The study, published open access in Discover Artificial Intelligence, was designed to challenge a widespread assumption in the marketing literature: that AI-driven marketing technologies uniformly improve sales performance. Most existing evidence reports a single average effect, which says little about whether the association between a technology and performance differs between lower- and higher-performing firms. The authors argue that this aggregation can hide precisely the heterogeneity that matters to managers deciding whether, and how, to deploy AI in their sales organizations.

To move beyond averages, the team employed quantile regression, a statistical technique that estimates effects at different positions in the conditional distribution of the outcome rather than only at the conditional mean. They estimated models at the 10th, 25th, 50th, 75th, and 90th percentiles of revenue growth, using ordinary least squares as a benchmark. Because the sample contains only 50 observations with five predictors, the extreme quantiles are informed by relatively few tail observations. Inference therefore relied on a nonparametric pairs bootstrap with 2,000 replications, reporting percentile 95 percent confidence intervals for every model.

A crucial part of the work was a data-integrity audit that uncovered a subtle but serious problem in the original specification. The dataset included two derived measures, marketing efficiency and revenue per salesperson, both of which are deterministic functions of revenue growth. Including such ratios as predictors creates target leakage: the model effectively contains the dependent variable by construction. When marketing efficiency was included, the model attained an R-squared of 0.935, an apparently impressive fit that was largely mechanical, since the logarithm of marketing efficiency equals the logarithm of revenue growth minus the logarithm of marketing budget. The corrected model excluded both ratios and reported a far more modest R-squared of 0.137, which the authors describe as a more credible representation of the limited explanatory content of the small synthetic sample.

With the corrected specification in place, automated lead generation emerged as the only predictor whose bootstrap confidence intervals excluded zero. Its coefficient was negative in ordinary least squares at minus 0.123, with a 95 percent interval of minus 0.211 to minus 0.022, and negative at every quantile. The intervals also excluded zero at the 75th percentile, minus 0.301 to minus 0.024, and the 90th percentile, minus 0.320 to minus 0.002. The coefficient grew in absolute size from minus 0.024 at the 10th percentile to minus 0.244 at the 90th, meaning the negative association appears stronger in the upper part of the conditional revenue-growth distribution.

This pattern runs counter to much of the prevailing literature, which portrays AI-powered lead generation as a tool for streamlining lead qualification and expanding sales pipelines. The authors point to studies noting that companies using marketing and sales automation frequently overwhelm sales staff with low-quality leads, diverting time from prospects that could actually convert. Research on lead management systems also indicates that performance depends on lead follow-up intensity, adaptive selling, and salesperson skills rather than sheer lead volume. The authors stress, however, that their data do not measure lead quality or sales workload, so this interpretation remains tentative. One explanation worth exploring, they suggest, is that faster-growing firms may have less spare capacity to absorb additional low-quality leads.

The other variables fared less well. AI usage showed small coefficients that varied in sign across quantiles, negative in ordinary least squares and at the 25th, 50th, and 90th percentiles, slightly positive at the 10th and 75th, with all confidence intervals including zero, even at the 90th percentile where the interval spanned minus 0.218 to 0.049. Marketing budget produced positive point estimates across all models, but no interval excluded zero. Sales team size and personalized campaigns changed signs across quantiles with wide intervals that consistently included zero. The authors note that AI usage and automated lead generation both exhibited broad variation in the sample, so the null result for AI usage cannot be attributed simply to insufficient adoption variability, while the narrow range of personalized campaigns may partly explain its imprecise estimates.

Robustness checks reinforced the exploratory character of the findings. Huber robust regression and ridge regression, with the ridge penalty selected by generalized cross-validation, preserved the signs of all five coefficients but differed in magnitude, with ridge substantially shrinking the estimates. Diagnostic tests on the corrected ordinary least squares model, including Breusch-Pagan, Durbin-Watson, Jarque-Bera, and variance inflation factor statistics, found no significant heteroskedasticity, non-normality, or multicollinearity, though the near-threshold Breusch-Pagan result and the small sample further motivated the bootstrap approach.

The authors are explicit that the study is analytical and hypothesis-generating, not causal or population-representative. The data are a stylized, country-neutral, cross-industry simulation of 50 firms, not drawn from any specific country, sector, platform, or historical period, and the original generation seed, distribution parameters, and generator code were not supplied, so the synthetic observations cannot be independently regenerated. The quantiles describe positions in the conditional revenue-growth distribution and should not be read as measures of firm age or life-cycle stage. External factors such as economic policy uncertainty, competitive intensity, consumer attitudes, and trust are absent from the simulation and remain plausible omitted variables.

For managers, the message is caution rather than prescription. The authors advise against inferring any universal optimal AI level from these coefficients and instead recommend staged pilots with pre-specified success criteria covering incremental revenue, conversion, forecast error, lead-acceptance rate, sales-cycle time, salesperson workload, and human override frequency. A practical scale-or-stop rule is to expand AI deployment only when incremental contribution remains positive after total technology, integration, training, and monitoring costs, and when lead quality and customer-experience indicators do not deteriorate. Lead-generation automation, in particular, should be governed by downstream conversion and capacity metrics rather than lead volume alone. The principal next step, the authors conclude, is validation in larger, longitudinal, industry-specific datasets that capture macroeconomic conditions, consumer responses, lead quality, and implementation context, alongside deposition of the original simulation generator so the synthetic sample itself can be reproduced.

Subject of Research: Distributional analysis of associations between AI-driven marketing technologies and firm sales performance

Article Title: Beyond the average association between artificial intelligence and sales performance

Article References: Sarıoğlu, C. İ., de Bruyn, C., Karacan, R., Kayacan, E. Y., & Yılmaz, S. (2026). Beyond the average association between artificial intelligence and sales performance. Discover Artificial Intelligence, 6(1), Article 1627. https://doi.org/10.1007/s44163-026-02543-y

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02543-y

Keywords: artificial intelligence, sales performance, quantile regression, automated lead generation, revenue growth, marketing analytics, target leakage, bootstrap inference, synthetic data, human-AI complementarity, sales force automation, exploratory study