GenAI Value in Telecom Depends on Open Architecture

Building a GenAI demo is easy. Making it work inside a live telecom environment, where OSS, BSS, billing, CRM, network systems and regulatory constraints are tightly interconnected, is something else entirely.

When you introduce GenAI into telecom, you’re stepping into years of accumulated complexity.

That’s why, when we designed Cognitive Core, we made one decision early: it had to be open from the beginning. In my experience, adding openness later always makes integration harder, not easier.

Open integration works with existing environments

In telecom, because there’s always ongoing modernization, and constantly evolving environments, there’s an assumption that meaningful AI value requires replacing systems. But when a GenAI system is open by design, supporting standard interfaces, such as MCP and REST, and adapting to additional protocols when needed, it’s able to work with the systems operators already have. This makes it possible to introduce GenAI incrementally while broader transformation programs continue.

Open data enables reliable context

With GenAI, a lot of attention is paid to prompts, but in telecom, it’s context that’s the harder challenge. GenAI must understand how customers, products, services and network elements relate to each other, and if those relationships aren’t clear in the data, even a strong model won’t perform well. I’ve seen situations where the model was blamed when the real issue was fragmented data across systems.
Reliable context requires coordinated access across environments, and open integration makes that possible. Without it, responses may sound fluent, but they won’t be consistent.

Open architecture goes beyond a single LLM

Models will evolve, new ones will emerge, and costs and regulations will shift. So, if a platform is tightly tied to one provider, that dependency becomes a risk. Industry initiatives like the GSMA’s Open-Telco LLM Benchmarks reflect the need for consistent model evaluation, and architecturally, the model should be replaceable without redesigning the system around it. We deliberately separated the model layer from orchestration so Cognitive Core can work with models such as GPT or Gemini, open-source models like Llama or Mistral, or models hosted within a customer’s own environment. We also use a mix of frontier and smaller LLMs that can be fine-tuned with telco-specific data which makes the overall solution more efficient.

Open agents show that one size doesn't fit all

Telecom problems are complex, and a practical way to manage complexity is to break them into smaller parts. The same applies to GenAI: a single agent can answer questions, but it won’t have deep expertise across customer care, billing, sales and network operations at the same time.

We use specialized telco-specific agents, focused on areas such as Sales or Billing, which are part of the pre-built domain-specific agent libraries within Cognitive Core. They can operate independently or collaborate when workflows cross domains. These GenAI agents interact with users, and with each other, through structured agent-to-agent communication. In that respect, it’s not very different from how human teams work today, (but instead of person-to-person, it’s agent-to-agent). As a multi-agent solution, it can integrate within broader multi-agent ecosystems, including agents from other providers.

Open access to users delivers a better experience

Technology enables capabilities, but ultimately, it’s the user experience that determines whether people will choose to adopt it. GenAI interacts through voice, chat, web and messaging platforms, and regardless of the channel, the experience needs to feel natural. In voice, responses should be concise. In chat, interaction shouldn’t be limited to text; structured tables, buttons, images and guided options make responses clearer and more actionable. GenAI must be able support multiple channels including users’ preferred channels of choice like WhatsApp and allow users to provide feedback so the system can improve over time.

Open partnerships support flexible deployment

Telecom AI doesn’t operate in isolation. Hyperscalers such as Microsoft, Google and AWS are central to modern telecom environments. Openness here means having the ability (and flexibility) to work within an ecosystem, run in cloud, hybrid or fully on-prem deployments and integrate with hyperscaler platforms.  

Open metrics and KPIs make value measurable

Operators are embracing domain-specific agents, but they still want to stay in control. Observability, governance and clear LLM access-management are essential, and that starts with monitoring system usage, response times and token-consumption. But measuring activity isn’t the same as measuring impact – operational KPIs, such as AHT (average handling time), CSAT (customer satisfaction) and NPS (net promoter score), show whether meaningful value is being created. Open monitoring, tracing and out-of-the-box reporting are essential for ensuring this information is accessible and understandable, so measurement should be built into the system rather than added later.

Security is the one area that definitely isn't open

Above all, security needs to be embedded into the heart of any application and can’t be just a test performed at the end. Telecom AI systems handle sensitive customer and network data, so encryption, access control and governance are tightly controlled by design. So, while integration can be flexible, the opposite is true for security boundaries.

Open architecture makes GenAI change possible and manageable

GenAI will continue to evolve, and architecture that isn’t designed to adapt will struggle. 
We built Cognitive Core so models, systems and agents can evolve independently, and that’s what open architecture enables. For me, openness isn’t an ideology – it’s a practical way to keep GenAI usable as telecom environments continue to change.