Originally published on LinkedIn:
https://www.linkedin.com/pulse/i-came-away-from-day-2-dtw-ignite-slightly-annoying-nikolaj-fl%C3%B8jgaard-ztqnc/
Agents are the easy story. Changing the operating model is the hard one.
Not because the AI story is weak. It isn’t. If anything, the direction is becoming much clearer.
The industry is moving beyond generic AI assistants and toward something far more useful: domain-specific agents orchestrated around real work, grounded in enterprise data, controlled through trust layers, and routed through model gateways that balance cost, privacy, risk, and sovereignty.
That’s a much better conversation than “let’s add a chatbot.”
But the question I kept coming back to wasn’t whether telcos can build agents.
Of course they can.
The harder question is whether telcos can change their operating model enough for those agents to matter.
From AI demos to agentic operations
One of the most interesting patterns I saw wasn’t a single AI assistant sitting on top of everything.
It was a structured agentic architecture built around:
- Domain-specific agents focused on real operational areas
- Orchestration layers that determine which agent acts, when it acts, and with what authority
- Context layers that connect agents to systems, workflows, tools, and business meaning
- Trust layers with permissions, approvals, audit trails, and escalation paths
- LLM gateways that route workloads across models, clouds, and execution environments
That stack matters.
In a telco environment, agents cannot simply be clever. They must be operational.
They need to understand the domain, respect permissions, explain their actions, recover from failures, and know when to hand work back to a human.
The LLM gateway is becoming a strategic control point
One platform view that stood out combined agent frameworks, templates, RAG capabilities, LLM gateways, MLOps, monitoring, and infrastructure across public and private cloud.
That feels close to the right mental model.
For telcos, model selection won’t become a religious debate. It will become a portfolio decision.
Some use cases require low cost. Some require low latency. Some require strong privacy. Some require local execution. Some require the best frontier models available. Some require enough transparency for operators, auditors, security teams, and regulators.
This makes the LLM gateway far more than a technical convenience.
It becomes a control point for:
- Cost management
- Governance and policy enforcement
- Sovereignty requirements
- Model diversification
- Operational risk management
It may not be the flashiest part of the stack, but it could become one of the most important.
Guardrails are part of the product
The same applies to trust and governance.
A production agent cannot simply be “helpful.”
It must know what it is allowed to do. It must know when not to act. It must leave an audit trail. It must escalate cleanly. It must operate within tightly controlled permissions.
And it must survive contact with reality: legacy systems, incomplete data, manual exceptions, and operational complexity.
This is where many AI programs will either become real or quietly become expensive side projects.
Guardrails are not something you add after the fun part.
They are part of the product.
The real transformation is process, not tooling
One slide that stayed with me described the evolution from tools, to agents, to agentic factories.
Tools help individuals move faster.
Agents automate pieces of work.
Agentic factories require organizations to redesign how work actually flows.
That’s where the real transformation sits.
If a process is fragmented, political, manual, and full of hidden approvals, agents won’t magically fix it.
They may simply automate the mess faster.
This is why I keep coming back to the same conclusion:
AI usually fails because of people, culture, incentives, ownership, governance, and operating models—not because the model wasn’t smart enough.
The model is the visible part.
The hard work is data quality, process ownership, architecture, integration, accountability, and the willingness to stop doing things the old way.
The diversification question remains unresolved
Another recurring theme was diversification.
Telcos know they cannot rely solely on connectivity forever.
The pressure is obvious:
- Networks remain highly capital intensive
- Customer expectations continue to rise
- Pricing power is limited
- Investment cycles never really stop
So the diversification story makes sense:
Data centers. Sovereign cloud. AI infrastructure. Managed services. Digital identity. Enterprise platforms. Industry ecosystems.
Diversification is the uncomfortable part of the telco AI story. Connectivity alone is probably not enough to fund the next transformation cycle. If telcos want to become serious players in data centers, sovereign cloud, AI infrastructure, enterprise platforms, and industry ecosystems, they need more than ambition.
They need capital discipline, commercial focus, and a clear view of where they can actually win. Otherwise diversification becomes another expensive transformation narrative layered on top of a core business that is already under pressure.
All of that is logical.
But the question I keep asking is simple:
Can enough telcos generate enough new revenue to fund the transformation they are describing?
I’m not sure.
The opportunity is real.
So is the cost.
It’s one thing to talk about AI-driven efficiency, autonomous operations, faster delivery, and new revenue streams.
It’s another thing entirely to fund the platforms, modernize the architecture, clean the data, retrain the workforce, change the operating model, manage the risk, and still deliver shareholder returns while the core business remains under pressure.
That tension may be the most interesting part of the telco AI story right now.
My takeaway
I liked the ambition.
I liked the shift from generic AI toward domain-specific agents.
I liked the focus on orchestration, context, trust layers, guardrails, LLM gateways, and model diversification.
I liked the practical examples around autonomous networks and customer service troubleshooting.
And I especially liked that the conversation is moving from:
“Can we build it?”
to
“Can we operate it?”
Because I don’t think the winners will be the companies with the loudest AI story.
The winners will be the companies that turn agentic capability into operating leverage.
They will simplify workflows instead of wrapping them in AI.
They will treat guardrails as architecture rather than paperwork.
They will use model gateways to control cost and risk.
They will build agents around real domains instead of abstract demos.
And they will be honest about whether diversification is a genuine business strategy or simply another transformation slide.
Agentic AI in telecom is no longer a vague future concept.
The building blocks are becoming real.
Now the question is whether business models, culture, and operating models can move fast enough to make those building blocks matter.
Good sessions today from Paul Bosch, Dr. Rainer Deutschmann, Nathan Bell, and Astrid Kyhl.
Very inspiring day.
Still not convinced the economics are easy.
The real test is not whether telcos can build agents. It is whether they can use AI to create enough operational leverage and new business to justify the investment.