Pre-sale AI support statistics across four service comparisons https://offloadrep.com/guides/pre-sale-ai-support-statistics-across-four-service-comparisons/ ## Quick answer In a published retail field experiment, AI pre-sale replies increased sales by 16.25% compared with unavailable live support. Against human-only replies, the estimate was 3.71% and statistically inconclusive. AI with human escalation produced an 11.53% sales estimate against human-only support, significant at the study’s 10% threshold. The comparison group changes what a merchant can conclude (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). A merchant evaluating AI replies first needs to identify what happens to the shopper’s question today. An unanswered question, an available human and a service with escalation create different comparisons. This analysis follows the randomized sales and purchasing results, shows the uncertainty around the human comparisons and identifies the outcome a merchant would need to test locally. ## Key Takeaways The no-live-help comparison produced a 16.25% sales estimate In the published pre-sale retail experiment, AI replies increased sales per assigned consumer by 16.25% relative to unavailable live support, with 44,614 consumers in that comparison (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4#A3.T1), Jun 2026). The human-only comparison remained inconclusive The AI-versus-human-only sales estimate was $0.0701 per assigned consumer with a standard error of $0.0967. A normal 95% interval calculated from these published inputs spans approximately −$0.119 to +$0.260 (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4#A3.T1), Jun 2026). The purchase lift was 13.1 additional purchasers per 1,000 The AI-versus-unavailable-support purchase-probability effect was 0.0131, equivalent to 13.1 additional purchasers per 1,000 assigned consumers. The paper reports a 21.70% relative increase (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4#A3.T2), Jun 2026). Hybrid versus human support warrants a qualified reading AI with human escalation produced an 11.53% sales estimate against human-only support. The published result reached the 10% significance threshold, while a normal 95% interval calculated from the coefficient and standard error spans zero (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4#A3.T1), Jun 2026). ## Which service would the shopper otherwise receive? An AI assistant can fill an empty service slot or change an existing service. Those are different investments. A shopper who gets an answer to a delivery question may complete an order they would otherwise leave. The comparison with an available human agent requires its own evidence. Figure description: Pre-sale AI support statistics: AI versus unavailable help adds 13.1 purchasers per 1,000 with interval 8.16 to 18.04; AI versus human help is negative 0.768 with interval negative 5.844 to positive 4.308, so equivalence is not established. Figure 1: Filling an unanswered slot and replacing a human are different decisions. Source: Generative AI and Sales Productivity, Table C2 (https://arxiv.org/html/2510.12049v4#A3.T2), Jun 2026. Original analysis of the cited source. Source periods and calculation limits appear in the text. The retail experiment makes this distinction unusually clear. Its main comparison put AI pre-sale replies against an automated message saying that live support was unavailable. Sales per assigned consumer rose by 16.25%. That result is relevant to stores whose pre-purchase questions currently go unanswered, including questions that arrive outside the hours their team covers (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). Three supplementary comparisons tested human replies and a hybrid service with escalation. Reading them alongside the headline gives a much more useful basis for a staffing decision than quoting the largest lift alone. ## What do the four service comparisons establish? Against unavailable support, AI replies added an estimated $0.274 in sales per assigned consumer. AI with human escalation added $0.422 in its separate comparison with unavailable support. Both estimates were statistically significant at the study’s 1% level (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). Against human-only replies, the AI estimate was $0.0701 per assigned consumer. Its standard error was $0.0967. Our approximate 95% interval runs from a $0.119 decrease to a $0.260 increase, so the experiment leaves substantial uncertainty around that difference (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). The hybrid service added an estimated $0.218 per assigned consumer against human-only replies. That comparison reached the paper’s 10% significance threshold. Its approximate 95% interval runs from a $0.005 decrease to a $0.441 increase. It is a promising result that calls for a measured pilot, especially before a merchant changes an effective human service (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). ## Did more shoppers buy or did their baskets become larger? The main AI-versus-unavailable comparison increased purchase probability by 0.0131. That is 1.31 percentage points, or 13.1 additional purchasers per 1,000 assigned consumers. The paper reports the same estimate as a 21.70% relative increase because it divides the absolute change by the control group’s purchase rate (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). Those two expressions describe the same result. A forecast that treats 21.70% as an extra 217 purchasers per 1,000 would greatly overstate the observed change. Keep the original probability difference beside the relative headline whenever the figure enters a business case (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). The hybrid-versus-unavailable comparison added 17.5 purchasers per 1,000 assigned consumers. Basket-value estimates in the four comparisons were statistically inconclusive. The available evidence points more clearly to changes in whether consumers purchased