Can learning where to use AI help you win customers?
In an experiment with early-stage startups, did practical examples of using AI at work help firms gain customers? Startups taught through concrete AI use cases were 11 percentage points more likely to acquire paying customers than the control group. The finding suggests that learning where AI fits into work can matter for customer acquisition.
ACADEMIC WORK REVIEWED
Mapping AI into Production: A Field Experiment on Firm Performance
ResearchersHyunjin Kim, Dahyeon Kim, Rembrand Koning
INSEAD working paper · 30 March 2026 · 2026/20/STR · Working paper · Source language: English
English review: 31 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
Does learning through examples where AI fits into work help startups complete more tasks and acquire customers?
The researchers examine how AI benefits in individual tasks might translate into firm performance. Their mapping problem means discovering where and how AI creates value within the production process. The experiment tested whether learning concrete applications from other firms changes that discovery process and business outcomes.
Explaining the context · DijitalPi commentary
DijitalPi commentary: When assessing a training budget, consider which work the team could do differently as well as how many tools the course introduces. This study is relevant to marketing because it measures customer acquisition, but the largest differences in AI use appeared in product development and product/strategy design. Reading the findings as evidence of better advertising campaigns would therefore be inappropriate.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
The authors recorded their research plan before running the field experiment. It took place within INSEAD’s AI Founder Sprint, a three-month online programme supporting startups around the world. The sample comprised 515 early-stage firms selected from July 2025 applicants that consented to research and completed baseline data collection. The programme began in the week of 25 August. Firms were randomly assigned within 32 groups defined by geography, initial progress and baseline AI use: 255 received the intervention and 260 were controls. A median is the middle value when values are ranked. The firms’ median founding year was 2024, and the median team size was 4. The experimental window was ten weeks. Use-case and task analyses covered 507 firms, and endline measures covered 479.
Understanding the method · DijitalPi commentary
From the third week, the intervention group attended compulsory workshops using cases about how firms reorganised production around AI. Controls discussed general entrepreneurship cases, such as defining customer profiles and validating ideas, in the same format. Technical training, credits for using AI services, mentoring and investor access were the same for both groups. The comparison therefore evaluates the case instruction rather than providing access to AI.
Data came from founders’ weekly progress reports. Two reviewers who were unaware of the experiment and group assignments checked new, distinct AI uses. A third resolved disagreements. Completed tasks, paying customer acquisition and revenue were tracked separately. The Venture Progress Index combined indicators for product launch, customer acquisition, revenue generation and investment raised. Revenue and investment were founder-reported. The researchers used a calculation that changes how much influence monetary amounts have on the result. This is called a logarithmic transformation.
03 / RESEARCH FINDINGS
Case instruction increased customer acquisition. The strength of the revenue finding varied across analyses
The researchers reported an average of 8.8 distinct AI uses in the intervention group and 6.1 in the control group. The paper describes this as 2.7 additional uses and roughly a 44% increase. Average completed tasks were 20.9 and 18.5. The estimate accounting for the randomisation groups was +2.3 tasks (p<0.05). The estimate for internal tasks was +2.2 (p<0.01), with no significant difference in external tasks. The probability of acquiring paying customers was 11 percentage points higher, expressed in the source as an 18% relative increase (p<0.01). The authors reported 1.9-fold higher total revenue from the logarithmic analysis, but this result met only the p<0.1 threshold. There was no significant effect on whether a firm generated revenue. The estimate using raw dollar amounts was negative and non-significant. Alternative analyses limiting the influence of extreme values also produced mixed results. Revenue effects were small across most of the distribution and largest at the 90th and 95th percentiles. There were also no significant average effects on the probability of raising investment or the amount raised.
Reported distinct AI uses
8.8 / 6.1Intervention and control group averages respectively. These count different applications, not the number of tools used.
Probability of acquiring paying customers
+11 percentage pointsThe intervention group’s difference relative to controls. This does not mean an 11% increase in customer numbers or revenue.
DijitalPi commentary: When buying training or advice, write down which step in your own work the example case could change. Assess completed internal work, customer interactions and paying customer acquisition separately. External tasks showed no significant difference. Do not assume that progress in internal work means progress at every stage of selling.
Random assignment supports a causal interpretation of the case instruction within this programme. It does not establish that every additional AI use delivers the same return. Among revenue-generating firms, baseline median monthly revenue was $1,000, and revenue outcomes were founder-reported. Although the authors describe the revenue finding as robust to alternative analyses, Table A3 presents a more mixed picture. Do not transfer the 1.9-fold estimate into your business budget or revenue forecast as an expected gain.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
Customer acquisition and revenue figures were reported by the founders. This working paper shows that, with equal access to tools and support, instruction through concrete AI cases can increase the variety of AI applications, completed internal tasks and paying customer acquisition among early-stage ventures. It is not a peer-reviewed journal article; the ten-week results do not establish lasting effects, outcomes in mature businesses or applicability to Türkiye and Turkish-language use. INSEAD provided financial support. The authors designed and ran the experiment within their institution’s accelerator and wrote some of the teaching cases themselves. No separate conflict-of-interest statement was found in the source. Free resources supplied by partners were provided equally to both groups.
Our open question at DijitalPi: which work step would you choose to make AI’s role concrete for your team, and how would you separately measure whether progress in that internal work carries through to customer acquisition?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: preparing product information for publication
An illustrative e-commerce business has access to AI tools but struggles to prepare product information for sale.
Would another tool help, or would showing where the current tool fits be more useful?
Open the recommendation for example 01+
Start with a case explaining a specific step in preparing product information
Write down the steps from supplier information to a checked product description. Where comparable teams are available, keep tool access the same and give one a concrete workflow example and the other general usage guidance. Track completed, approved product records separately from newly acquired paying customers. The study did not test this e-commerce application. Producing content alone does not demonstrate customer acquisition.
EXAMPLE 02
B2B software: a prototype before customer meetings
An illustrative B2B software startup is struggling to prepare a product prototype for customer meetings.
Will completing more internal work with AI lead to more customer meetings?
Open the recommendation for example 02+
Treat prototype production and external meetings as separate objectives
Identify the prototype steps where AI will produce drafts and the team will review them. Record completed prototype tasks, customer meetings held and paying customer acquisition separately. Additional tasks in the study were concentrated in internal work, with no significant difference in external tasks. Do not interpret faster prototyping as an increase in meetings or contracts.
EXAMPLE 03
Tourism: reorganising quotation preparation
An illustrative tourism business repeatedly rearranges the same information across documents when preparing customer quotations.
Which outcome should I use to assess AI training?
Open the recommendation for example 03+
Measure quotation preparation separately from customers’ purchasing decisions
Identify the information gathering, drafting and human review steps in preparing a quotation. Build training around a concrete example of these steps and record completed quotations and outputs needing correction. Track paying customer acquisition separately. Do not use the paper’s 1.9-fold revenue estimate as an expected return. The study did not evaluate this tourism application or its results in Türkiye.
TRY IT WITH YOUR TEAM
Choose one work step
Design a small application comparing concrete case instruction with general training while keeping tool access constant. Measure completed internal work and paying customer acquisition separately. This is a proposal that DijitalPi has not yet implemented.
Ajans AI Dönüşümü provides consultancy for agencies reorganising their work around AI and automation. It does not guarantee the customer acquisition gain reported in this study. Explore Agency AI Transformation ↗
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences