‘Frontier’ telcos like Vodafone, AT&T, and Telefónica are deploying hundreds of AI agents across their operations, says Microsoft – to automate processes, accelerate sales, and drive a new kind of operational agility. It has a weeks–to-minutes use-case with Vodafone. But telcos are also slowing in the AI race because of three big problems.
In sum – what to know:
Days to minutes – Vodafone is using an AI agent to automate parts of its RFI/RFP process, producing drafts in minutes instead of the 10–30 days typical in many telcos.
Agents everywhere – Microsoft says ‘frontier’ firms like AT&T and Telefónica have deployed hundreds of agents everywhere from sales and finance to network ops.
Barriers to scale – Data governance, enterprise skills, and agent management are emerging as the key obstacles telcos must solve to unlock the full ROI of agentic AI.
Here’s a good business case for agentic AI in telecoms – as deployed by Vodafone in the UK, as told to RCR by Microsoft at MWC in Barcelona. It is drawn from an expanding roster of Microsoft projects with ‘frontier’ telcos, all-in on AI, which also (anecdotally) includes AT&T and T-Mobile in the US and Telefónica in Spain. On paper, this stream of work with Vodafone might just be the most powerful and elegant example of how AI goes to work for telecoms. “In most telcos, you might be talking 10 working days, maybe upwards of 30. Whereas this takes minutes.”
So says Rick Lievano, chief technology officer for telco industry clients at Microsoft. “Ten days is unacceptable in today’s world,” he says, in the back room of a busy stand in Barcelona. “These frontier companies are the templates, for sure – having expanded outside of their contact centers. Because AI is not just the contact center anymore. AT&T has deployed over 400 agents across its business. Same with Vodafone.” And so we hear about Microsoft’s work with Vodafone Business to automate the UK firm’s RFI/RFP process – at least to the point a draft is on the table.
Lievano says: “These product catalogs are exceedingly complex, right? Especially B2B catalogs. This is not an e-commerce catalog with a few hundred items; this is an enterprise telco catalog with tens of thousands of them. Because any offering, with any tweak, becomes a new product. So they are complex, and these incoming RFIs/RFPs request this QoS and this SLA, and mandate a new fiber circuit or a wavelength, say, can’t go via this geography; and all of that has to be selected and documented through the response process.
“What happens today is that RFI/RFP sits in a queue until someone picks it up, has a coffee, defines a response, goes to lunch, and picks it up again three days later from another queue of tasks; and then they start to respond. They review a library of similar requests, a set of knowledge-based articles, all the procedures for how to respond, all the proper checkpoints – like the manager associated with that account. So it is a really long process – from first receipt to first draft. Whereas Vodafone has an AI agent to pick all of that up, and start the process autonomously.”
He explains: “It receives the request, understands the request, checks the system, notifies the stakeholders. And it doesn’t stop there. It has access to the same information: company policies, boilerplate text, past responses, pricing information, product catalog. You name it, it has access; and it produces a draft, 90-percent finished, in just minutes – which shows what’s standard and what’s changed, and hands over to the account manager to see it out, to ensure compliance and pricing. The point is it is not starting at zero anymore.”
And he adds: “It means you can be much more agile – and agility is money. You can be faster than your competitors – and so, guess what, you can be better positioned.”
AI FOR EVERYONE
More than that, you don’t even need much training, it seems. Low-code / no-code software tools (like Copilot Studio) offer a graphical “drag-and-drop” coding platform with a generative AI language interface so sales directors and account managers, say, can dictate terms on their terms – to create such an RFI/RFP agent from scratch. It prioritizes domain knowledge both ways, in a parallel of the old IT/OT discussion – so the deep tech is abstracted and democratized (and popularized / scaled), and put to work in service of the specialist business knowledge.
“Agents won’t be created by engineers. Engineers know their own domain – how to run a network, how to write code. They don’t know the business. The best people to create agents are the people doing the work, who need the help. Which is the idea with copilots – agents, in multiples – in support of pilots. The pilot is the sales rep – who knows the processes, documents, policies. A software engineer in a lab doesn’t know any of that. And now they have the tools to create the agent, and choose how it’s triggered – whether it’s conversational, or attached to an inbox.”
There are a “million ways” to trigger an agent, says Lievano. Again: total flex for total opportunity, within the limits the tech affords. The Vodafone case is a neat example of a decent telco fix – “to drive an RFP to completion, or at least to a first draft”, he says, before it enters a more familiar workflow – where a human “reads it, confirms it, approves it, and moves it to the next phase, to yield a customer offer”. As an anecdote, it also articulates a bigger story about the rise of AI agents among these so-called ‘frontier’ firms (over 400 at AT&T, plus whatever the rest are doing).
The demos on Microsoft’s booth (as demo’d to RCR) are more expansive; there are lots of AI agents, talking together variously, to help with trouble shooting, ticket management, field operations – to drive autonomy in the network itself. Leivano reflects: “The contact center was the front door, and it was only natural you’d see generative AI bloom there because it is a treasure trove of language. But that was the lab, where the experiments began. The frontier telcos are expanding everywhere – into sales, finance, HR, legal, and into more specific telco domains like the network.”
The “messaging”, he says, is all about the “ROI on intelligence and trust”. Which, very clearly, is a downstream conversation relative to the industry’s early call-centre experiments, and a sign, conversational or commercial, that the industry is maturing, and looking to stack AI use cases inside their operations. “Every telco is saying, ‘Hey, I’m investing, I’m driving, I’m creating pilots; so show me the ROI’. At the end of the day, that’s what it’s about. We totally hear you; we understand. And we are providing a set of guidelines based on best practices at these frontier telcos.”
