Every sales leader buying AI this year is asked some version of the same question: should this tool help my reps do the work, or do the work instead of them? Vendors answer it for you. One pitch promises a "copilot" that makes every seller better. The next promises an "agent" that runs prospecting while your team sleeps. Both pitches treat augmentation and automation as a product category you pick once, at purchase.

That framing is the source of most disappointing AI rollouts in sales. The real question is smaller and more useful: for each step inside a selling job, who should lead the work, who should check it, and how do you know? This article answers that question with the evidence we have, and gives you a pattern for making the call step by step.

In plain terms, augmentation means a person and an AI system work on the task together, and the person keeps the decision. Automation means the AI system performs the task, and a person may never look at the individual output. Raisch and Krakowski, in the Academy of Management Review, put it the same way: "automation implies that machines take over a human task," while "augmentation means that humans collaborate closely with machines to perform a task" [1].

The first article in this knowledge base, Sales AI: How to Define It by the Selling Job, Not the Tool, introduced the Delegation Ladder: Inform, Draft, Act, Decide. Augmentation lives on the lower rungs (Inform and Draft), where a human approves before anything reaches a buyer or a system of record. Automation lives on the upper rungs (Act and, rarely, Decide). This article does not redefine the ladder. It explains how to choose the rung, and why the choice should be made per step rather than per tool.

The selling fundamentals matter here. The companion article in the Sales Agent knowledge base, Trust and Credibility in B2B Sales: What the Evidence Actually Supports, defines trust as a buyer's willingness to act on your information without independently verifying it. Buyers are not short of information; they are short of people whose claims survive checking. That tells you where automation is cheap and where it is expensive. Work that never reaches the buyer (logging, enrichment, summarizing, routing) can tolerate occasional errors because a person or a downstream check catches them. Work that the buyer reads, hears, or relies on spends your credibility every time it goes out.

Raisch and Krakowski add a second point that is easy to miss. Their central argument is that "augmentation cannot be neatly separated from automation." The two are "interdependent across time and space," and "overemphasizing either augmentation or automation fuels reinforcing cycles with negative organizational and societal outcomes" [1]. In a sales team, the cycle looks like this. A team automates first-touch emails. Reps stop writing them. A year later, nobody on the team can write a strong opener when the automated sequence stops working, and the only fix anyone can imagine is more automation. The reverse cycle is just as real: a team that insists every step stay human-reviewed never learns which steps the AI handles reliably, and its reps drown in approvals.

So the line between augmentation and automation is not a property of the tool. It is a design decision about each step of the work, and it moves over time as you measure.

The most useful single study on this question is a preregistered meta-analysis by Vaccaro, Almaatouq, and Malone in Nature Human Behaviour (2024), covering 106 experiments and 370 effect sizes published between 2020 and mid-2023 [2]. Two findings matter for sales.

First, human and AI combinations beat humans alone by a medium to large margin (Hedges' g = 0.64), but on average they performed worse than the best of humans or AI alone (g = -0.23) [2]. In other words, augmentation reliably helps the person, but it does not reliably beat simply letting the stronger partner do the job.

Second, the direction depends on who is stronger. When the human alone outperformed the AI alone, the combination beat both (g = 0.46). When the AI alone outperformed the human, the combination did worse than the AI by itself (g = -0.54) [2]. The type of task mattered too: decision tasks (choosing from a fixed set of options) showed losses from combining, while creation tasks (producing open-ended content) leaned positive, though that estimate rested on only 34 effect sizes and was not statistically significant [2].

Translate that into sales work. A lead-scoring or routing decision, where a model has been validated on thousands of past outcomes, is the kind of fixed-option task where a human second-guessing every call can make results worse. Writing a discovery recap or a business-case narrative, where the human holds context the model cannot see, is the kind of creation task where working together tends to help. Neither rule replaces measurement, but they tell you where to start.

Two field studies show that augmentation lands unevenly across a sales or service team. Brynjolfsson, Li, and Raymond studied 5,179 customer support agents given a generative AI assistant. Issues resolved per hour rose 14% on average, with a 34% improvement for novice and low-skilled workers and minimal impact on experienced, highly skilled workers [3].

