AI is being rapidly adopted across the customer journey, presenting both opportunities and risks. As regulators increasingly engage with the potential impacts of the technology, they must strike an effective balance between mitigating harms and promoting responsible innovation.
Conversational chatbots for customer support represent the most common consumer-facing application of AI in telecoms today. Most major operators have now deployed chatbots, providing a simple and familiar entry point for customers to begin interacting with the technology.
Applications that rely on recent advancements in generative AI and agents are currently at a relatively early stage of deployment but hold significant potential for improved customer experience and support for vulnerable groups, as well as for protection from harmful online content.
AI could pose significant risks for consumers, particularly regarding data privacy, bias and the black box nature of the decision-making process. Bad actors can also exploit AI to increase the volume of fraudulent communications, widening the pool of potential victims.
Regulators are increasingly exploring the risks and opportunities of AI for telecoms customers, with ACMA in Australia an early mover in this area. Arcep in France has taken the most forward-looking approach, raising concerns about the possible threat AI could pose to the open internet.
To date, direct regulatory interventions targeting AI’s role in the customer journey have been few and far between. However, the FCC in the US has been notably active, introducing rules to both mitigate risk and enable the technology to directly benefit end users.
Effective oversight will require regulators to keep pace with the rapid evolution of AI and navigate a delicate balance between over- and under-regulation. Internationally-agreed principles could offer a relevant blueprint, although their voluntary nature may limit practical application.
Across the telecoms customer journey, operators are predominantly using AI to support existing customers
Artificial intelligence (AI) has gained significant traction over the past few years, particularly due to the rapid adoption of chatbots such as OpenAI’s ChatGPT, which attracted over 100m users globally within two months of its November 2022 launch. Within the telecoms sector, AI is in use across network operations, with Deutsche Telekom and Vodafone just two examples of operators employing the technology to optimise site deployments and/or spectrum management. Consistent with broader industry trends, the main consumer-facing applications of AI in telecoms are centred around customer service and support – see Figure 1. The continued deployment of AI chatbots in this area reflects the technology’s increasing maturity, while also offering a straightforward and familiar entry point for consumers to interact with AI.
Operators tend to implement AI most frequently in the “use” phase of the four-stage customer journey, typically through conversational chatbots, which are increasingly relied upon to handle customer interactions and have become more sophisticated thanks to advancements in generative AI. This includes Vodafone’s TOBi chatbot, which processes around 1m interactions each month. Another example is Deutsche Telekom’s voice-enabled Ask Magenta Voice, which has been available via calls on mobile phones since 2023, allowing customers in Germany to speak freely to the digital assistant rather than typing, offering customers greater flexibility and improving accessibility for those who might struggle with a traditional chat function. In Australia, Optus is using Google and Anthropic’s large language models (LLMs) to develop customer-facing AI tools, including Sally AI, a digital assistant that guides customers through home internet setup and similar tasks through the end user’s handset camera.
In contrast, during the initial “discover” phase, consumers are more likely to use their own AI tools (such as ChatGPT or Google’s AI Overviews) to find and compare different tariffs and service bundles. In the “buy” phase, the adoption of AI at the point of purchase remains low, particularly where AI agents might make a purchase on a consumer’s behalf. This hesitancy is likely due to the financial commitment of multi-year telecoms contracts, where correcting an unsuitable purchase could be difficult, coupled with the current relative lack of maturity of AI agents in comparison to other generative AI tools. In the “renewal” phase, AI can alert operators to early signs of churn, such as reduced engagement, or develop personalised offers to incentivise consumers to continue, renew or potentially upgrade their service.
