TELCOS

AI at telco scale, governed at telco scale.

Tier-1 deflection on millions of customer interactions. Network ops alarm correlation across hundreds of thousands of network elements. Field tech copilots that know the equipment and the ticket history. Cost economics that work at the volumes telcos actually run.

Telcos are the volume operators of generative AI. A tier-1 European mobile operator handles 30–80 million customer interactions per year across web, app, voice and chat. Network operations centres process tens of millions of alarms per month. Field technicians close hundreds of thousands of tickets. The unit economics of frontier-model calls at this scale are decisive — a single design choice between a frontier model and a fine-tuned SLM is a five-to-eight-figure annual cost difference.

The regulatory perimeter has tightened: NIS2 brings telcos in scope as essential entities, with 24-hour incident reporting windows. The EECC governs network operations and customer rights. ePrivacy and GDPR govern the customer-facing surface. National regulators (AGCOM, Ofcom, ARCEP, BNetzA) impose conduct rules on retail offers and outage communication.

Kosmoy is the operating layer that lets the telco move at volume without losing control of the policy point or the cost ledger. The Gateway enforces PII redaction, prompt-injection defence and model selection at every call. The Cost Tracking module attributes every model call to the team, app and customer interaction. The Action Capsule contains the agents that take customer-facing actions.

What this industry runs into.

Volume economics

Frontier models at 30M-conversation scale dwarf any other operating cost line. The router decision — frontier vs fine-tuned SLM — is the difference between a viable AI program and a runaway bill.

OSS / BSS modernisation drag

Customer-facing AI is bolted onto OSS/BSS systems built in different decades. Agents that read across billing, CRM, network inventory and provisioning need a managed integration surface.

Network operations complexity

5G networks have orders of magnitude more elements than 3G/4G. NOC analysts can't process the alarm volume manually. AI-assisted alarm correlation is no longer optional.

ePrivacy and customer consent

Personalised AI-driven retention offers and behavioural recommendations sit on top of consent and ePrivacy obligations. The agent has to know the consent state, not assume it.

Regulatory landscape.

The regulations that shape AI in telcos — and where each one bites on AI deployment.

NIS2Network and Information Security Directive 2· EU

Telcos are essential entities. AI infrastructure must meet incident reporting (24h/72h windows), risk management and supply chain security obligations.

ePrivacy DirectiveDirective 2002/58/EC + national transpositions· EU

Personalised offers driven by AI on top of communications metadata require explicit consent. Tracking and profiling rules apply.

EECCEuropean Electronic Communications Code· EU

Customer rights, contract transparency, switching and porting — all bound on AI-driven retention or sales motions.

GDPRGeneral Data Protection Regulation· EU

Telco customer data is personal data at scale. Profiling under Art. 22 covers AI-driven retention/upsell decisions. Data minimisation pressures retrieval scope.

EU AI ActRegulation (EU) 2024/1689· EU

Customer-facing chatbots must disclose AI nature (Art. 50 transparency). High-risk classifications less common but possible for credit-related decisions on post-paid offers.

National regulatorsAGCOM (IT), Ofcom (UK), ARCEP (FR), BNetzA (DE)· National

Each imposes retail conduct, outage communication and number portability obligations that bound AI use in customer touchpoints.

Use cases that are actually shipping.

Tier-1 customer support deflection

Customer opens chat: 'why is my bill higher this month?'. The agent reads the bill, the previous month's, recent plan changes, usage spikes and tariff promotions, and drafts a structured explanation with the specific line items that drove the change. If the customer accepts the explanation, the conversation closes; if not, hand-off to a human agent with full context.

Tier-1 deflection rises 20–35% on bill-related volume. Cost per resolution drops by an order of magnitude on cases the SLM handles. Customer NPS on bill-related touchpoints improves because the agent answers in seconds, not minutes.

Powered byLLM GatewayLLM RouterCost TrackingGuardrails

Network ops alarm correlation

NOC analyst opens the morning queue: 50,000 alarms across 100,000 network elements overnight. The agent correlates by topology, time, alarm type and known runbooks, and presents a ranked list of 30–50 incidents with the suggested next runbook step. Analyst decides; the agent never auto-resolves an alarm.

MTTR on common incident types drops 25–40% because the analyst spends time on diagnosis, not categorisation. NOC headcount stays flat as network complexity grows.

Powered byAction CapsuleInsights DashboardAgent Registry

Field tech copilot

Tech arrives at a customer site for a fibre installation. The copilot on the tablet reads the work order, the property's prior service history, the equipment manuals for the OLT/ONT involved and the current ticket-resolution playbooks. Tech has the context plus the procedure surfaced automatically.

