May 21, 2026
40% of AI Service Interactions Now Complete Without a Human. What Does That Mean for the Other 60%?
When an AI agent resolves a customer’s issue without any human involvement, it counts as a win. However, as those wins accumulate, a hidden problem is taking shape on the other side of the interaction log.
Salesforce’s latest State of Service research based on responses from 3,075 service professionals, finds that 40% of case resolution interactions involving AI are now completed entirely autonomously. It also means that 60% of interactions still require a human to step in, and the nature of that work is changing in ways most organisations have not fully reckoned with.
The Work Left Behind
The interactions AI starts but cannot finish are cases that involve enough complexity, ambiguity, or emotional weight that a human agent still needs to step in. As AI takes on greater volumes of routine work across service channels, the interactions left for human agents are becoming increasingly difficult by default. Straightforward queries are resolved autonomously, but everything else waits for a human.
Agentic AI adoption has risen from 39% to 66% of service organisations in a single year, according to the Salesforce report. The pace of that expansion means many organisations are redesigning how their teams function in real time, without much runway to prepare.
The data shows that 97% of service representatives are engaged in some form of upskilling, through certifications, online courses, conferences, which suggests awareness of the problem. But awareness and readiness are not the same thing.
A Different Kind of Job
The skills Salesforce identifies as most important for the emerging role of the human agent are AI oversight and judgement (knowing when to trust or override AI outputs), adaptability, and complex problem-solving. These are not skills that most customer service training programmes have historically prioritised. Until recently, contact centre work was optimised around volume and speed. The job is now moving towards something closer to triage and exception management which requires a different approach.
Customers still prefer human agents for emotionally charged or high-stakes interactions, which means the 60% of cases that do reach a human are likely to include a disproportionate share of difficult conversations. Agents handling those interactions need to be equipped in terms of the emotional and cognitive demands involved.
The Knowledge Problem
There is also a risk that organisations underestimate how much institutional knowledge gets embedded in routine interactions. When AI absorbs the volume, it also absorbs the repetition that once helped new agents build familiarity with products, customer behaviour, and edge cases. The learning curve does not disappear, it only becomes harder to design around.
The Salesforce report notes that 92% of service leaders with AI say it improves their ability to coach at scale, which is encouraging. But coaching assumes there is something stable to coach toward.
Human Agents Need a Defined Role
Humans must own customer understanding and experience design, with AI used selectively to reduce effort and build confidence. If organisations treat human agents as a fallback for cases AI cannot handle, rather than as skilled practitioners with a distinct and valued function, they risk hollowing out the very capability they need most when things go wrong.
The operational case for agentic AI is not in doubt. Seventy percent of organisations report measurable value within 60 days of deployment. Customer satisfaction is the top KPI improved, ahead of handle time and first-response metrics. The direction of travel is not going to reverse.