Fraud prevention, online content moderation and AI agents represent emerging applications of the technology
While chatbots are currently the most common consumer-facing application of AI in telecoms, several other use cases are emerging, for which widespread implementation is expected over the coming years:
Fraud prevention: Telecoms infrastructure is often exploited as a vector for cyber scams, which consumers typically experience via calls, SMS, emails or other online channels. AI can be used to detect and protect consumers from malicious activity, for example by monitoring network traffic for unusual deviations (e.g. a sudden surge in international calls) and using pattern recognition to compare this activity against historical attack data. This allows the system to proactively flag suspicious behaviour and alert consumers to potential fraudulent communications;
Online content moderation: AI can be used to moderate user-generated content (UGC) and improve online safety. By automatically scanning and filtering text, images and videos, AI can flag and recommend the removal of content that violates platform guidelines;
Accessibility features: AI offers the potential to improve accessibility for customers who face difficulties with using telecoms services due to their specific needs. This includes people who speak a foreign language, as well as vulnerable or elderly customers who often require more tailored or specific guidance when encountering an issue; and
AI agents: Moving beyond basic chatbots, agentic AI would be the next step in telecoms customer service, where agents can autonomously take action to resolve issues, requiring minimal human oversight. In particular, the buy phase holds untapped potential, where AI agents could function as sales assistants, guiding customers to the most suitable plan for them and then completing the sales transaction.
While AI could improve the customer experience, ongoing human oversight is needed to mitigate the risk of friction and frustration
AI has the potential to significantly improve the customer experience, particularly by providing instant, 24/7 support. Generative AI chatbots, in comparison to traditional chatbots that have scripted, basic responses, offer a more human-like interaction, enabling personalised responses and an increased speed of service, which could be especially beneficial for vulnerable groups. For example, older customers who might have difficulty diagnosing technical issues with telecoms products or services can be guided by conversational AI tools, making support more accessible and effective.
AI agents could go further to improve customer support, as they have the ability to not only respond to queries but to take autonomous action to fix specific issues. Unlike AI assistants, which can sometimes frustrate customers, AI agents can respond more comprehensively by accessing customer history and past conversations, allowing them to assist customers with actions such as upgrading or cancelling telecoms services. One NZ in New Zealand is a leading example having implemented Salesforce’s AI agents (Agentforce) in its customer service operations, using customer data to provide more personalised interactions, answer queries and schedule appointments.
However, operators also need to be cautious when using AI chatbots and agents for customer support, as not all consumers will be receptive to this model of engagement. A reliance on automated service without sufficient human support or oversight could frustrate customers, potentially alienating some and leading to higher churn rates, especially if chatbots fail to resolve problems. Operators should also be aware of the risk of hallucinations, whereby generative AI produces inaccurate or nonsensical responses, carefully outlining the level of authority and autonomy they grant these systems, and ensuring that there is sufficient and continuous human oversight.
AI is a double-edged sword when it comes to cybercrime and online safety, helping to both tackle and enable scams and harms
The widespread use of AI in telecoms poses risks for end users’ privacy, safety and security. Chatbots process vast amounts of personal data, which may include sensitive or financial information. Transparency is another key concern, requiring operators deploying AI to ensure that users are fully informed about what data is being collected and how it is being used. AI systems can also exhibit bias, which could lead to issues such as misidentifying fraud or the unfair disadvantage of specific groups. Operators should prioritise fairness during AI implementation, putting measures in place to identify and mitigate systemic biases. AI decision making is often referred to as a “black box”, meaning users can see the final output but not the internal reasoning process. This lack of transparency could have serious implications, especially with AI agents. Users could face difficulty contesting decisions made, or actions taken, on their behalf, or determining what data the agent has accessed.
AI is a double-edged sword with respect to fraud. Given the high amount lost to scams each year, the potential for AI to detect and prevent scam messages or phone calls represents a significant opportunity for the telecoms sector. In the UK, Virgin Media O2 has used AI to detect and flag 1bn suspected scam and spam calls. By analysing calls from unknown numbers in real time, the system alerts customers to potential scams before they answer. Calls labelled “suspected scam” are answered 42% less often and last 89% less time than unflagged calls. The operator also piloted an “AI Granny”, a human-like chatbot designed to answer fraudulent calls, therefore wasting the scammers’ time. Although Virgin Media O2 has not deployed this tool on a more permanent basis, it used the data gathered from the spam calls to raise awareness around common tactics used by criminals to defraud consumers.
On the other hand, AI has also exacerbated cybercrime, with the World Economic Forum (WEF) suggesting that AI-enhanced fraud is up to 4.5 times more lucrative than traditional methods. AI produces more convincing scams through techniques such as voice cloning, deepfakes and highly personalised phishing or social engineering messages. Using AI can also allow criminals to increase the scale at which they send scam communications, thereby increasing the pool of potential victims.
