When Interest Did Not Become Adoption: The Training Test
For the Georgian small firms in this study, the training experiment asks a focused question: did a practical course that lowered the apparent difficulty of digital commerce change what firms actually did? The answer is important precisely because the firms had already expressed interest. Readers can use the result to distinguish a failure of a particular training offer from a general claim that skills never matter, and to see why motivation and organizational capacity remain plausible explanations when knowledge-building produces no measured adoption.
A plausible intervention met a practical bottleneck
The training was not a vague invitation to become more innovative. The Georgia Innovation and Technology Agency offered a three-day, face-to-face program between May and November 2019. It addressed the ordinary tasks that might stand between a small business and a digital sales channel: increasing visibility through familiar online platforms, understanding customer profiles, registering on e-commerce platforms, and thinking through a business model, finance, grants, and procurement. The average cost was about $130 per business.
That design gives the intervention a clear theory of change. A firm might be connected to the internet and still lack the procedural knowledge to create a credible business presence. It might not know which information a customer needs, how to register on a platform, or how to translate an existing business into an online listing. Instruction should reduce that implementation cost. Once the cost falls, a firm with enough expected demand ought to be more likely to act.
The delivery format also matters. Training was initially offered to groups of three to five businesses, with individual sessions at the firm’s premises for businesses that could not attend at the designated times. The program was therefore designed to be more accessible than a purely self-directed resource. It still required an owner or manager to make time for three days, which is not a trivial requirement for a small operation. That tension—help made available, but attention still needed—is part of the intervention rather than an irrelevant detail.
The sample had already passed one motivational filter. The first 567 firms were surveyed from six regions after indicating interest in participating in e-commerce training, and another 291 businesses registered interest after an online portal opened. These 858 firms were not businesses selected at random from the entire Georgian economy. They were firms that had signaled some willingness to consider the opportunity. If even this group did not attend or change behavior, the result would be more informative about the limits of a broad training offer than a result drawn from firms with no stated interest at all.
Yet interest is not a commitment. Saying that a course sounds useful is a low-cost action; attending it for three days requires rearranging work, travel, and attention. The study found that only about 56% of the firms assigned to the training offer actually attended. The gap is not proof that the remaining firms lacked ability or did not value digital markets. It does show that the intervention had to overcome the same kind of prioritization problem that later digital adoption would require.
This is the first point at which the study’s intention-to-treat logic matters. The relevant comparison is between firms assigned to receive the offer and firms assigned to pure control, not between the people who attended and a hand-picked set of nonattenders. Attendance was itself affected by the practical appeal and accessibility of the offer. If we compare attendees with nonattendees directly, we risk attributing their pre-existing differences in motivation or capacity to the course.
The assignment estimate therefore asks a grounded policy question: what was the effect of making this training available to interested Georgian firms under ordinary delivery conditions? It combines the possibility that instruction helped participants with the possibility that many firms did not take it up. That is often the effect a program sponsor needs to know. A course that works only after overcoming substantial participation friction may still be valuable, but its real-world impact will not look like the effect of perfect attendance.
The measured result did not show a shift in digital engagement
The short-run follow-up came in 2021. Researchers summarized several indicators in an e-commerce engagement index rather than relying on one convenient measure. The components included online selling, computer and internet use for business, the quality of the internet connection, social-media or platform presence, business email, a website, online banking, online marketing, and delivery by mail. Combining related outcomes reduces the chance of highlighting one accidental result among many tests, while also making the outcome broader than “received an online order.”
Assignment to training had a small and statistically insignificant effect on the index. The paper reports no effects across the individual components, either. Firm size also did not show a meaningful improvement. The careful interpretation is not that every treated firm remained exactly unchanged. It is that the average difference between firms assigned to training and firms assigned to control was too small and too uncertain to support a positive effect on the measured digital behaviors or firm size.
That distinction between “no measured effect” and “nothing happened” is central. Statistical uncertainty can conceal a modest effect, especially when an intervention reaches only part of its assigned group. But the pattern across the composite index and its components makes a large, broad transformation less credible. If the program had substantially unlocked online participation for the average assigned firm, one would expect at least some corresponding movement in the outcomes designed to capture that participation.