First-visit fix rate rises 8–15%. Field-time-per-ticket drops 10–20%. Newer techs reach senior productivity within a quarter rather than a year.

Powered byRAG-in-a-BoxAction Capsule

Churn risk outreach

Customer's contract is approaching renewal and signals (NPS detractor, recent complaint, price-comparison search) suggest churn risk. The agent drafts a personalised retention proposal — a tariff change, a roaming bundle, a device upgrade — within the consent boundary the customer has granted. A human reviews and authorises before send.

Retention rate on at-risk segments rises 5–15 percentage points. Cost per save drops because the agent drafts at scale; humans review the most consequential offers.

Powered byAgent BuilderGuardrailsInsights Dashboard

5G enterprise slicing Q&A

Enterprise customer asks: 'can your 5G slice deliver 50ms latency for our manufacturing site in Bergamo with 99.99% availability?'. The agent reads the network capacity, regional coverage, slice templates and SLA history, and drafts an indicative answer. Account team validates and binds.

Enterprise sales response time on slice questions drops from days to minutes. Account teams concentrate on the 20% of opportunities that need human judgement.

Powered byRAG-in-a-BoxAgent Builder

Agent governance

Where telcos agents need extra discipline.

Telco agents operate at volumes that make Cost Tracking, LLM Router and Insights Dashboard non-negotiable. The Router routes the simple 80% of conversations to a fine-tuned SLM running on telco infrastructure; the harder 20% reach a frontier model. Cost Tracking attributes every call to the channel, app and ultimately the customer interaction — when the CFO asks 'what does our AI cost per chat session?', the answer is auditable.

Customer-facing agents that take actions — porting, plan change, dispute submission, retention offer acceptance — run in Action Capsules with explicit allowed-action scope. The Agent Registry tracks every agent persona deployed across channels (mobile, web, IVR, retail, partner) with its risk class and operational metrics.

Chatbot use cases

Chatbots, by surface and risk class.

Telco chatbots are the highest-volume chatbot deployments in any industry. Customer-facing, partner-facing, internal — every channel has at least one. Governance has to scale with the volume, not lag behind it.

Mobile app self-service

Plan, balance, top-up, roaming, troubleshooting. Routed by LLM Router; PII guardrails strict; cost-attributed per session.

IVR + voice agent

Voice-first conversational AI for inbound calls. Hand-off to a human agent when the conversation classifier indicates frustration or complexity. Full transcript stored.

Partner / dealer portal

Dealer asks 'what's the commission structure on this enterprise plan?'. Citation-grounded answer from the partner portal docs.

Internal NOC / field copilot

Operations-facing chatbots for runbook navigation, manual lookup, ticket triage. Action Capsules where the chatbot can take incident-resolution steps; otherwise read-only retrieval.

How Kosmoy fits.

Telcos benefit from Kosmoy's volume-tested architecture more than any other industry. The Gateway and Router process millions of calls per day per deployment. The Cost Tracking module is what most telco CFOs ask about first — and what makes the AI program defensible at quarter-end.

Deployment fits the telco hosting posture: Kubernetes-native, single-tenant, in the operator's own cloud or on-prem infrastructure. Open-weight models on operator GPUs cover the bulk of customer-facing volume; frontier models are reserved for genuine complexity. Customer call transcripts and CDR-adjacent data never leave the operator's perimeter.

Module questions, answered straight.

Does Kosmoy work at our chat-session volumes (tens of millions per year)?

Yes. The Gateway and Router are designed to process millions of calls per day per deployment. The architecture is horizontally scalable; large telcos run multiple regional clusters with a federated AI Inventory at the group level.

How do you keep AI costs predictable?

The LLM Router routes simple Q&A to fine-tuned SLMs running on operator infrastructure (low marginal cost) and reserves frontier-model calls for complex cases. Cost Tracking attributes every call to the channel, app and segment. Budget alerts fire before spend escapes; hard caps where policy demands.

Does the Action Capsule slow down the customer-facing path?

No measurable latency difference at telco volumes. The Capsule is a Kubernetes pod with the Gateway as its egress; the network hop is the same as any internal microservice call. Pre-flight authorisation is a sub-millisecond check.

Can we mix open-weight and commercial models?

Yes — most telcos do. Llama / Mistral fine-tuned on operator data covers the volume tier; Azure OpenAI / Anthropic via the Gateway covers the complexity tier. The Router decides per-prompt which provider serves it.

Bring telco-scale AI under one operating layer.

See how the AI Gateway, LLM Router and Cost Tracking handle millions of customer interactions and tens of millions of NOC alarms.

Book a demoEmail sales@kosmoy.com