Similarly, AI could have costs and benefits in terms of automating online content moderation. The technology can remove harmful content much faster and on a wider scale, limiting the number of people exposed to it. Further, AI can alleviate the psychological toll placed on human moderators, who often have to view graphic or distressing material. However, AI moderation has limitations. Harmful content often requires a nuanced understanding of complex social and cultural contexts, which AI may struggle to interpret. The effectiveness of AI moderation can also vary across languages, performing better in languages in which AI has more training, such as English, which could result in harmful material in other languages slipping through the net.
Generative AI could also have a negative impact on online safety, particularly as AI tools have triggered a substantial increase in synthetic content. Since AI models are trained on existing online material, and much of this material is increasingly becoming AI-generated, the quality of these models is likely to degrade over time, leading to a vicious cycle of an oversaturation of AI content, which continually declines in quality. Even more concerning, generative AI has already led to a proliferation of illegal content being disseminated online, including non-consensual intimate images and child sexual abuse material (CSAM). Ofcom is currently investigating X for potentially violating the UK’s Online Safety Act after reports that its Grok AI chatbot was being used to create and share demeaning sexual deepfakes.
Regulators have identified bias, transparency and privacy as the major risk areas for AI
Regulators have become increasingly aware of the risks and opportunities of AI for telecoms consumers. The Australian Communications and Media Authority (ACMA) was a first mover in this respect (see Figure 2), highlighting risks such as bias, a lack of transparency and AI-driven personalisation creating “echo chambers” or “filter bubbles” on digital platforms. According to ACMA, personalised AI-algorithms could make users more susceptible to misinformation, or alternatively, recommend content that is highly provocative, emotive or extreme, potentially harming individual well-being while also fuelling broader social polarisation. In 2023, the Body of European Regulators for Electronic Communication (BEREC) raised similar concerns around potential bias and a lack of reliable training data. However, it also expected AI adoption to become the norm in telecoms over the next decade, with key opportunities in areas such as network optimisation.
Arcep has arguably undertaken the most future-facing thinking, suggesting that AI, particularly search tools (such as Google’s AI Overviews), pose a fundamental threat to the open internet by reducing user agency. By generating summarised responses directly on the search results page, these tools reduce traffic to other websites, meaning that companies will need to optimise their content for AI summaries rather than traditional search optimisation. Arcep suggested that this could ultimately reduce users to a purely receptive role online, rather than allowing them to actively choose the content they consume. This is especially concerning given that it is currently not possible to disable AI Overviews on Google search, meaning that users cannot choose to disengage with AI search functions, and many will take the AI-generated summary at face value and choose not to click on any links to source material. As Arcep has noted, AI could therefore effectively dictate which content is viewed – an issue exacerbated by the black box effect of the decision-making process. In a worst case scenario, providers could pay for the prioritisation of their websites in AI search sourcing, which would pose a further threat to the open internet. Arcep’s analysis should also provide a useful steer for BEREC, whose forthcoming report on the competition dynamics of AI markets may consider the technology’s impact on the ability of end users to access and distribute information and content online, without unlawful interference or discrimination.
Taking a similarly forward-looking approach, the UK’s Digital Regulation Cooperation Forum (DRCF) – a cross-sectoral forum comprising Ofcom, the Competition and Markets Authority (CMA), the Information Commissioner’s Office (ICO) and the Financial Conduct Authority (FCA) – released a 2026 paper specifically focusing on the future of agentic AI. The DRCF flagged various risks associated with the technology, in particular regarding transparency and the potential for these agents to become black boxes. The DRCF also outlined the risk of algorithmic collusion, where AI agents might learn to achieve coordinated outcomes, irrespective of the original intent of the developer or deployer. In January 2026, Ofcom also launched a consultation on the potential impacts of AI on telecoms consumers, aiming to set out its findings in the second half of 2026.
