Ten agents for every seller, and most sellers will not feel a thing

On July 28, Gartner published a prediction with the punchline built in. By 2028, it says, AI agents embedded across the commercial function will outnumber human sellers ten to one. In the same release, it predicts that fewer than 40 percent of sellers will say those agents improved their productivity. Hold both numbers in your head at once, because the pairing is the story. The analyst firm whose conference keynotes did as much as anyone to accelerate the agent buying cycle is now putting a number on the gap between what sales orgs are deploying and what sellers will actually feel, and the number is brutal. A ten-fold automation of the commercial function, and a majority of the people it was supposed to help shrugging.

Dan Gottlieb, a VP Analyst in Gartner’s sales practice, said the quiet part in the release: “Sales organizations are moving quickly toward a future where AI agents are embedded across the commercial function, but more agents will not automatically mean more productivity.” His warning to sales leaders is that without the right data foundation and seller experience, they risk “agent sprawl, with more digital activity, but little improvement in seller impact.” Agent sprawl is a usefully sticky coinage, and I expect it to age the way “shelfware” did, because it names something every rep on a modern floor has already met and nobody had a word for. The pipeline-scoring agent that marketing bought. The call-summarizer that came bundled with the conversation-intelligence platform. The follow-up drafter someone in ops turned on in March and nobody has looked at since. The meeting agent that posts summaries to a channel nobody reads twice. Each one arrived with a business case. Almost none of them arrived with an owner.

Sprawl is what buying looks like when nobody owns the count

The adoption numbers from this spring explain how an org gets to ten agents per seller without ever deciding to. KPMG’s Q2 Pulse survey, run between April 28 and May 25 across 204 US executives at companies with a billion dollars or more in revenue, found employee adoption of AI agents jumped from 23 percent to 56 percent in a single quarter. That is one of the steepest enterprise adoption curves anyone has measured for anything. The same survey found that while 66 percent of those organizations have monitoring dashboards and 61 percent have approval processes, only 26 percent report full real-time visibility into what their AI systems cost to run. Read that as a purchasing pattern rather than a technology story. Companies more than doubled their agent footprint in ninety days, built the approval paperwork, and skipped the meter.

The supply side is pushing just as hard. HubSpot’s Q2 earnings, reported August 5, showed its Data agent passing 16,000 activated customers, up 80 percent quarter over quarter, its Prospecting agent near 17,000, and more than 55 percent of its Professional and Enterprise customers using agents or its Breeze assistant. Every major vendor in the sales stack now ships agents as the headline feature, which means a sales org acquires agents the way a house acquires cables: one plausible purchase at a time, no inventory, and one day a drawer that will not close.

The problem is that an agent is closer to a very junior hire than to a feature. It acts. It sends things, scores things, drafts things, logs things. A feature that nobody uses just sits there costing its license fee. An agent that nobody owns keeps producing output, and its output lands on somebody, and in a sales org that somebody is usually the rep. Ten unowned agents per seller is not automation. It is ten streams of unaccountable activity converging on the one person in the building whose calendar was already the scarcest resource.

What the sub-40 percent actually measures

The prediction that fewer than 40 percent of sellers will report a productivity gain sounds like an indictment of the technology, and I do not think that is quite what it is. Sellers judge a tool by one test: did it hand me back hours I can spend selling, or did it hand me new work dressed as help? A surprising share of what gets deployed as seller productivity fails that test on contact. The agent that drafts follow-up emails still needs every draft read, because the rep’s name is on it and the rep has been burned by a confidently wrong paragraph before. The agent that enriches accounts produces a richer record the rep must now reconcile with what the champion actually said on the call. The scoring agent reorders the queue, and the rep, who has a decade of pattern memory the model does not, quietly re-sorts it back. The outreach agent drafts a sequence the rep rewrites from scratch, because the agent has never heard this account’s voice on a call and the rep has.

It helps to picture what the agents that clear the bar actually look like, because the productive minority share a shape. They sit at a point in the workflow where the rep was doing pure retrieval or pure transcription, work with no judgment content at all, and they finish the job rather than starting it. An agent that assembles the full pre-call brief, account history, open tickets, last three conversations, and the champion’s recent public statements, into one page the rep reads in ninety seconds has replaced forty minutes of tab-hopping with no review burden, because the rep was always going to sanity-check a brief anyway. The failing majority sit one notch higher, at points where the work had judgment in it, and they convert the rep from author into editor without removing the underlying cognitive load. Editing a wrong thing into a right thing is not faster than writing the right thing, for anyone experienced, and sellers are experienced. The placement decision, which rung of the workflow the agent sits on, predicts the felt outcome better than the sophistication of the agent does, and placement is a management decision, which is precisely why this is a leadership story and not a procurement story.

Gottlieb’s phrase draws the exact line that matters: more digital activity, little improvement in seller impact. Activity is what the dashboards count, and by the dashboards, agent-heavy orgs look magnificent. Touches are up, response times are down, coverage is total, and every activity chart slopes the right way. Impact is what the seller feels at 6 p.m., and the seller’s accounting is different. Review work is still work. Supervision is still work. An agent that converts one hour of writing into forty minutes of checking has saved twenty minutes and will be experienced, correctly, as marginal. An agent that acts on an account without the rep knowing has not saved time at all. It has created a small diplomatic incident the rep discovers later, in the buyer’s tone.

