Most organisations did not decide to deploy AI agents. They just kept extending the pilot until production became the only honest description.
According to IBM's 2025 CEO Study, which surveyed 2,000 chief executives across 33 countries and 24 industries, 61% confirm they are actively adopting AI agents and preparing to scale them across the organisation. Deloitte's 2026 State of AI in the Enterprise report, drawing on responses from 3,235 senior leaders across 24 countries, finds that close to three-quarters of companies plan to deploy agentic AI within two years. One in four already has more than 40% of their AI experiments operating in production. A further 54% expect to reach that level within six months.
This is not a wave approaching. It is a wave that arrived while most organisations were still debating what sort of surf lessons to book.
What changed
Agentic AI crossed from demonstration into deployment because the underlying infrastructure finally caught up with the ambition. Multi-agent orchestration, the ability to chain AI models together so they plan, act, and adapt across complex workflows, moved from research papers into enterprise toolkits during 2025.
The clearest signal is the Model Context Protocol (MCP), an open standard that allows AI agents to connect to and operate external tools, APIs, and data sources. Launched in November 2024 with approximately 2 million monthly SDK downloads, MCP reached 97 million monthly SDK downloads by March 2026. For comparison, the React npm package took roughly three years to reach 100 million monthly downloads. MCP got there in 16 months.
When agents can reliably call APIs, query databases, execute code, and hand tasks to other agents without a human in the loop for each step, a great deal of what previously required manual effort starts to look like automation waiting to be switched on. The production threshold gets crossed not because anyone makes a formal decision to cross it, but because the use cases stop looking like demos and start delivering measurable returns.
MCP is now supported as a standard by every major AI provider: Anthropic, OpenAI, Google DeepMind, Microsoft, and Amazon Web Services. This cross-industry adoption means the tooling ecosystem is building on stable, vendor-neutral infrastructure rather than proprietary integrations that need to be rebuilt when platforms change.
The governance gap that nobody is talking about loudly
Deloitte's figures contain one number worth reading slowly: only 21% of companies planning to deploy agentic AI report having a mature model for agent governance.
Read that the other way. Roughly four in five organisations are deploying, or plan to deploy, autonomous AI agents that will take actions, make decisions, and operate across their business systems, without yet having worked out how to govern that behaviour reliably.
Gartner's June 2025 analysis reinforces this. Their prediction: more than 40% of agentic AI projects will be cancelled before the end of 2027, most of them due to escalating costs, unclear business value, or insufficient governance infrastructure. This is a separate finding from their projection that 40% of enterprise applications will feature embedded AI agents by end-2026, which reflects the apps side of deployment.
Both predictions can be true simultaneously. A technology can be in widespread production deployment and still have a failure rate that makes a thoughtful person pause.
A technology can be in widespread production deployment and still have a failure rate that makes a thoughtful person pause.
The organisations that navigate this gap well are not the ones that avoided deploying agents. They are the ones that built governance frameworks alongside deployment: clear definitions of what agents are authorised to do, audit trails for agent actions, and humans in the loop for the decisions where the cost of an error is high.
Where this trajectory ends
Gartner projects that by end-2026, 40% of enterprise applications will have embedded AI agent capabilities, up from under 5% at the start of 2025. Gartner's longer-range analysis suggests agentic AI could drive approximately 30% of enterprise application software revenue by 2035, an estimate they place at over $450 billion.
These numbers describe a technology becoming structural to how enterprise software works, not a feature layer on top of it. The difference matters for how you think about building skills and operational models around AI.
Deloitte also found that worker access to AI rose 50% in 2025. That figure does not describe AI replacing workers. It describes AI being distributed to workers at scale, and the organisational question shifting from whether to use AI to how to use it effectively.
What this means if you are paying attention
For most professionals, the appropriate response to this data is not anxiety about autonomous agents. It is recognition that the organisations setting the pace right now are not the ones with the largest budgets. They are the ones where individuals understand how to use these tools well, and where leaders understand how to structure work around them.
Understanding what MCP is and what it enables, knowing how to design a multi-step agent workflow, being able to identify which tasks are worth delegating to an autonomous system and which require human judgment: these are no longer advanced topics for specialists. They are the operating knowledge required for anyone who wants to work effectively inside an AI-augmented organisation.
IBM's CEO data found that while 61% of executives are actively scaling agents, they are moving faster than their technical and governance infrastructure can support. The constraint is not ambition. It is not budget. It is people who understand the technology well enough to build it properly, govern it sensibly, and get value from it before the next governance incident forces a rethink.
The pilot phase is over for the organisations running the surveys. The question is whether it is over for you.