than to a dependable increase in what each purchaser spent (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). ## Which unanswered questions belong in the business case? A merchant can begin by identifying where a useful answer is currently missing. For an abandoned checkout, that might be a question about the selected size, a delivery date, the contents of a bundle or how to use an existing discount. The relevant test is whether timely, correct assistance changes completed orders among shoppers eligible for that assistance. The control experience should reflect what those shoppers genuinely receive today. If a human already answers the same questions promptly, the human-only comparison is the appropriate reference. If the question normally sits unanswered until the next working day, an unavailable-support comparison is more informative. A pilot should preserve the shopper’s current cart and the merchant’s actual information. An answer about a different variant, an unsupported delivery promise or a second discount can change the transaction in ways the experiment did not study. Record those failures separately from whether a message was sent or opened. ## How should human escalation be evaluated? The hybrid condition let AI begin the conversation and pass unresolved issues to human agents. That is a more specific service design than adding a generic contact link. The escalation path must deliver the question, the relevant cart or product information and the work already completed so the shopper can continue without starting over. The appropriate escalation boundary still depends on the store and the question. A merchant can define that boundary around issues the assistant can verify, such as a published product attribute, and issues that require a person, such as an exception to a policy. Then compare both the purchase outcome and the cost of resolving the question. Track how often escalation is needed, how long shoppers wait and whether the final answer corrects the initial one. These measurements make a disappointing pilot useful. They help distinguish missing coverage, poor information and an ineffective handoff instead of treating every lost order as a model problem. ## What should a merchant measure before applying the result? The source comes from a single cross-border retail platform and its specific service arrangements. It gives evidence that pre-sale assistance can change sales under those conditions. The recovery rate for another merchant’s iMessage programme remains to be measured. For a local test, assign eligible shoppers before they choose whether to engage. Compare sales or contribution per assigned shopper over a defined purchase window. Report how many assigned shoppers actually received assistance, and keep the assignment result visible even when a separate analysis examines engaged conversations. The useful question is how much the proposed service improves the experience shoppers already have. A clear baseline, accurate answers and a measured escalation path make that question answerable. ## Which reported measures support these findings? The figures below preserve the source’s outcome labels and units. Read comparisons within their stated experiment, population and period. Reported measure | Value and unit | Source AI vs unavailable support, sales | 0.274 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs unavailable support, sales relative effect | 16.25 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs unavailable support, sales standard error | 0.097 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, sales | 0.0701 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, sales relative effect | 3.71 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, sales standard error | 0.0967 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, sales | 0.422 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, sales relative effect | 25.03 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, sales standard error | 0.115 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, sales | 0.218 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, sales relative effect | 11.53 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, sales standard error | 0.114 USD per assigned consumer | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs unavailable support, purchase probability | 0.0131 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs unavailable support, purchase probability relative effect | 21.7 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs unavailable support, purchase probability standard error | 0.00252 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, purchase probability | -0.000768 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, purchase probability relative effect | -1.03 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI vs human-only support, purchase probability standard error | 0.00259 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, purchase probability | 0.00358 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, purchase probability relative effect | 4.82 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs human-only support, purchase probability standard error | 0.00304 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, purchase probability | 0.0175 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, purchase probability relative effect | 28.99 % | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) AI plus escalation vs unavailable support, purchase probability standard error | 0.00291 probability difference | (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026) ### How are the additional calculations reproduced? These are arithmetic comparisons of the published inputs. They retain the same sample and limitations as the inputs and do not create an additional experiment. Calculation | Expression | Result and unit Additional purchasers per 1,000, AI vs unavailable support | .0131 * 1000 | 13.1 purchasers per 1,000 Additional