents on top of legacy BSS & OSS
Most telcos have some combination of Amdocs, Netcracker, Salesforce, and Oracle, along with custom systems built in-house. When agents need to act within these systems, the overarching challenge is AI integration. There might be APIs that allow reads but not writes, no reliable way to track state change, batch interfaces available even when the agent needs a synchronous response, and inconsistent identity across systems.
This is why it’s often the tool layer that decides whether an agentic pilot reaches production in a telecom environment, instead of the model. Integration platforms like MuleSoft typically become the layer that bridges the integration gaps, providing agents access to a governed set of actions guided by permissions and logging, instead of directly accessing databases and sensitive controls. This layer is tied less to agentic development and is more akin to middleware development and integration engineering.
Guardrails, evaluation & observability
Four things must come together before an AI agent, or agentic system, can take a production-stage decision or act in a live environment.
- Action scoping – Decides what an agent can do and what its limits and constraints are. This includes order types, value thresholds, regions, and time windows.
- Evaluation harness – Gauges agent performance against known cases, establishing a baseline. This helps track changes in agent behavior, triggered by changes in instructions or the underlying model.
- Rollbacks – Ensure that even permitted actions can be reverted to a stable, former state if they don’t lead to desired results.
- Audit trails – Made up of an agent’s (and agentic system’s) inputs, decisions, actions, and approvals. This helps in determining where things went wrong (if needed).
Collectively, these four things turn agents that only seem to work into agents that work. This is useful for both change advisory boards and regulators.
Why most telecom agentic AI pilots do not reach production
There are five failure modes that telcos must understand, and none of them are related to
model capability.
- Missing or inadequate tool layer: If an agent can read but not end, a human must take an action. That erodes the distinction between generative AI and agentic AI.
- No evaluation baseline: Without measuring the current process against relevant KPIs, there is no way to prove improvement with agentic deployment or justify scaling.
- Unclear action authority: Agents that are not given the authority to act autonomously (without human approval) may serve as a glorified advisor and nothing more.
- Data residency influencing model choice: An agentic pilot built around a frontier model may not scale if the model sits outside the region or perimeter of customer data.
- No owner in operations: If no individual or team in the production environment takes responsibility for an agentic deployment, it may remain an innovation-focused pilot.
Governance, data residency & regulatory readiness
Agents that act on customer data naturally inherit every obligation that applies to that data. Elements like consent and regulations govern not just what an agent stores but what it processes and “sees.” Audit logging cannot be selective and must capture agent decisions alongside system events and tool calls, because regulators can ask why certain decisions were made and where the reasoning deviated from the acceptable path.
Under the EU AI Act, systems making high-impact decisions about individuals carry additional obligations, and some telecom decisioning can fall within this scope, depending on the function.
Regional data residency often determines model choice. In places where data cannot leave a jurisdiction, locally hosted open-weight models emerge as a practical option. This can significantly change the capability of agentic systems and should be addressed while the agents are being designed. Internal approval paths matter too, since security, data protection, and operations each hold a veto.
Agentic AI engineering for telecom operators
Engineering reliable agentic systems across complex telecom environments requires strong governance and oversight. In practice, that entails building the layers that decide whether agents reach production: integration and tool layers over BSS and OSS, agentic AI development and orchestration, evaluation harnesses and agent testing, and the governance controls that let operations teams accept agent actions.
This combination of AI engineering, systems integration, and operational governance is what turns agentic capability into production-ready automation.
Conclusion
The distance between a working pilot and an agent authorized to touch a production order is not model capability. It is a clean tool layer, measured evidence, defined action authority, and an owner in operations. Operators that build those four things move agents into production. Those that do not run pilots indefinitely.
TABLE OF CONTENTS
Ready to move beyond the pilot?
Our AI engineers combine end-to-end agentic capabilities with telecom industry experience to build resilient and scalable agentic systems.
Related blogs
AI/ML
Building an AI Integration Strategy
Learn a practical 9-step AI integration framework to define outcomes, overcome organizational barriers, measure ROI, and build AI solutions that...
AI/ML
Agentic AI in the Telecom Industry
The telecom industry is embracing agentic AI for multiple operational and customer-facing use cases, while navigating legacy systems, integration, and...
AI/ML
How is AI being used in real estate in 2026
Discover how AI is transforming real estate, from asset management to contract intelligence, with practical use cases and signs that...
AI/ML
Enterprise AI Agent Security: Lessons from the OpenAI Incident
The recent OpenAI testing incident offers a timely reminder that AI agent security needs to go beyond the model itself....
AI/ML
Top 10 offshore software development companies (2026)
In a recent Mauloa podcast, 10Pearls CEO Imran Aftab shares practical lessons on AI-native transformation, leadership, engineering excellence, and building...
AI/ML
AI Code Security: Who’s Testing AI-Generated Code?
This article compares the top 10 custom software development companies for 2026. The rankings are based on the average review...
AI/ML
Navigating the AI Shift: Lessons from Imran Aftab
In a recent Mauloa podcast, 10Pearls CEO Imran Aftab shares practical lessons on AI-native transformation, leadership, engineering excellence, and building...
AI/ML
How AI Is Reaching the Charity Sector
AI is reshaping how charities work and how people donate. The real opportunity lies less in new tools than in...
Get in touch with us
Global digital transformation and product engineering partner.