AI in sales - where agents belong and where they do not
The question a revenue leader hits after a first AI pilot is usually some version of the same one: which parts of the funnel do I actually hand to an agent? The tools can draft a thousand outreach emails in an afternoon and then sit useless when a six-figure deal stalls because the buyer got nervous. The capability is real and badly uneven, and the unevenness does not track the funnel.
The cut between agent work and human work runs along one question: does the next action need information synthesized, or a buyer's confidence transferred?
Short answer: AI agents earn their keep at every stage where the task is information-dense and relationship-light: prospecting research, lead scoring, outreach sequencing, meeting prep, and CRM hygiene. They stall where buyers need a human to reduce risk and build confidence: late-stage negotiation, multi-stakeholder consensus, and the moment a deal goes off-script. Fund agents for volume. Keep humans for conviction.
Why "which funnel stage" is the wrong question
Carving the funnel into an agent half and a human half does not work, because the same stage holds both kinds of work.
Take discovery. Pulling the org chart, the recent funding round, the tech stack, and three relevant case studies is information synthesis, and an agent does it in seconds. Sitting across from a VP who is quietly worried this purchase will be the one that gets blamed if the quarter misses is confidence transfer, and an agent cannot do it at all. Both happen in the same meeting. A stage-based split would assign the whole thing to one side and get half of it wrong.
The buyer data draws the real line. In a May 2026 survey of 645 B2B buyers, Gartner found that 69% turn to sales reps to validate AI-generated insights, and that buyers working with a rep were 28 percentage points more likely to advance to the next step and 32 points more likely to feel confident in their decision. Buyers will let a machine gather and even interpret. They want a person to stand behind the answer when money is on the line.
So the question to ask of any task is not where it sits in the pipeline. It is which of two jobs it does. Does it assemble or interpret information? Give it to an agent. Does it move a human from doubt to commitment? Keep a person on it.
What do AI sales agents own well?
Agents own the information-dense work, and the productivity case for handing it over is strong.
The research layer is already gone. Gartner projects that by 2027, 95% of sellers' research workflows will begin with AI, up from under 20% in 2024. The same study names the tasks agents handle well: account research, personalized messaging, and signal monitoring. These share a property. The output is a synthesis of data a person would otherwise spend hours collecting, and the quality bar is "accurate and relevant," not "trusted by a nervous buyer."
The chief revenue officers (CROs) running this in production land in the same place. In a SaaStr session, leaders from Databricks, Windsurf, Perplexity, and Owner described AI handling research, personalization, outreach, and CRM while humans keep judgment, negotiation, and the final stretch that needs intuition. Windsurf scaled its go-to-market team from 3 to 75 people in under a year using AI-enabled recruiting and onboarding. The volume work was the part that scaled.
ElevenLabs is the sharpest version of the same move. On Harry Stebbings' 20VC podcast, the voice-AI company's go-to-market lead Carles Reina described scaling from zero to more than $350 million in annual recurring revenue in roughly three years on a sales team built around AI agents, with reps carrying a quota of twenty times their base salary and more than 80% of them hitting it (20VC, via Business Insider, 2026). A $100,000 base means a $2 million number. A thin layer of people clears a bar that size because the agents carry the research, the outreach, and the follow-up underneath them.
The revenue follows when the agent's output reaches the rep as a recommendation rather than a report. Gartner found that sales orgs providing AI-enabled next-best-action guidance are 2.6 times more likely to achieve commercial growth. One B2B tech company described by HBR saw a 6% lift in response rates and projected $50 million in incremental annual revenue from using agentic AI to improve lead conversion. The pattern holds: point agents at the high-volume, information-heavy tasks, and the floor of the team comes up.
A caution sits next to the upside. Gartner predicts that by 2028, AI agents will outnumber human sellers by 10 to 1, yet fewer than 40% of sellers will report that agents improved their productivity. The orgs that win are the ones that gave agents the information work and stopped there. The same discipline decides which workflows to automate first anywhere else in the business.
Where does a human seller still have to be present?
A human has to be in the room wherever the next action moves a buyer from doubt to commitment.
That is a short list, and it is expensive to get wrong: late-stage negotiation, multi-stakeholder consensus, the pricing exception, and the moment a deal goes off-script and the playbook runs out. Gartner's research names the human-differentiated capabilities directly. They are empathy, judgment, contextual understanding, and value framing. None of those is an information problem. Each is the work of getting a person who is uncertain to act.
Buyer preference is moving toward the human in exactly these moments, not away. Gartner predicts that by 2030, 75% of B2B buyers will prefer sales experiences that prioritize human interaction over AI, particularly in complex or high-stakes transactions. As routine touchpoints get automated and commoditized, the human conversation becomes the thing buyers actively want when the decision is large. The scarcer it gets, the more it is worth.
This is also where the operators draw the line. The SaaStr CROs reserve the final stretch, the part that needs intuition, for people. Agents lift the floor of the team. The closer who reads a room and reframes the deal is still the ceiling, and no amount of synthesis substitutes for the person who absorbs the buyer's risk and says, in effect, I will stand behind this.
Working through this for your own org? Sign up for the newsletter to get the next piece.
How do you draw the line in your own org?
Run every funnel task through one test: synthesis or confidence. If the task assembles, interprets, or monitors information, an agent can own it. If the task exists to move a specific human from uncertainty to commitment, a person owns it.
