Industry
Communications, Media and Information Services
Connect. Curate. Comprehend: Unlocking Value Beyond Boundaries
Highlights
- Telecom operations are no longer limited by the ability to automate. They are limited by decision quality — knowing where to look, when to escalate, and how to act with confidence.
- TCS is fine-tuning NVIDIA Nemotron open models on operators’ own network, engineering and language data, inside the operator’s own trust boundary.
- A sensing-led architecture — monitor the entire network efficiently and apply deep diagnostics only where needed — makes autonomy economically scalable without runaway inference costs.
- Voice AI in African languages shifts telecom from customer service to genuine digital inclusion, with sovereignty and data ownership retained by the operator.
On this page
On this pageinpage
Automation is not autonomy
Most operators have automated a great deal. Scripts run, workflows fire, and tickets close—but they automate known tasks, not autonomous decisions. Autonomous network operations must continuously sense changing conditions, reason over telecom-specific context, select and coordinate actions within operator-defined policy, validate outcomes, and escalate when human judgement is required. That is why operations can still feel reactive: the hard part is not execution, but deciding where to act first, when to escalate, and how to justify the decision afterwards.
Three gaps show up consistently. Assurance runs on batch reviews, so anomalies are confirmed hours after customers have already felt them. Field engineering knowledge sits locked inside design documents, handover packs and site photographs that no system reads end to end. And customer-facing voice channels do not work in the languages a large share of subscribers actually speak.
These look like three problems. Architecturally they are one: general-purpose AI does not understand telecom, and telecom data cannot be sent to general-purpose AI. The answer in every case is the same — take capable open models, fine-tune them on the operator’s own data within the operator’s own boundary, then use governed agents to translate that domain-specific intelligence into policy-aware autonomous decisions and actions, with evidence, auditability, and human-in-the-loop control where required.
That is why TCS is building with NVIDIA Nemotron. Open models give operators a foundation they can adapt to their own network, engineering, language and policy context, while retaining control over where sensitive data is processed, how models are evaluated and where they are deployed. Open weights also make sovereign and in-tenancy deployment viable. Because Nemotron spans Nano, Super and Ultra models, TCS can match model capability and cost to each task instead of using the largest model for routine work. And the surrounding NVIDIA stack provides a path from experiment to production, not from experiment to demo.
TCS is applying this foundation to three telecom challenges: validating that network sites were built as designed, helping operations teams sense and diagnose emerging network issues, and extending digital services into the languages customers use every day. Each has a different objective, but all follow the same path to governed autonomy: operator-specific models and data, agents grounded in operational context, policy-controlled access to tools and systems, traceable evidence, and human oversight for consequential decisions.
Turning site evidence into engineering insight
Before a new or upgraded radio access network (RAN) site can enter service, the operator must verify that it was built as designed. Field crews document the installation through photographs, equipment records, checklists, test results and commissioning documents, often assembled into a site-handover package for engineering review. Today, validating that evidence against approved design documentation is manual and inconsistent, and often happens long after the engineer has left the site. Deviations in equipment count, placement, labelling or earthing therefore surface late, when correcting them can require a second truck roll.
TCS is fine-tuning the NVIDIA Nemotron Super 3.5 VL model to reason across drawings, schedules and photographs together — recognising RAN components, reconstructing how a site is actually connected, and flagging where the build diverges from the design. Operator-specific detail that changes site by site, such as naming conventions and approved substitutions, is supplied through retrieval rather than baked into the model, so a convention change does not trigger a retraining cycle.
Just as important is what the system does not do. It makes no automated compliance approval or rejection. Every finding carries a confidence score and a traceable evidence reference, low-confidence cases route to human review, and the determination stays with the operator’s engineers. The pilot is being measured against the current manual process using detection precision and recall, review effort per site, validation turnaround time, and repeat truck rolls avoided. Where equivalent test conditions are available, TCS will also compare the fine-tuned Nemotron model with relevant baseline models and report inference efficiency and cost. Quantified results will be published only after operator validation. The business case is straightforward: catch deviations while the crew is still on site, and convert a manual audit backlog into a prioritised exception queue.
