Artificial Intelligence deployment in telecom

We’ve noticed recent news about Artificial Intelligence (AI) planned deployments by telecommunications voice service providers. It’s an interesting topic, and we thought we’d share some highlights and background info. Let’s have a look.

AI in telecom: an illustrative example

There have been many examples of AI planned deployments, but let’s use a recent announcement from Orange Business in Paris, France, as an example. The announcement describes their plans to “leverage trust and AI to reimagine enterprise communications.” Highlights include:

• Branded Calling. Orange Business is rolling this out in France and the United States this year. It enables enterprises to display their name, corporate logo, and call reason on a recipient’s mobile screen.
• Deepfake Detection. The solution will detect fake audio in calls. The portfolio of solutions will extend to other forms of communications to detect fake images, video, and documents.
• AI-Augmented Customer Care and Agentic Telephony. This Live Intelligence solution will embed AI into contact center and CRM environments. The benefits claimed include faster response times, optimized costs, and improved customer satisfaction.

AI risks and concerns

The announcement acknowledges the risks of conversational AI, such as hallucinations and accuracy errors. Here’s their take on this issue:

“Trust has become the most valuable currency in the digital economy. At the same time, AI is transforming how work gets done. We believe the future of enterprise communications must combine both.” — Usman Javaid, Chief Product and Marketing Officer, Orange Business.

You might be thinking, “Hey, wait a minute. Tell me more about these risks.” Here’s a quick overview to get you up to speed on the issues. We’ll start with a basic explainer.

How AI LLMs work

AI includes a wide range of machine learning and automated actions driven by analytics. We’re going to briefly cover one type of AI that’s been getting lots of attention: LLMs (Large Language Models) and conversational AI agents processing text. This explanation is very brief and perhaps over-simplified, but it should give you a basic idea of how it works.

LLMs are trained with source data, such as all the webpages that can be crawled on the internet. The crawler converts the text it finds into tokens, a process called encoding.

The tokens are then analyzed using Natural Language Processing to compute the relationships between them. In 2017, researchers at Google came up with a new way to do this, called a Transformer Architecture, which they described in a paper entitled Attention is All You Need. It’s a very technical paper, but all you need to know is that it describes a newer, faster, better way to train LLMs that made the recent AI boom possible.

When you use an LLM in an interactive session with an AI agent, your input prompt text is encoded into tokens and token relationships, then compared to the LLM database of tokens and token relationships. Word by word, the software uses the token relationships to formulate a token response and decode it into text. Think of it as a fancy version of autocomplete.

Optimism and enthusiasm, then some disappointment

When one first uses an AI agent powered by an LLM, it seems amazing, like a conversation with a real person. Can’t blame anyone for tremendous optimism and enthusiasm. This will change the world we live in.

If one uses AI for a while, however, one will sometimes feel disappointed. Here are some examples of recent AI-related disappointments.

• On March 10, 2026, news broke that Amazon had experienced a trend of incidents in recent months involving “Gen-AI assisted changes.” In one incident, the Amazon website and shopping app went down for nearly six hours. They resolved to have AI-assisted work by junior and mid-level engineers reviewed by senior engineers.
• In October 2025, the Center for AI Safety and Scale AI published a study that examined whether AI agents were able to complete computer-based work. The study used freelance projects that were posted on and completed by contract workers on Upwork, a website for freelancers. The study tested the success rates of six different AI agents in completing the work. The researchers found that the best of the AI agents could only achieve a 2.5% success rate. The authors believe this shows that most economically valuable remote work currently remains far beyond the capabilities of AI.
• In February 2026, the 5th U.S. Circuit Court of Appeals sanctioned an attorney for using AI to draft a legal brief and failing to verify the accuracy of the content. The Court found 21 instances of fabricated quotations or serious misrepresentations of law or fact.

What we think this means

What does this mean? Is AI a bunch of hype?

We don’t believe so. In our experience, we’ve found that AI can be a valuable resource. However, AI isn’t perfect. Just like autocomplete, sometimes it misses the mark and gives you something you didn’t want and cannot use.

So, how should one use AI? Here are a few observations and suggestions:

• When using an LLM AI model that’s highly trained in a particular topic, it can do a great job of quickly bringing you lots of useful information.
• If you’re exploring a niche topic with an AI LLM model that was not well-trained for that, then it can miss the mark.
• AI agents are typically trained to respond with a confident, enthusiastic, “can-do” tone of expression. Don’t let that lure you into trusting a response without verifying it.
• Be thorough with your input prompts. AI is more likely to give an accurate answer if you give it the context it needs. Better to over-explain than under-prompt.
• Keep the conversations short. AI chats have a context window, which is a limit on how much content they can track. AI agents can lose the plot in a long chat. Start a new one if it’s getting too long or if the discussion starts going in a different direction.
• We think you should give AI a chance. Just verify the information that it provides.

TransNexus solutions

TransNexus is a leader in developing innovative software to manage and protect telecommunications networks worldwide. The company has over 25 years of experience in providing telecom software solutions including toll fraud prevention, robocall mitigation and prevention, CDR and call analytics, advanced call routing, billing support, STIR/SHAKEN, and branded calling.

Contact us today to learn more.