The jagged frontier comes to sales development
Abstract. Two field experiments changed what we know about AI at work: generative AI lifted the least experienced workers most, and it improved quality dramatically on tasks inside its capability frontier while degrading judgment on tasks just outside it. This review translates both findings into a delegation discipline for sales development teams: which pipeline tasks to delegate, which to direct, and which to keep fully human.
The question
Every SDR organization is now making the same decision, mostly implicitly: which parts of the job belong to the AI system, and which belong to the human. The implicit version of this decision fails in two opposite directions. Teams that under-delegate keep reps typing all day. Teams that over-delegate ship confident garbage at scale. The research gives us something better than instinct.
The evidence
Brynjolfsson, Li and Raymond (2023) studied a generative AI assistant deployed to roughly five thousand customer support agents. Average productivity rose 14 percent, but the distribution is the finding: the least experienced agents improved most, while the most experienced improved little. The assistant effectively transferred the experts’ patterns to the novices. For sales development, a function with chronically high turnover and long ramps, this is the strongest argument yet that new reps should learn with the system from their first week, not after they have “mastered the basics.”
Dell’Acqua and colleagues (2023) ran 758 BCG consultants through realistic tasks with and without AI. On tasks inside the model’s capability frontier, AI users produced work judged over 40 percent higher in quality. On tasks deliberately designed to sit just outside the frontier, AI users were 19 percentage points more likely to produce wrong answers than the control group, because the output looked right. The frontier is jagged: it does not track task difficulty in any intuitive way, so workers cannot feel where it is. They have to learn it.
The older human-automation literature predicted the failure mode. Parasuraman and Riley (1997) catalogued automation misuse and complacency; Lee and See (2004) showed trust in automation must be calibrated to actual reliability, not to fluency or confidence. A generative model is fluent everywhere and reliable somewhere, which is precisely the combination that miscalibrates trust.
The mechanism
The three-verb discipline follows directly. Delegate the tasks well inside the frontier, where output is reliably correct: retrieval, enrichment, scheduling, execution timing. Direct the tasks on the frontier, where drafts are valuable and errors are plausible: copy, account tiering, research summaries. Here AI proposes and a human disposes. Own the tasks beyond the frontier or too consequential to delegate: live conversations, qualification judgment, the final say on anything a buyer sees. And because complacency is the documented drift, every Delegate and Direct lane needs a named review point with a sample rate, an owner, and a cadence.
Implications for practice
A delegation map is a team document, not a personal habit. It versions as the frontier moves. New hires read it on day one. Managers audit against it. The Institute scores it in certification because it is the clearest single artifact of AI-era competence.
In the curriculum this paper underpins PA 120 (AI Orchestration I) and the whole 300 level. The practical instrument it produces, the delegation map, is the first artifact in the Certified Pipeline Practitioner portfolio.