Ask a large company whether it is trying AI agents and the answer is usually yes. Ask whether agents run every day across the business and the answer changes.
Trying is common. Scaling is not.
McKinsey's 2025 survey of 1,993 respondents found that 23% were scaling an AI agent system somewhere in their organisation and a further 39% were experimenting. That is 62% at least trying. KPMG's tracking of US firms with $1bn or more in revenue shows the share that had deployed agents rising from 11% in early 2025 to 26% by the end of the year. Deloitte found that only 25% of organisations had moved 40% or more of their AI pilots into production, and only 21% said they had a mature model for governing agents.
These surveys use different definitions and samples, so they cannot be added together into one trend. They do point the same way: interest is high, everyday use is much lower.
What gets in the way
The blockers people name are not about whether the AI is clever. They are about the work around it:
Complexity and integration. In KPMG's survey, 65% named the complexity of agentic systems as the top barrier.
Data readiness. An agent that reads out-of-date prices or policies gives out-of-date answers.
Security and privacy. Agents act on what they read, so who can see and change what matters.
Unclear return. Forrester found only 15% of AI decision-makers saw a profit lift in the past year, and expects 25% of planned AI spend to slip into 2027.
Gartner has forecast that more than 40% of agent projects will be cancelled by the end of 2027. That is a forecast based on a webinar poll, not a count of cancellations. It is best read as a warning that some products labelled "agentic" are old chatbots with new branding.
Customer service is moving first
Salesforce's 2026 survey of 3,075 service professionals found 66% of service teams using at least one AI agent, up from 39% the year before, and 70% of adopters reporting value within 60 days. It is a vendor survey, so read it as a direction, not proof. It does fit what most businesses see: calls and chats are high-volume, bounded and easy to measure.
What a smaller business can do
Smaller firms use AI less than large ones. The US Census Bureau found 37% of firms with 250 or more staff using AI against under 20% of firms with four or fewer. The upside is that you do not have to untangle many pilots at once.
Pick one bounded job. After-hours calls, bookings or order status. Measure it for 60 to 90 days before adding more.
Write three rules before launch. What the agent may never decide alone, when it hands over to a person, and who reviews its conversations each week.
Check your data first. Prices, opening hours, policies and customer records must be right.
Buy a working product for a defined task rather than building your own agent platform.
Set targets in advance. Agree a resolution rate and a satisfaction score, and track cost per resolved conversation.
Keep a clear route to a person. Customers expect it: 87% say human access is essential when AI is used for service (Gartner, 2026).
Read the full report
This article is a short companion to our industry report, From pilot to production: how AI agents get beyond the demo. It has the charts, the myth-versus-reality section and the full source list.
View the full report on our Resources page →
This article summarises published research as of 23 October 2026 and is for general information only. Some figures are reported through secondary sources, and several come from companies that sell AI products.
