The Effects of Voice-Based Digital Training and Generative AI on the Self-Employed: A Field Experiment in Tanzania

In low-and-middle-income countries (LMICs), self-employment represents the primary form of labor. “Subsistence entrepreneurship” (Cho, Robalino, and Watson 2016) is a challenging livelihood. Many microenterprise owners operate in crowded local markets with limited capital, while facing frequent economic shocks. Against this backdrop, and with little formal training or reliable guidance, the self-employed must regularly make significant financial and business decisions that determine the viability of their enterprises.

One particularly important decision is whether and how to use credit. Borrowing can enable business owners to finance productive investments, smooth cash flow, weather shocks, and take advantage of growth opportunities. At the same time, borrowing entails significant risks if investments fail to generate the expected returns or repayment obligations cannot be met.

We test the impact of voice-based digital tools designed to help small business owners learn more about credit, make informed decisions about when and how to use it, and strengthen their general financial literacy and business skills. Both interventions are delivered in Kiswahili through Viamo's interactive voice response (IVR) platform in Tanzania—known as Elimika (“to learn”)—which users reach by dialing a toll-free number.

The first intervention is a digital training (DT) program. It entails a series of five dramatized interactive lessons that teach listeners about: financial credit; loan products, terms, and costs; responsible borrowing; how to apply for and qualify for a loan; and using a mobile phone for business, including making and accepting Lipa kwa Simu (“pay-by-phone”) payments.

The second intervention provides access to AVA (“Ask Viamo Anything”), a voice-based generative AI assistant on the Elimika platform that responds to users' spoken queries on demand in Kiswahili. Unlike the structured digital trainings, AVA offers guidance tailored to users’ own questions, whether about business and loans or other topics, as they arise and become relevant. Participants received five toll-free calls per month over six months.

A third intervention combines the DT modules and access to AVA.

Participants are randomly assigned to one of four groups: (i) Control; (ii) T1: Digital Training only; (iii) T2: AVA only; or (iv) T3: Digital Training plus AVA.

2025-12-12

2026-07-31

We pre-specify three primary outcomes. Each is measured through an endline phone survey. We report each treatment arm versus Control and AVA-versus-DT (T2 vs T1).

First, we analyze the effects on interest in and demand for financial credit; its use for productive purposes (e.g., to support business or income-generating activities); and responsible borrowing practices (e.g., participants consider the costs of borrowing before taking out loans).

Second, we examine changes in financial literacy and knowledge, including knowledge of credit and loan terms (interest and fees, repayment periods, how to qualify, consequences of non-repayment) and registration requirements for acquiring a Lipa kwa Simu till number. We expect the DT modules that directly provided structured training on financial credit and credit products to produce the largest gains in underlying knowledge.

Third, we analyze effects on business self-efficacy (i.e., confidence in one's ability to solve business problems) and access to business guidance (i.e., whether respondents report having a source of advice when business questions or problems arise). We expect AVA to produce the largest gains here as the digital assistant can serve as a business coach and provide advice when enterprise owners feel they have no other source to turn to.

(1) Credit interest, demand, and responsible borrowing: This covers whether participants borrowed money in the past six months and specifically whether they borrowed for business or income-generating activities. It also includes the size of the largest loan and whether participants have fully repaid it.

Additionally, the domain measures prospective demand for credit through a hypothetical borrowing scenario, in which respondents are asked: i.) how likely they would be to apply for a loan if they needed funds to invest in their business (e.g., to restock inventory or take advantage of a supplier discount); and ii.) to measure responsible borrowing, whether respondents report that they would only do so when the expected returns justify the cost of borrowing and when they are confident they can meet their repayment obligations.

Finally, among current or future business owners, we analyze participants’ intentions of applying for any kind of loan to invest in their business or enterprise over the next six months at the extensive and intensive margins.

To assess whether borrowing may be leading to indebtedness, we assess participants’ repayment difficulty (whether they are behind on a current loan and negative coping actions) measured for the subsample for whom these items were fielded.

(2) Financial literacy and knowledge: This domain measures participants' understanding of credit and how mobile credit products work. We ask a series of factual questions and score whether each is answered correctly: that the fee or interest on a loan represents the cost of borrowing; that repaying a mobile loan on time still means paying back more than the amount borrowed; the typical flat fee on a merchant loan (about 11%) and the typical repayment period (about 30 days); the best reason to take a loan (a productive investment the borrower can repay on time); what qualifies a small business owner for a digital merchant loan (a record of business activity through Lipa kwa Simu); and that failing to repay on time results in a penalty. We sum the correct answers into a credit knowledge index.
A second measure captures participants’ understanding of what is required to register for a Lipa Namba account—which is necessary for merchants to receive business loans from mobile network operators. This is derived from an open-ended question in which participants are asked to name the minimum requirements for Lipa Namba registration (a registered SIM card, a national ID, and a tax identification number).

We expect the digital training modules, which provide structured instruction on credit and credit products, to produce the largest gains on these indices.

(3) Business self-efficacy, skills, and guidance: This domain measures participants' confidence in running their business and their access to advice. We capture two dimensions of self-efficacy, each by agreement on a five-point scale: whether participants feel they can usually figure out a solution when problems arise in their business (problem-solving confidence), and whether they feel they have the skills they need to make their business successful (skill confidence).

We also measure access to business guidance: the number and types of sources participants turned to for advice about their business in the past six months, whether they report having no one to turn to when business questions or problems arise, and whether they cite Elimika and AVA as a source of advice.

