AI Vendors Are Winning, But Are Their Customers?

The AI industry is having a remarkable year. NVIDIA continues to report record growth, Anthropic reached a $965 billion valuation last month, and Elon Musk is even planning to launch an orbital AI data centre. From CX AI perspective, Salesforce has just agreed to pay $3.6 billion for the AI customer-service platform Fin, while vendors such as Zendesk and Genesys are positioning AI as the centrepiece of their future growth. Across the sector, the message is consistent: artificial intelligence is creating enormous value.

For enterprise customers, however, the picture is more complicated. Some are reporting genuine returns, while many others remain stuck in pilot programmes, struggling to prove the outcomes their boards are asking for. AI ROI is starting to separate enterprises from one another almost as much as it separates buyers from sellers. Ultimately, vendor and customer outcomes are intrinsically connected, but real returns may be slower to show up than vendor growth.

There is little doubt that AI is paying off for the companies selling it. Genesys closed fiscal 2026 with nearly $2.6 billion in annual recurring revenue, up more than 35 per cent year on year, with over 70 per cent of its customers now using its AI capabilities. Salesforce's AI and data business, anchored by Agentforce, already generates close to $3.4 billion in annual recurring revenue, up over 200 per cent year on year, and its newly agreed Fin acquisition adds an established agentic customer-service platform to that push.

What is perhaps most striking is that AI is no longer being treated as a feature or product category. Across the customer experience market, it is increasingly becoming the organising principle behind vendor strategy. Product roadmaps, acquisitions, pricing models and growth narratives are all being reshaped around AI, particularly autonomous and agentic systems.

Zendesk is betting heavily on this vision, recently declaring the chatbot era over at its Relate 2026 conference in favour of an Autonomous Service Workforce priced on verified outcomes. The company’s CEO, Tom Eggemeier, was explicit: “We are entering the age of the autonomous service workforce”.

Whether buyers are converting that investment into meaningful production deployments is far less settled. Forrester's State Of Agentic AI, 2026 found that three-quarters of enterprise leaders say they are adopting agentic AI, yet only a small minority have it running in meaningful production beyond what the analyst firm calls “agentish” chatbots, with scaled multi-agent systems rarer still.

Dun & Bradstreet's 2026 AI Momentum Survey points to a similar divide between adoption and outcomes. While 97 per cent of organisations report active AI initiatives, only 24 per cent say those investments are delivering broad or strong returns. The findings suggest that adoption is still running well ahead of value realisation.

Adoption, deployment and value creation are three distinct stages, and most organisations appear further along the first than the other two. Having an AI initiative is no longer unusual. Generating a measurable business outcome from it, however, still is.

One reason for this is that AI often exposes organisational weaknesses that existed long before the technology arrived. Poor data quality, fragmented systems and unclear ownership can all limit performance. Governance requirements create another challenge, particularly in regulated industries where AI outputs must be monitored, audited and explained before they can be deployed at scale.

Many organisations are also discovering that AI is not a plug-and-play productivity tool. While vendors can deploy new capabilities quickly, capturing value often requires redesigning workflows, retraining employees and redefining performance metrics. These changes take time and can delay the benefits that boards and investors expect to see.

As a result, there is often a significant gap between AI adoption and AI value. Buying AI has become relatively easy. Transforming an organisation to take advantage of it remains considerably harder.

Plenty of organisations are proving it can be done. Klarna's AI customer service assistant now handles work equivalent to more than 853 full-time agents, has saved the company $60 million, and has cut response times by 82 per cent while reducing repeat issues by a quarter. Notably, Klarna also walked back its boldest replacement claims and reintroduced human agents for complex cases, suggesting the returns came from disciplined deployment rather than wholesale automation.

JPMorgan tells a similar story. The bank has gone so far as to tie engineers' performance reviews to AI tool adoption after its internal coding assistant lifted productivity by 10 to 20 per cent across tens of thousands of developers, according to a Reuters report last year. A widely cited NBER study of 5,179 customer support agents found comparable gains in contact centres more broadly, with a generative AI assistant lifting issues resolved per hour by 14 per cent, concentrated among newer and lower-skilled workers.

Evidently, AI can generate significant returns. The real question is no longer whether it works, but why some organisations are capturing so much more value from it than others.

The organisations pulling ahead tend to share a few traits: reasonably clean, accessible data; executive sponsorship and clearly defined objectives; and, crucially, a willingness to redesign workflows around AI rather than simply bolt it onto existing processes. They also tend to measure success by business outcomes rather than activity.

Forrester's own research points to the same pattern, urging enterprises to invest in orchestration before adding more agents, redesign work rather than just tooling, and treat every agent as a governed identity with clear ownership and logging. The common thread is that these organisations treat AI as business transformation, not procurement. Technology alone rarely creates the value; the value comes from changing how work actually gets done, which is a far harder thing to buy than software.

A useful way to think about the resulting landscape is three groups. AI Winners, like Klarna and JPMorgan, are achieving measurable ROI, scaling deployments and expanding investment. AI Improvers are seeing real but incremental gains, the kind the NBER study documents, without yet reaching transformational impact. AI Tourists are running extensive pilots that struggle to reach production, the pattern Forrester describes as being caught between promise and payoff.

A qualitative study of agentic AI adoption published on arXiv found a comparable split among the companies it interviewed, with most still running simple AI assistants and only one operating genuine multi-agent orchestration in production. Most enterprises are therefore much earlier in their AI journeys than the headlines might suggest.

The first phase of the AI boom proved the technology works. The second phase proved enterprises are willing to buy it. The third phase, now underway, will determine whether organisations can consistently capture enough value to justify continued investment.

For vendors, AI is already delivering results, with growth, retention and acquisition spending all pointing the same way. For customers, the picture is becoming more nuanced rather than simply negative. Some are generating meaningful returns and pulling further ahead of their competitors, while others remain caught between experimentation and impact.

So are AI vendors' customers winning too? Some clearly are. If the gap between AI leaders and AI laggards continues to widen, it could eventually test some of the assumptions underpinning today's AI investment boom. Vendors can continue to report strong growth for a time, but enterprise spending ultimately depends on customers generating returns.

The next phase of the market may therefore be defined by how many can prove it is creating meaningful business value. If too few can, the virtuous cycle that has powered AI's rise - vendor growth, customer investment and proven business outcomes - could begin to lose momentum.