Ten agents per seller. One CRM they all believe.

Gartner has now made the same prediction twice, eight months apart, and the second time it brought a name for what goes wrong.

The prediction: by 2028, AI agents will outnumber sellers by ten times, and fewer than 40% of sellers will say those agents improved their productivity. Ten times the digital labor, and most of the humans it was bought for will not feel it. The name for the gap between those two numbers is agent sprawl.

What is agent sprawl?

Agent sprawl is what happens when an organization adds AI agents faster than it fixes the systems those agents read from, producing more digital activity without more seller impact. Gartner named it in a July 28, 2026 release, where VP Analyst Dan Gottlieb warned that without the right data foundation, workflow integration and seller experience, CSOs risk exactly that outcome. The agents work. The number does not move.

Search the term today and you will find it defined almost entirely by security and data governance vendors, who read it as an inventory problem: too many agents, no registry, nobody sure who owns what or what it can reach. That reading is correct and it is not the one that costs you a quarter.

Gartner published this prediction twice

This is the part worth noticing, because almost nobody has. The same 10x and sub-40% forecast first ran on November 18, 2025, under a different analyst, Melissa Hilbert, with a softer frame. Her words were “value ceiling”: beyond a certain point, more AI does not mean more productivity.

Eight months later the same forecast came back with a harder name, a different analyst, and something the first version did not have: evidence. The July release carries a survey of 210 CSOs and senior sales executives fielded from January through February 2026. A prediction that gets restated with fresh data behind it is not a content refresh. It is a firm telling you the thing it guessed at is now measurable.

Why doesn’t more automation mean more productivity?

Because agents inherit whatever your CRM already believes. Gottlieb put it in one line that should be on a wall in every RevOps team: if those systems are fragmented, the agents will scale the fragmentation.

That is the whole mechanism. A seller who opens an account record with four competing versions of the same field pauses, squints, and asks someone. An agent does not pause. It reads the field, treats it as true, acts on it, and writes the result somewhere another agent will read. Ten agents per seller means ten times the throughput on that judgment, and no one squinting.

The readiness numbers say most companies are not close. A Harvard Business Review Analytic Services survey of 1,574 enterprise IT leaders, sponsored by Cloudera and published March 5, 2026, found just 7% say their data is completely ready for AI. Gartner has separately predicted that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear value and inadequate risk controls. Those cancellations will not be model failures. They will be data failures with a model on top.

The IT definition and the revenue definition are different problems

An IT team worried about sprawl asks: how many agents are running, who deployed them, what can they reach. Those are real questions and a registry answers them.

A revenue team has a worse problem, and a registry does nothing for it. Your agents can be fully inventoried, correctly permissioned, individually approved, and still quietly wrong together, because they all read the same bad field. The risk is not the agent nobody knows about. It is the twelve agents everybody knows about that agree with each other about something false, and produce a forecast, a routing decision and a comp statement that are all internally consistent and all built on a value somebody fat-fingered in 2023.

Inventory does not catch that. Only looking at the data does.

What a context layer actually requires

Gartner’s first recommended action is to own AI-forward sales infrastructure: a centralized context layer connecting enterprise data, systems and seller judgment. It is the right instruction and it is abstract enough to be bought rather than built, which is how most of these go wrong.

Concretely, at the field level, a context layer means an agent can answer four questions about any value it reads. Which field is authoritative when three of them disagree. When was this last verified, and by what. Who owns it when it is wrong. What breaks downstream if it changes. In the instances we get called into, the usual answer to all four is nobody knows, because the field count grew for years and nothing was ever retired. One engagement we worked came out of migration carrying 1,532 fields, with common fields living in four or five lookalike versions. No context layer sits on top of that honestly. You resolve it first or the layer just formalizes the confusion.

This is also why we argue that agents should not get write access to forecast, routing or comp fields until they have earned it field by field. Sprawl is bad when agents read a broken system. It is considerably worse when they write back into it.

Where should a revenue leader actually start?

With the measurement, not the purchase. Gartner predicts that by 2028, CSOs who overhaul data, automation and user experience will be five times more likely to gain ROI from AI than those choosing quick fixes. Five times is a large enough gap that the sequencing matters more than the tool selection.

The same July survey found 60% of CSOs say their revenue number is largely driven by elements outside their control. Some of that is the market. Some of it is that the system telling them what is happening cannot be trusted, which is a different problem with a different fix. Salesforce’s State of Sales, 7th edition, a survey of 4,050 sales professionals published in February 2026, found high-performing sales organizations prioritize data hygiene at roughly 1.5 times the rate of underperformers, 79% against 54%. That is a correlation rather than a cause, and it is still the clearest tell in the category about what good teams do first.

Before you add the eleventh agent, find out what the ten you have are reading. Count how many fields feed your forecast. Check how many have a named owner. Check when each was last verified. If those answers are uncomfortable, the agents will not fix it. They will scale it, ten times, on schedule, by 2028.

More agents is a decision. A data foundation is a prerequisite.