purchasers per 1,000, hybrid vs unavailable support | .0175 * 1000 | 17.5 purchasers per 1,000 AI vs human sales, lower 95% limit | .0701 - 1.96 * .0967 | -0.119432 USD per consumer AI vs human sales, upper 95% limit | .0701 + 1.96 * .0967 | 0.259632 USD per consumer Hybrid vs human sales, lower 95% limit | .218 - 1.96 * .114 | -0.00544 USD per consumer Hybrid vs human sales, upper 95% limit | .218 + 1.96 * .114 | 0.44144 USD per consumer ## Does an inconclusive human comparison establish a safe replacement? The AI-versus-human purchase estimate is negative 0.768 purchasers per 1,000 assigned consumers. Its approximate 95% interval is negative 5.844 to positive 4.308, calculated from the published coefficient and standard error. This does not prove AI reduces purchases, but it also does not rule out a decline of two purchasers per 1,000. A merchant whose replacement decision requires losses smaller than that worked allowance would not have resolved its threshold with this interval. (Generative AI and Sales Productivity researchers, Table C2 (https://arxiv.org/html/2510.12049v4#A3.T2), Jun 2026). By comparison, the AI-versus-unavailable-support purchase estimate is 13.1 additional purchasers per 1,000, with an approximate 95% interval of 8.161 to 18.039. Those are separate randomized comparisons with different samples and baselines. Their contrast in evidential meaning helps identify whether the local project is filling an unanswered service slot or replacing an effective service; subtracting the two effects would not create a new head-to-head experiment. ### What should a replacement decision require? Set an acceptable loss allowance in the outcome that matters, along with factual-quality and escalation requirements. Then compare the proposed service with the actual human service. A statistically inconclusive difference is not a demonstration of equivalence or noninferiority. Cost savings can affect the commercial decision, but they do not change the outcome interval or erase a wrong answer. A coverage decision has another comparator. When eligible questions normally receive an unavailable message, an accurate automated answer may create value by making help available at all. Track those eligible unanswered cases separately from questions a staffed team already resolves. A portfolio-wide lift can be diluted or exaggerated when the business case applies the unavailable-support result to all conversations. ### Is human escalation an independently established upgrade over AI alone? The four published comparisons do not supply a direct AI-plus-escalation-versus-AI-only estimate in one shared population. The $0.422 and $0.274 sales figures each compare against unavailable support in separate samples. Their difference of $0.148 is therefore not an identified escalation effect. A merchant deciding whether to add escalation needs an appropriate direct comparison, together with the cost and quality of actually completing the handoff. Keep the first verified answer, the unresolved question, the handoff and the eventual outcome as separate events. A contact button is not the same exposure as a human receiving the conversation context and taking responsibility for the remaining decision. A local test should inspect incorrect confident answers and abandoned handoffs, not only completed interactions. The replacement-threshold calculations (https://offloadrep.com/guides/pre-sale-ai-support-statistics-across-four-service-comparisons/data/support-comparator-thresholds.csv) and calculation (https://offloadrep.com/guides/pre-sale-ai-support-statistics-across-four-service-comparisons/analysis.py) preserve the absolute purchase estimates and their 95% intervals. The two-per 1,000 allowance is a worked decision threshold, not a source recommendation or proof that a present-day support replacement is safe. ## Methodology and Sources We transcribed the four separate randomized comparisons in Tables C1 and C2 of the June 2026 version of Generative AI and Sales Productivity. We converted purchase-probability differences into additional purchasers per 1,000 assigned consumers and calculated approximate 95% intervals as coefficient plus or minus 1.96 standard errors. The four comparisons use different samples. We do not subtract their percentage lifts to estimate a head-to-head treatment effect (Generative AI and Sales Productivity researchers (https://arxiv.org/html/2510.12049v4), Jun 2026). Generative AI and Sales Productivity researchers: Version 4, 1 June 2026. Citation months identify the cited publication or version. Where a citation explicitly says “source checked”, November 2026 is the verification month, and the original study period remains stated in the text. Source observations, calculations and the proposed merchant applications have distinct meanings. AI vs human-only sales and conversion are statistically inconclusive. The hybrid vs human-only sales result is only marginal at p<.1 and its normal 95% interval crosses zero. The study does not measure iMessage abandoned-checkout recovery or this portfolio product. ## Glossary Sales per assigned consumer: Sales value divided by every consumer allocated to that condition. Hybrid support: AI replies with the human escalation offered in the experiment. Baseline: The service the comparison group received. Relative effect: The difference expressed as a share of the baseline. Confidence interval: A range expressing sampling uncertainty under the model. ## Sources Generative AI and Sales Productivity researchers: Generative AI and Sales Productivity: Field Experiments in Online Retail (https://arxiv.org/html/2510.12049v4) Generative AI and Sales Productivity researchers: Table C1, pre-sale service sales results (https://arxiv.org/html/2510.12049v4#A3.T1) Generative AI and Sales Productivity researchers: Table C2, pre-sale service purchase and basket results (https://arxiv.org/html/2510.12049v4#A3.T2) Check the source figures The data file retains source labels, values, populations and limitations for this analysis.