We expect AVA to produce the largest gains in problem-solving confidence and access to guidance, as the assistant can serve as an on-demand business coach. The structured training, by contrast, may do more to build participants' sense of having the underlying business skills.

Secondary outcomes include weekly income; mobile-for-business practices; willingness to pay for the voice services; and perceived usefulness of DT and voice-first GenAI.

Weekly income covers self-reported total earnings from all income-generating activities over the past seven days. It is recorded as an open amount and coded into bands. Because the study period is only seven months at this data collection point, we treat income effects as exploratory.

We analyze the winsorized inverse-hyperbolic-sine of reported weekly earnings as our primary income measure, and the enumerator-coded income band, which should be less prone to digit-entry error, as a robustness check.

Mobile-for-business practices focuses on adoption and use of Lipa kwa Simu and Lipa Namba merchant payments (whether participants have a till number, how recently they obtained it, and how often they use it) and use of the phone for business (recording sales, communicating with customers, and advertising). We combine the payment-adoption items and the phone-for-business items into standardized indices. We treat this as secondary rather than primary because the relevant content appears in the later digital-training modules, which fewer participants completed. Second, AVA provides on-demand responses to users' questions rather than promoting specific financial products. Consistent with this design, administrative query data indicate that participants rarely asked AVA about Lipa kwa Simu, focusing instead on credit, loans, and broader business questions.

Willingness to pay: We measure hypothetical willingness to pay using a single-bounded dichotomous choice experiment. Participants were randomly assigned a hypothetical per-call price and asked whether they would be willing to pay that amount for an additional call to the Elimika service. We use these responses to estimate the demand curve for the service and compare willingness to pay with its actual per-call price.

Perceived usefulness: Among users, we measure how useful and trustworthy they found AVA and whether they would use it again or recommend it; and, for the training, its perceived relevance and whether it led participants to change anything in their business.

Our study represents a four-arm, individual-level randomized controlled trial. Microenterprise owners and aspiring entrepreneurs registered on Viamo's IVR platform were recruited and consented through an IVR baseline survey, and 3,000 (balanced by gender) were randomly assigned to one of four arms: Control, Digital Training only (T1), AVA only (T2), or both combined (T3). Assignment was stratified on baseline characteristics, including registered gender and self-reports in the baseline IVR survey of whether the participant currently operates a business, prior borrowing in the past 12 months, and consent to share contact details for a follow-up phone survey.

Treatments were deployed over approximately seven months (mid-December 2025-mid-July 2025). Platform administrative data enables tracking of treatment uptake and engagement. Endline outcomes are measured in a phone survey administered by IPA-Tanzania to the 1,757 participants who consented at baseline to be contacted for follow-up. This consenter sample is the study's analysis sample, and the primary intention-to-treat effect is defined among these follow-up consenters. Because treatment assignment was stratified on follow-up consent, the consenter sub-population is balanced across arms by design.

Primary inference centers on three contrasts: T1 versus Control (the effect of digital training), T2 versus Control (the effect of the AVA assistant), and T2 versus T1 (the on-demand assistant relative to structured training). Because access to AVA in the combined arm was substantially reduced by its placement in the IVR menu, we focus on T2 versus Control to test the AVA mechanism.

Not available

Randomization was done in office by a computer. We used the randtreat command in Stata to assign the 3,000 participants in equal proportions to the four arms, stratified (blocked) on four baseline variables: registered gender; consent to share a phone number for the follow-up phone survey; whether the participant currently operates a business or income-generating activity; and whether the participant had taken any loan in the prior twelve months.

Individual. Each participant, a registered Viamo platform user recruited via IVR, was randomly assigned to one of the four arms. There is no clustering; assignment and treatment are at the individual level.

No

No clustering.

3,000 individuals were randomized. Endline survey outcomes are collected from the 1,757 participants who consented at baseline to a follow-up survey, with a target of 1,200 completed phone interviews; this is the endline analysis sample. (Administrative engagement measures are available for all 3,000 randomized participants.)

Randomized: 750 per arm: 750 Control; 750 Digital Training only (T1); 750 AVA only (T2); 750 Digital Training + AVA (T3).

Follow-up consenters (the analysis frame): 440 Control, 439 T1, 438 T2, 440 T3 (1,757 total).

Endline target: approximately 300 completed interviews per arm (1,200 total).

Power calculations assume an endline sample of approximately 1,200 completed interviews (about 300 participants per study arm), a two-sided significance level of 5 percent, and 80 percent statistical power. For comparisons between a single treatment arm and the control group (e.g., Digital Training vs. Control), the study is powered to detect effects of approximately 0.23 standard deviations for standardized continuous outcomes. Comparing all treatment arms pooled against the control group increases precision, allowing detection of effects of approximately 0.19 standard deviations. For binary outcomes, such as whether a participant borrowed during the past six months, and drawing from the baseline borrowing rate of 20 percent, the study can detect differences of about 9–10 percentage points for single treatment-versus-control comparisons and 7–8 percentage points for pooled treatment-versus-control comparisons. We also pre-specify analyses by gender. If we are able to reach 150 participants per treatment arm within each gender, these subgroup analyses are powered to detect effects of approximately 0.32 standard deviations. They are intended primarily to explore heterogeneous treatment effects and are not powered for precise tests of treatment-by-gender interactions. To improve precision, all specifications will include randomization-stratum fixed effects and, when available, the baseline value of the outcome. We will also control for baseline mobile-money use as a pre-treatment measure of mobile and financial fluency. Because baseline demographic measures are limited, we control for respondent gender, age, and education as measured at endline.