Most tasks sort cleanly. The ones that do not are usually two tasks wearing one name. "Discovery call" is research synthesis plus relationship-building, so split it: the agent preps the brief, the rep runs the conversation. "Proposal" is document assembly plus value framing, so split it: the agent drafts and prices the body, the rep owns the narrative and the room it is presented in. The test does not assign a meeting to one side. It assigns the actions inside the meeting.
The frameworks already in market support the same cut without naming the mechanism. BCG describes three levels of AI involvement in selling: augmented, assisted, and autonomous, and reports that roughly 7 in 10 sellers already use general-purpose AI for tactical tasks while managers say its potential to improve deal conversion and customer relationships is largely untapped. McKinsey's four highest-value B2B gen AI use cases are meeting support, next-best-action guidance, RFP response automation, and smart pricing, with a $0.8 trillion to $1.2 trillion productivity prize across all of sales and marketing. Every one of those four use cases is information synthesis feeding a human decision. The line is the same. The data just keeps confirming it.
The synthesis-or-confidence map
Use this to sort your own funnel. The left column is the task. The middle is which job it does. The right is the owner.
| Funnel task | Synthesis or confidence | Owner |
|---|---|---|
| Prospect research and list building | Synthesis | Agent |
| Lead scoring and signal monitoring | Synthesis | Agent |
| Outreach sequencing and follow-up | Synthesis | Agent |
| Meeting prep and account briefs | Synthesis | Agent |
| CRM hygiene and forecast prep | Synthesis | Agent |
| Proposal and RFP drafting | Synthesis | Agent drafts |
| Discovery conversation | Both | Agent preps, rep runs |
| Validating an AI-generated insight | Confidence | Rep |
| Multi-stakeholder consensus | Confidence | Rep |
| Late-stage negotiation and pricing exceptions | Confidence | Rep |
| A deal that goes off-script | Confidence | Rep |
One number is worth flagging before you set a hard rule. Some vendors report figures like 73% larger deal sizes and 80% higher win rates with AI assistance; those are self-reported by the vendor's own customers, not independent research, so treat them as directional. We have found no empirical source for a deal-size threshold above which human-only closing wins. Where you set that line is editorial judgment, not a published finding. Set it from your own win-loss data, not a benchmark.
Frequently asked questions
Which sales tasks should AI agents handle first? Start with the information-dense, relationship-light work: prospect research, lead scoring, outreach sequencing, meeting prep, and CRM hygiene. Gartner projects that by 2027, 95% of sellers' research workflows will begin with AI. These tasks have the clearest synthesis output and the lowest trust requirement, so the handoff is low-risk and the productivity gain is immediate.
Will AI agents replace sales reps? No, though the ratio shifts hard. Gartner expects agents to outnumber sellers 10 to 1 by 2028, while 75% of B2B buyers will prefer human interaction in high-stakes deals by 2030. Agents absorb the volume work. Reps move toward the confidence-transfer moments, the negotiations and consensus-building that buyers still want a person to own.
How do I know if a sales stage is ready to hand to an agent? Run one test. If the task assembles, interprets, or monitors information, an agent can own it; if it exists to move a specific buyer from doubt toward commitment, keep a person on it, because that is confidence work and a machine cannot do it. Most stages contain both. Split the stage into its tasks rather than assigning the whole thing to one side.
Why do AI sales tools often fail to improve productivity? Gartner found that fewer than 40% of sellers will report agents improved their productivity even as agents outnumber them 10 to 1 by 2028, and the usual reason is aim: the agent gets pointed at a whole funnel stage, or at the confidence-transfer work, instead of the information synthesis it is good at. Agents pay off on synthesis. Aimed at trust, they stall.
What does an AI-native sales team look like organizationally? Agents own the research, scoring, sequencing, and CRM layer; reps focus on the conversations where buyers need conviction. Windsurf scaled its go-to-market team from 3 to 75 people in under a year using AI-enabled recruiting and onboarding. The structure pairs a thin layer of agents handling volume with reps redeployed onto the deals that need a human.
We write about what it takes to run an AI-native commercial org. No tool reviews, no hype. The decisions a CRO or CEO actually has to make. Join the newsletter to get the next piece.
Sources
- Gartner, "Predicts 2026: Leading Sales in the Age of AI Contradictions" (November 2025). gartner.com
- Gartner, "69% of B2B Buyers Turn to Sales Reps to Validate AI-Generated Insights" (645 B2B buyers, May 2026). gartner.com
- Gartner, "Sales Organizations That Provide AI-Enabled Next Best Actions Are 2.6x More Likely to Achieve Commercial Growth" (227 CSOs, May 2026). gartner.com
- Gartner, "By 2030, 75% of B2B Buyers Will Prefer Sales Experiences That Prioritize Human Interaction Over AI" (August 2025). gartner.com
- McKinsey, "An unconstrained future: How generative AI could reshape B2B sales" (2024). mckinsey.com
- BCG, "How AI Agents Will Transform B2B Sales" (400 US sales reps and managers, October 2025). bcg.com
- HBR, "How Successful Sales Teams Are Embracing Agentic AI" (September 2025). hbr.org
- SaaStr, "Sales in the Age of AI: The Playbook from the CROs of Databricks, Windsurf, Perplexity, and Owner" (2025). saastr.com
- Carles Reina (ElevenLabs) on 20VC with Harry Stebbings, reported by Business Insider, "Sales reps at the $11 billion AI startup ElevenLabs have to bring in 20 times their base salary, or they're out" (February 2026). businessinsider.com via aol.com
By Christopher Kliebenstein. We build and run AI-native workflows for commercial operators.