But MWC is a telco show, of course, and Microsoft has designated ‘frontier’ firms everywhere, in every industry. Telecoms has been good at AI (or good at pursuing AI efficiencies), traditionally; but it is not such a leader, anymore. The firm has a research paper with analyst house IDC that polls 4,000 business leaders from all sectors, and says ‘frontier’ firms have put AI to work in seven different business functions, on average, and two third (67 percent) are monetizing “industry-specific” AI in one way or another (58 percent use “custom AI”, rising to 77 percent by 2028).
AI AGENTS TRIPLE
In two years, the number of frontier companies using agentic AI will “triple”, the report suggests (without clearly giving a base-line start). The pair has a best-practices presentation, as well – also worth reading. More helpful, Lievano says telcos have seen a 2.8-times return on their agentic AI investments so far. But telcos are slowing, generally or relatively, and every other industry is accelerating. Lievano comments: “Initially, telcos were some of the first to adopt generative AI specifically, but… we don’t see that same acceleration that we saw two years ago.”
On the stand, Microsoft says the par score for network autonomy, early 2026, is somewhere between two and three – out of five on the TM Forum scale (Levels 1-5) for core-operational telco AI. It is worth noting that even claims from frontier firms about Level 4 capabilities, with a high-degree or ‘self-organizing’ and ‘self-repairing’ are mostly siloed in certain operational functions, siloed in certain operating companies. Hence, the importance of course to talk about best practices, also getting into top-line revenue generation, at the likes of AT&T, T-Mobile, Telefónica, Vodafone.
Hence why the message about the relevant simplicity, in use and design, is important, too. Leivano comments: “They are embedding AI across their businesses. We might talk to the networks team, and we might talk to the finance team. Different conversations, same foundation – you design in a similar manner, even though the agents are grounded and instructed differently, and connected to different data and different tools to do different things. But the architecture is very much the same.” But there are obstacles, as well – three “compounding” ones, says Lievano.
The first (and last; isn’t it always?) is data – or orchestration of it. “People say the problem is data silos, but that’s not really the issue; it’s governance.” Telcos are rich in data, but such abundance masks a structural problem: the information is not consistently defined, trusted, or connected across systems. Customer records might live in CRM platforms, billing platforms, product catalogs, and network databases – each with different identifiers and naming conventions. What one system calls a ‘customer’, another might call a ‘subscriber’, says Lievano.
Functionally they refer to the same entity; structurally they are treated as different things. The challenge is less about where the data sits – whether in the cloud, whichever cloud, or on-premise; or increasingly, distributed across the continuum with multiple vendors – and more about how it is discovered, understood, and managed. AI agents, like every horizontal digital-change tech, only work as well as the information they can access and trust. If the systems cannot reconcile identities, schemas, and permissions, the agent can’t reliably stitch together the right picture.
And then they lose accuracy and trust, and the house comes down. The answer, says Lievano, is not to centralize everything in a single repository. “We have no expectation that all the data moves onto Azure,” he explains. Instead, the aim is to connect to it wherever it lives, derive its structure and metadata, and build a semantic cloud-SaaS layer (Microsoft Fabric) on top that reconciles relationships between datasets. In that sense, the governance challenge is about creating an operational ontology – a shared understanding of what the data means and how it relates.
AI GOVERNANCE
The second obstacle is the skills gap, discussed already. Lievano adds: “Like we discussed: anybody can create an agent. You don’t have to be a developer. You don’t have to have an Azure subscription. If you have a Microsoft 365 subscription, which is 90 percent of telcos, or more, then you already have access to everything you need to create and host the agent – without needing anything else. So anybody can self-serve the agents they need.” Again, it is not a knee-jerk overhaul, about hiring AI techs, but about prioritizing in-house skills to make the tech solution scale.
Although, there is plenty of irony in such a position, given the headcount reduction in telecoms right now, to a greater or lesser degree enabled by AI-geared efficiency drives. But that is a separate discussion. For Microsoft, the pitch is that low-code / no-code tools are easily accessible and easily usable for operational teams to embed knowledge and build agents. Which leads directly to the third challenge: governance again – but this time for the agents themselves. This is what Cisco talked about at MWC, also, and where it is seeking to hang its hat in the new AI sales discourse.
Lievano says: “Right now, IT groups are horrified that all these agents are being unleashed across their enterprise systems and networks. They have no visibility into what they are doing.” Questions quickly multiply. How are they identified? What can they access? What are they using? What are they costing? Without control, agents will swarm and riot, effectively; without proper governance, their management will be impossible. Expensive AI will do more harm than good, and trust will be eroded. That kind of message.
Microsoft’s answer is an emerging management framework (Microsoft Agent 365; in preview now, for launch “soon”), a centralized catalog where enterprise agents can be registered, monitored, and controlled regardless of where they originate. IT administrators can assign identities and permissions, track activity, and manage costs – while still giving reign for the wider organization to create and deploy agents as required. “That was the missing link to enable us to go to the next level of the agentic revolution where we now have thousands of agents being deployed.”
These frontier telcos, which Microsoft has anointed, have been grappling with this problem, Lievano says. Many are building their own governance frameworks as the number of agents grows into the hundreds. The aim now is to standardize that control layer so companies can scale the model safely. ““We’re saying, ‘No, you don’t need to do that; we’re going to provide the capability.’ We’re working with many telcos that have tried to address the problem, and taken their feedback into the development of Agent365; they’re among our first customers.”
Broadly, the enterprise conversation about AI has matured, says Lievano. “It’s not really about whether you should be creating agents anymore. That question has been answered. The discussion now is: how do I manage them, how do I control them, and how do I make sure they are delivering value?” Which, in the end, takes us back to Vodafone’s RFP example. The tech is impressive (minutes instead of weeks), but the lesson is structural. If operators can solve the governance questions, the business solution is actually pretty simple: agility. Which is money – like the man says.
