CSPs aren’t realizing AI’s full value yet

As communications service providers (CSPs) look beyond proofs of concept to widespread deployment of generative and agentic AI, IBM’s Eoin Coughlan, Global CTO for Telco, Media and Entertainment, and Rahul Kumar, Senior Partner and Global Industry Leader for Telco Media and Entertainment, emphasize that true enterprise-wide adoption hinges on strong leadership, robust governance and clear evidence of business value.

“Proof of concept, by definition, says it’s a concept that has not been proven yet,” says Kumar. “Conceptually, GenAI has been proven.” Now, telcos need to shift their mindset to proof of value, he says, adding: “The word value is very important.”

Earlier this year, IBM's Institute for Business Value and Oxford Economics surveyed 106 CSP executives globally about their AI strategy, initiatives and outcomes. The results have been published in a new report, Telecommunications in the AI Era. In addition to analyzing the survey, IBM drew on telco case studies and the company’s consulting work with CSPs worldwide.

In the study, nearly two-thirds of respondents (64%) admitted that their AI initiatives have not yet delivered the full expected value. However, operators are beginning to realize some benefits from AI deployment. The graphic below shows the AI-attributable improvement respondents have seen during the past year in four high-priority KPIs.

Strategy and leadership

Kumar stresses that AI initiatives must be tied to strategic priorities. “Doing AI is not a standalone thing – you don’t do AI,” he explains. “You think in terms of what the technology [can do]. It’s AI today; it could be a different technology tomorrow.”

Telcos need to understand how the use of AI impacts business KPIs like improving customer experience and cutting opex, Kumar adds. “It’s very important for the organization to understand that, get behind it, and start doing things and thinking in a way which we call ‘AI first’,” he says.

Both Kumar and Coughlan emphasize that scaling GenAI starts at the top. “If senior leadership doesn’t make AI a priority, it won’t scale,” says Coughlan. This means that leaders need to make time for employees to identify repetitive tasks that can be automated using AI. “Let’s be honest: Most people are very busy each day and have to prioritize their workload,” says Coughlan. By offering support, leaders can encourage staff to find creative ways to use AI.

Kumar points to IBM’s own AI transformation – dubbed ‘IBM as Client Zero’ – as evidence of what’s possible with strong leadership. In 2023 the company’s CEO, Arvind Krishna, set a mandate for the executive team to make IBM “the most productive company in the world through use of AI, automation and simplification of business,” Kumar says.

Today the company is driving $4.5 billion in productivity gains as a result. This happened “through a mix of eliminating complexity, simplifying the business, extreme automation and using AI-first principles, driving AI into all aspects of our business and operations,” according to Kumar.

Governance is key

Kumar and Coughlan discussed AI governance at length in an interview with TM Forum Insight for an upcoming report on how optimizing AI governance can help CSPs scale AI. Technical governance challenges, especially around data quality and compliance, must be addressed, according to Coughlan.

“If you look at projects that do not have governance, you will see a lot of failures, because they won’t get past legal; they won’t get past a risk assessment; and they won’t go into production,” he says.

“If you don’t get the data right, the rest doesn’t matter,” Coughlan adds, noting that automation compounds the negative impact of poor data. “With any bad data that causes a bad decision, the effect of that bad decision is now multiplied because of automation,” he explains.

Coughlan advocates for a federated data architecture and centralized governance platforms that can monitor model performance, compliance with regulations like the EU AI Act and ethical standards. “You need explainability and data lineage,” he says. “Regulators won’t accept ‘AI is too complex’ as an excuse.”

Kumar agrees and notes that while strong technical governance is essential, it’s also important for CSPs to guard against organizational governance becoming a bottleneck. “If every initiative has to go through the AI center of excellence, that complete centralization is going to choke innovation,” he explains. “It’s needed at the beginning, but then once…you’re confident about the approach, you [need to] start decentralizing.”

Agentic AI and autonomous networks

Both executives discussed agentic AI, with Kumar noting that its deployment raises strategic management questions. “How much control are we willing to cede to agents?” he asks. “There is a natural tendency [for CSPs] to be risk averse.” He adds that as agents proliferate, operators should not focus simply on automating steps in a process; rather it’s about “redefining how that work gets done”.

Coughlan highlights the need for secure agent-to-agent (A2A) communication protocols. “If we don’t get the security and access privileges right between agents, you have nothing,” he says.

Indeed, the lack of A2A interfaces is a key reason for delays in adoption of agentic AI by telcos, along with infrastructure-related data challenges. That was the finding in another new report, this one a collaboration between IBM, TM Forum and Oxford Economics. The report, Navigating autonomous networks, is based on a survey 200 telecoms executives across six countries using a structured questionnaire aligned with TM Forum’s autonomous network (AN) maturity framework.

The research finds that 57% of telecom executives view cloud and AI as pivotal in advancing from basic automation to fully autonomous networks. However, when considered separately, only about one-third (35%) recognize the importance of generative and agentic AI.

This perception gap is reflected in operational adoption: Nearly half of respondents (55%) have moved network operations to cloud environments, yet only 19% have integrated AI into three or more network or OSS functions, which is where most closed-loop systems reside.