In November 2025 Gartner published a prediction with two halves, and the second half is the one worth reading twice. By 2028 it expects AI agents to outnumber sellers by a factor of ten. In the same prediction, it expects fewer than 40 percent of sellers to report that those agents improved their productivity.
Why a forecast can contain its own failure
A firm publishing a bullish adoption forecast has no commercial reason to attach a disappointing outcome to it. That Gartner did makes the second figure more interesting than the first.
The implied mechanism is familiar from every previous wave of sales tooling. Deployment is a purchasing decision and can be made quickly by a small number of people. Productivity is a work-design outcome and requires deciding what stops, who owns what, and which measures change. The first can run years ahead of the second, and the gap is where the disappointment accumulates.
The same shape, seen elsewhere
This is the McKinsey finding in a different dress. Adoption near universal, attributable impact rare, and the difference explained by whether the work was rebuilt rather than by what was bought.
Gartner has also predicted that by 2027, 95 percent of sellers' research workflows will begin with AI, against under 20 percent in 2024. Read alongside the productivity figure, that describes an industry where nearly every seller uses the same class of tool for the same task and most of them cannot show it made a difference.
What we take from it
Our position is that the number of agents is not a measure of anything. The measure that matters is which decisions stopped requiring a person and whether that was the right set.
A pod with three automated seats and a written boundary between production and judgment will beat one with thirty agents and no such boundary, because the second cannot tell you which of its thirty outputs was worth producing. That is not a prediction about 2028. It is observable now in any system that has been running long enough to be audited.
The practical implication for a firm buying in this market is to ask what a system is allowed to decide and what it hands back, before asking what it can do. A capability list is available from every vendor. A boundary is not.