A sales-specific study adds the shape of the curve. Luo, Qin, Fang, and Qu ran randomized field experiments with sales agents at two fintech companies, starting with 429 agents assigned to an AI or human coach [4]. The gain from the AI coach followed an inverted U: middle-ranked agents improved the most, while bottom- and top-ranked agents showed limited incremental gains. Bottom-ranked agents suffered information overload from the AI's feedback; top-ranked agents showed the strongest aversion to it. Restricting how much feedback the AI gave bottom-ranked agents substantially improved their performance. And in a third experiment with 451 agents, an AI and human coach working together beat either coach alone for both bottom- and top-ranked agents [4]. As one of the authors put it, AI and human coaches "are not dichotomous choices" [4].

That is augmentation designed well: the AI's role is tuned to the person, and a human stays in the loop where the AI alone underperforms.

In a preregistered experiment with 758 Boston Consulting Group consultants, Dell'Acqua and colleagues found that on 18 tasks inside AI's capability frontier, consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, with significantly higher quality. On one task deliberately set outside that frontier, AI users were 19% less likely to produce a correct answer [5]. The frontier is invisible from the outside. Within a single selling job, say preparing for a negotiation, summarizing the buyer's stated priorities may be well inside the frontier while recommending a concession strategy sits outside it. A tool-level decision cannot capture that. A step-level decision can.

The way people use AI is moving toward handing work over. Anthropic's first Economic Index, based on about one million Claude.ai conversations from late 2024 and early 2025, found 57% of tasks were augmented and 43% automated [6]. By August 2025, the share of "directive" conversations, meaning full delegation with minimal back and forth, had risen from 27% to 39%, and the September 2025 report stated that "this is the first report where automation usage exceeds augmentation usage." Among business use through the API, 77% of uses showed automation patterns [7]. The March 2026 report found augmentation rising slightly again on Claude.ai, and noted that users with six or more months of experience were less likely to simply delegate and more likely to succeed [8]. That is usage data from one AI provider, not a sales study, but the direction is clear: automation is growing fastest in business pipelines, while experienced individual users lean back toward collaboration.

Sales teams show the same pull. Salesforce's State of Sales survey of 4,050 sales professionals across 22 countries (vendor data, fielded August to September 2025) found that 87% of sales organizations use some form of AI, 54% of sellers say they have used agents, and sellers spend an average of 40% of their time selling [9]. Sellers expected agents to cut prospect research time by 34% and email drafting by 36% once fully implemented [9]. Note that these are expectations, not measured results, and come from a company that sells agents. The same survey found 51% of sales leaders with AI said disconnected systems were slowing their AI initiatives [9].

The evidence points to five ways the augmentation and automation decision goes wrong.

The design pattern this knowledge base uses for this decision is the Lead and Check Split. Break each selling job into its steps, and for every step assign two roles: who leads (produces the work) and who checks (verifies it before it matters). That gives four modes.

Three tests decide the mode for each step.

Each step's assignment is provisional. Re-run the tests each quarter, and move a step between modes only on evidence.

For reps. Ask for AI that checks your work, not only AI that does it. A second reader that flags weak claims or missing stakeholders protects your credibility on the steps where you are stronger. Where the AI drafts, treat yourself as the editor accountable for every sentence the buyer reads.

For sales managers. Expect AI help to land unevenly. Middle performers and newer reps tend to gain the most, while new reps can be overloaded and top reps tend to resist [3][4]. Tune how much the AI says to each tier, pair AI coaching with your own, and keep some steps human-led so the team keeps its skills.

For whoever builds or buys it. Do not evaluate "copilot versus agent" as a product choice. Map the job into steps, run the stronger-partner test on your own data, and assign modes step by step. Treat vendor time-saving figures as expectations until you measure them [9]. Fix data plumbing before automating anything that writes to the CRM or talks to a buyer.

Augmentation and automation are not rival strategies. They are two settings you assign to each step of the work, based on who is stronger, who sees a mistake, and what your team needs to keep knowing how to do.