The FCC has taken some of the most direct regulatory interventions to date
Some regulators have already taken action to protect consumers from AI-related harms. In February 2024, in the US, the Federal Communications Commission (FCC) classified AI-generated voices as “artificial or prerecorded voices” under the Telephone Consumer Protection Act (TCPA) 1991. As a result, AI robocalls now require the recipient’s prior express consent, making unauthorised calls illegal and subject to fines. This specifically targets AI-driven scam calls, such as those using AI voice cloning to impersonate others (including relatives or officials). The Telecom Regulatory Authority of India (TRAI) is considering similar restrictions that would require explicit prior consent and disclosure of AI use. AI calls would also be restricted to specific hours. However, despite the FCC’s implementation of these restrictions, Brendan Carr (Chairman, FCC) has called for regulators to adopt a “humble approach” to AI policy, suggesting that the FCC may be unlikely to impose stringent AI regulation in the near future.
There are also rare cases of regulators looking to implement AI themselves in order to actively benefit consumers. In this regard, the FCC is exploring the use of AI to improve telecoms relay services (TRS) for individuals with disabilities. The FCC’s proposal involves using automated technologies, such as automatic speech recognition to convert a hearing caller’s speech into text in real time, and text-to-speech tools to vocalise typed responses. These updates would allow some calls to be managed automatically rather than through a human intermediary. This would help improve accessibility for users who are deaf, hard of hearing or have speech disabilities, allowing them to communicate via traditional phone services. In line with the US Government’s general stance towards AI adoption, the FCC is seemingly leading the way in terms of regulating within a pro-innovation approach, as few other regulators have taken similar actions.
There is no straightforward pathway to regulating AI itself
Regulators must navigate the fine line of promoting AI innovation and managing the associated risks. As AI is adopted across the telecoms customer journey, applications will inevitably carry different levels of risk, and any resultant regulatory attention should be proportionate. This tiered approach is already evident in some frameworks, such as the EU’s AI Act, which outlines four risk categories: minimal risk (unregulated); limited risk (requiring transparency); high risk (regulated); and unacceptable risk (strictly prohibited) – see Table 1.
Compliance with obligations under the act may be challenging, however, in cases where certain applications fit into multiple risk categories. For example, an AI customer service chatbot would typically be classified as limited risk, requiring that users are informed that they are interacting with AI. However, if a chatbot were used by bad actors to target and manipulate vulnerable individuals (such as elderly people) into excessive spending, its classification would escalate to unacceptable risk, which is prohibited. Operators must therefore carefully consider AI deployments, implementing guardrails to prevent applications from migrating into higher risk categories. Regulators should be equally vigilant. As the designated national authority for the enforcement of the act, several regulators, such as ANACOM, BNetzA and Comreg, should be well-positioned to understand the role, and handle the implications, of AI in telecoms given their long-standing oversight of the sector.
Beyond the act, multilateral standards or guidelines could also serve as a model for regulators to follow, particularly where they seek to balance the potential consumer benefit against concerns such as privacy and fairness. In January 2025, the Global Coalition on Telecommunications (GCOT), which is made up of Australia, Canada, Finland, Japan, Sweden, the UK and the US, published six voluntary principles for responsible AI adoption in telecoms, including:
Innovation and competition;
Transparency, explainability and human oversight;
Privacy;
Fairness;
Security and resilience; and
Environmental sustainability.
However, while these principles could represent a blueprint in that they reflect growing international consensus, their voluntary nature could limit their practical influence. Regulators will therefore face a challenge in implementing frameworks or measures to protect priorities such as transparency, fairness and human oversight without stifling innovation. Given the rapid development of AI, proactive engagement with emerging issues will be essential for regulators to remain one step ahead and limit potential harms. Regulators should also monitor the impacts of existing interventions (such as those by the FCC) and engage in cross-border dialogue to learn from pioneers in this space.
At the same time, over-regulation remains a potential risk. Certain aspects of AI adoption, such as enforcing strict regulations on AI search tools to prevent threats to an open internet, could be both challenging and unpopular, potentially being perceived as overly prescriptive and anti-innovation. Further, the sheer scale of LLMs – some with up to one trillion parameters – and their black box nature mean that they are inherently opaque, making their precise operations extremely difficult to understand and therefore regulate. Ultimately, regulators must try to balance the risks of over- and under-regulation, allowing consumers to benefit from the opportunities that can come from AI while mitigating any risks.