From batch review to sensing-led autonomy
Autonomous network operations depend on more than agents. They depend on knowing where agents should focus. Across millions of cells and hundreds of network signals, it is neither necessary nor economical to apply the most computationally intensive reasoning to every event, everywhere, all the time.
TCS and NVIDIA are building a multi-fidelity AI sensor architecture that senses broadly, diagnoses selectively, and acts precisely. NVIDIA NV-Tesseract continuously analyses multivariate network telemetry to detect emerging anomalies, forecast changing conditions, and prioritise the hotspots that warrant attention. NVIDIA Nemotron then provides the telecom-specific reasoning layer, bringing together relevant telemetry, network topology, change history, and operator context to assess likely causes, evaluate remediation options, and recommend the next best action. Pilot evaluation covers operator-approved indicators across service availability, accessibility, retainability, mobility, traffic, latency and fault behaviour; the exact key performance indicator (KPI) set and baseline values remain operator-specific and will be published only if validated and approved for external use. Action executes through the operator’s existing optimisation frameworks, inside policy guardrails, with full auditability.
Holding it together is a persistent agent runtime built on NVIDIA NemoClaw blueprints, which retains memory of change history, anomaly drift and what previous interventions actually achieved — and uses that memory to decide when deeper inspection is warranted.
The result is a practical path from periodic, batch-based review to continuous, evidence-led operations. Operators can focus expert attention on the conditions that matter most, respond faster to emerging issues, and advance toward governed autonomy without replacing the operational systems they already use.
Voice AI that speaks the customer’s language
For an operator serving tens of millions of South African subscribers, a digital experience cannot be English-only. Customers need to be able to access services and resolve issues in the languages they use every day. TCS is fine-tuning Nemotron speech models for isiZulu and extending the approach across South African-accented English, isiXhosa, Sesotho, Xitsonga and Afrikaans, with a multilingual foundation designed to reach Francophone African markets later.
The objective is not simply transcription or translation. It is to create voice agents that can understand customer requests, use approved operator knowledge and service tools to help complete tasks, and escalate when confidence is low or human judgement is required.
For many African languages, the barrier to production is not only model capability. It is the limited availability of representative, consented speech data and the absence of widely accepted public benchmarks. Most are absent from the supported tiers of even the strongest open speech models, and for some, no public benchmark exists at all. TCS therefore structured the engagement as a six-week, gated de-risking exercise rather than a delivery sprint — the deliverable is evidence and a costed production plan, with an explicit go/no-go decision each week.
The pilot will compare baseline and fine-tuned models using operator-approved test sets, measuring speech-recognition error rates, intent accuracy and task-completion quality by language and accent. Operational evaluation will track human-escalation rates, manual handling avoided, average handling time and cost per completed interaction. Quantified improvements will be published only after the operator validates the results and approves them for external use.
One early finding reshapes the entire investment case. Related languages may transfer to one another: isiZulu and isiXhosa are both Nguni and substantially mutually intelligible, and Afrikaans is Dutch-derived. If transfer holds, production becomes two or three language-family programmes instead of six independent ones — which is often the difference between a fundable initiative and an unviable one.
A second finding is equally practical. The binding constraint on African-language voice AI is not graphics processing unit (GPU) capacity. It is human transcription by native speakers, which runs at several times real time and becomes the critical path long before compute does. Building that capability early gives the operator a more realistic view of cost, timeline, quality thresholds, and the conditions required to scale responsibly.
What makes this repeatable
Across all three programmes, the repeatable pattern is consistent: operator-specific models and data remain within the operator’s boundary, model capability is matched to the task, and consequential decisions remain evidence-led and human-governed.
The opportunity is to move networks from passive transport toward cognitive platforms that sense, reason and act — at a cost, risk and governance model operators can sustain.
Working with NVIDIA, TCS is helping operators build networks that are easier to interpret, safer to optimise at speed, and able to serve customers in the languages they use every day.