The question around enterprise AI is starting to change. For the past few years, businesses have been asking what they can do with generative AI. Now, the more useful question is how much real work AI agents are taking on, and whether organisations can make that work dependable enough to run at scale.
The commercial evidence is becoming harder to ignore
Recent earnings results point in the same direction, even if the companies measure AI activity differently. 8x8, for example, reported AI adoption more than doubling as CX demand accelerated, with customer adoption of its AI solutions growing 121% year-on-year. Intelligent Customer Assistant interactions were also up more than 88%.
Microsoft’s AI business surpassed $37 billion in annual revenue run rate, up 123% year-on-year, with chief executive Satya Nadella describing agents as an emerging dominant workload, with the results pointing to a broader shift in enterprise AI investment. Meanwhile, Salesforce's Agentforce reached $1.2 billion in annual recurring revenue, up 205% year-on-year, while Agentforce and Slack had together delivered 3.8 billion "Agentic Work Units" to date.
The figures are not directly comparable. Microsoft's numbers cover a much broader AI business, so they demonstrate growing demand for AI infrastructure and services rather than proving that customers are running autonomous agents. 8x8's and Salesforce's measures are more closely tied to AI usage and agent activity.
Taken together, though, they suggest a meaningful change. Businesses are not just testing AI anymore. They are paying for it, using it and increasingly putting it to work in live environments.
What Salesforce's Index shows
Salesforce's latest Agentic Enterprise Index adds usage data developing among Salesforce customers from February 2025 through April 2026. It found the average number of activated agents per organisation nearly tripled, while the time from creating an agent to putting it into use fell 53%, to under two days. Agents are also taking on more varied tasks, with the average number of unique skills each agent could act on rising from two to six by the end of 2025.
Customer service provides an even clearer example. Agents handled 170 times more customer-service chats than in previous years, while the escalation rate to human agents remained around 32%. Customer satisfaction was also the outcome most associated with deploying service agents.
This is significant because the standard for a useful CX agent is changing. The question is no longer simply whether it can hold a convincing conversation. Can it understand the customer's situation, find the right information, take an action such as issuing a refund, and recognise when it should hand the case to a person? This shift from answering questions to actually resolving problems is one CX AI News explored in more depth earlier this year.
An agent that can only answer questions is still essentially a chatbot with better phrasing. One that can look up an order, apply a policy and process a refund is doing something much closer to the work of a human service agent.
Nubank offers a useful real-world example. Its engineers recently published a framework for building customer-support agents across more than 100 million users, covering five production deployments including card delivery and debt management. Their research highlights the importance of evaluation-pipeline quality to iteration velocity, showing that putting an agent into production depends on much more than the capability of the underlying model.
Production is not the same as scale
Forrester's State of Agentic AI 2026 adds an important qualification. Roughly three-quarters of enterprise leaders say they are adopting agentic AI, but more advanced implementations remain much less common, and genuinely scaled multi-agent systems are still rare.
There is a big difference between getting one agent to work in a live environment and running a network of agents reliably across an organisation. While evidence increasingly supports the move from experimentation to individual production deployments, the more challenging step is extending that success across multiple functions, systems and workflows without losing control.
Why orchestration is the harder problem
Consider a service agent handling a billing query. Its work may depend on a sales agent that generated the lead, a knowledge system supplying the relevant policy, and the CRM and billing systems it needs to access. A person may also need to step in when the case falls outside the agent's authority.
That is a very different challenge from making one agent answer questions correctly. Getting agents, people, data and applications to hand work between them without errors requires a common approach to permissions, monitoring, evaluation and governance.
Salesforce itself has argued that this requires organisations to rethink core IT architecture rather than simply bolt agents onto existing systems. Data quality, permissions, observability and evaluation are becoming practical limits on how far these systems can be deployed, alongside the need to control what agents are allowed to do.
What this means for CX leaders
The evidence increasingly points to a real shift from AI experimentation towards production use. The earnings results show growing commercial demand, while Salesforce's Index provides a window into how agents are being used in live environments and becoming more capable.
For CX, the move from AI that can answer a customer to AI that can resolve their problem is only the beginning. The real test will be whether organisations can trust that capability when the volume, complexity and consequences of the work increase.