I want to be fair to the other side of this, because the sub-40 number can also be read as a measurement problem. Sellers are not neutral witnesses to their own automation. A rep whose prospecting is now half agent-sourced has reasons to under-report the help, the same way nobody credits the CRM for a closed deal. Some of the felt disappointment will be status protection rather than honest accounting. But even discounting for that, a world where the median seller cannot feel ten agents’ worth of assistance is a world where most of those agents are pointed at the wrong work.

Count agents the way you count headcount

The operational fix follows from taking the ten-to-one ratio literally. If agents are going to outnumber your sellers, then the agent fleet is an org within the org, and it deserves the basic management primitives you would never skip for humans. Start with a census, because most sales leaders cannot currently produce one. Every agent acting anywhere in the commercial function, listed, with what it touches and what it costs to run. The KPMG finding that only 26 percent of enterprises can see AI costs in real time suggests the census alone will produce surprises, and the surprises will be recurring line items.

Then give every agent a named human owner, and make ownership mean something specific: the owner reads a sample of the agent’s output on a schedule, answers for its mistakes, argues for its budget at review time, and can state what it costs to run without looking it up. An agent nobody will claim is an agent you retire, on the grounds that if its output mattered, someone would have fought for it. Retirement needs to be a standing process rather than a cleanup project, because the acquisition side never stops. Vendors will keep shipping agents into your stack whether you ask for them or not, and a fleet that only grows is how you arrive at Gottlieb’s sprawl with a dashboard that says everything is fine.

The metric that should govern all of it is the one almost nobody instruments: seller-felt time. Ask reps, on a cycle, which agents returned hours to them and which added review load, and believe the answers over the activity data. The Gartner prediction is, at bottom, a forecast of that survey’s results across the whole industry, and a leader who runs the survey locally gets to beat the forecast. The orgs that land in the productive minority will not be the ones with the most agents. They will be the ones that treated agent count as a cost to justify rather than a progress bar to fill.

There is a sequencing discipline hiding here too. Gottlieb’s release names the data foundation as the thing that separates sprawl from productivity, and every rep knows why. An agent multiplies whatever record it reads, and most CRMs are archaeological sites, layers of half-true fields deposited by generations of reps under duress. Pointing ten agents at that substrate does not automate the sales motion. It automates the fiction, at scale, with confidence. The unglamorous work of making the account record actually true is worth more right now than the next three agents on the roadmap, and it is exactly the work that gets deferred because no vendor sells it as a launch.

There is a mirror version of all this for anyone whose job is selling agents rather than buying them, and it is worth a detour because the two sides of the market are about to collide. If Gartner’s ratio is even directionally right, your product is not entering a green field. It is entering a fleet, joining nine other agents already competing for the same seller’s tolerance, several of them abandoned, at least one of them actively resented. The buyer’s unspoken question has quietly changed from “what can this agent do” to “why will this one be different from the seven we already stopped checking.” The vendors who win that conversation will be the ones who walk in with an answer to the sprawl itself: a named owner model, a measurable seller-felt-time claim, and a willingness to help the buyer retire something else to make room. Selling an agent in 2026 without acknowledging the fleet it joins is like selling a meeting without acknowledging the calendar. The scarce resource was never capability. It was attention, and the incumbent claim on the buyer’s attention is the sprawl you are asking them to add to.

The part of the prediction I believe

I hold analyst predictions with a loose grip, this one included. A 2028 number is a directional bet wearing a decimal point, and Gartner has an obvious commercial interest in naming the disease after selling tickets to the party where everyone caught it. The ten-to-one ratio might land at six, or fifteen, and “fewer than 40 percent” is the kind of forecast that can never be graded cleanly because the survey that would grade it belongs to the forecaster.

But the shape of the claim seems right to me, and the reason is economic rather than technological. Agents are cheap to add and expensive to notice. Adding one is a checkbox in a vendor console; noticing what it actually does to a seller’s week requires management attention, which is the one input that did not get cheaper this year. Any system where expansion is nearly free and evaluation is costly will overshoot, and the overshoot will look exactly like the release describes: a commercial function humming with digital activity, sellers whose quota attainment has not moved, and a leadership team honestly unsure which of its dozens of agents are earning their compute.

The sellers themselves have a move available, and it is worth naming because most of the coverage treats them as passengers. A rep can run the census on their own book. Which of the agents touching my accounts do I actually trust, and which do I quietly route around? Bringing that answer to a sales leader, with examples, is the most useful piece of upward feedback available this year, because leadership is currently buying blind and knows it. The rep who can say “these two agents return me four hours a week, and this one nearly damaged my largest renewal” is doing the evaluation work the org skipped, and evaluation is scarce.

The ratio is coming either way. Vendors will ship the agents, procurement will approve them, and by 2028 some version of ten-to-one will be true in most orgs that sell software. Whether the sellers feel any of it was ever the optional part, and it gets decided by inventory discipline, named ownership, honest measurement, and the willingness to kill agents that only produce activity. Gartner’s release reads to me less like a forecast than like a fire inspector describing next year’s fire while pointing at this year’s wiring. The wiring is being installed right now, one plausible purchase at a time, in orgs that mostly cannot tell you what is already running. The count was never the achievement.

Ten agents for every seller, and most sellers will not feel a thing was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.