The attendance rate gives the result two compatible readings. One is about participation: a course cannot change the behavior of firms that do not attend, and 56% attendance limits the effect of the offer at the population level. The other is about the mechanism: even among the firms that attended, the study’s assignment comparison does not show that the training offer translated into a detectable average shift. The design does not identify the effect of learning for a hypothetical firm that attended perfectly and absorbed every lesson, but it does reject the easy assumption that offering broad instruction to interested firms is enough.
The outcome also speaks to the difference between capability and willingness to act. Training can reduce a technical obstacle without creating a customer, a deadline, or a reason to prioritize setup over existing work. A firm may leave a course knowing more and still postpone implementation. It may recognize that the process is feasible but not believe that online demand will repay the time. Or it may need a person, payment arrangement, or delivery routine that the course did not supply. The null result does not tell us which of these explanations dominates, but it makes a pure knowledge-deficit account incomplete.
The longer horizon did not reveal a delayed training payoff
The researchers returned to the firms in 2025, approximately four years after the training intervention. This follow-up tests a different possibility: perhaps the course did not produce immediate adoption because firms needed time to use the new knowledge. Training represents an accumulation of human capital, so its effect might appear later as digital markets became more attractive or as owners encountered a suitable opportunity. In contrast, a time-limited demand shock would be expected to act quickly and then fade.
The long-run training estimates again showed no observable positive effect on the e-commerce index, number of employees, or online revenue share. The paper therefore finds no evidence of a delayed “sleeper effect” in which training quietly accumulated value and later moved firms into digital markets. This strengthens the short-run conclusion for the tested offer: additional time did not compensate for the absence of an initial change.
But the long-run result has its own measurement boundaries. The composition of the e-commerce index changed somewhat between 2021 and 2025 as the questionnaire was adapted to capture longer-term engagement. Direct revenue evidence is especially limited because more than 60% of firms refused to report revenue in the 2025 survey. The researchers could examine online revenue share and other reported indicators, but those measures cannot support a precise estimate of how much extra income training generated.
The endline also cannot turn a null assignment effect into a universal verdict on education. The training was delivered in a particular place, to firms outside Tbilisi that had expressed interest, with a particular curriculum and attendance pattern. A targeted intervention that diagnoses a specific knowledge gap, provides continuing support, or reaches a different population could have a different effect. The evidence says that this broad, short course did not produce a measurable change under these conditions. It does not say that every skill constraint is imaginary or that all training is wasteful.
There is a second reason to resist an overly broad conclusion. A null effect on digital engagement does not prove that technical capability had no role in any firm’s decision. Some firms may have learned useful skills that they applied in ways the index did not capture. Others may have benefited but lacked demand. Still others may have been unable to attend. The experiment’s strength is not that it observes every latent mechanism; it is that it shows what the offered program changed, on average, in the outcomes the researchers measured.
This is how the result updates the central puzzle. Before the experiment, low online participation could reasonably be attributed to a lack of know-how. After the experiment, that explanation is weaker as a complete account for these firms. The combination of initial interest, a relevant curriculum, a real delivery effort, limited attendance, and no detectable short- or long-run improvement suggests that capability was not the only—or perhaps not the primary—constraint activated by this intervention.
What the training result leaves open
The fairest conclusion is narrower and more useful than “training does not work.” The Georgian experiment shows no meaningful average improvement from assigning this training offer to this interested sample, either by the 2021 follow-up or in the 2025 measures. It also shows that interest did not reliably translate into attendance, and that the absence of a later sleeper effect makes a simple delayed-capability story less persuasive.
What remains open is the source of the residual friction. If owners already understood enough to act but lacked a customer, the missing input was demand. If they saw a possible customer but could not fit the setup work into daily operations, the bottleneck was organizational. If they doubted that a new channel would pay, information about technique would not alter the expected return. These possibilities predict that a concrete opportunity might do more than another lesson—not because knowledge is irrelevant, but because the decision has more than one input.
The training result also changes how a later incentive should be interpreted. If a buyer’s offer succeeds, it would not simply prove that firms secretly possessed every skill they needed. It would show that the circumstances surrounding the task—its value, urgency, and accountability—can release action that instruction alone did not. If the offer fails too, then the explanation would have to account for a barrier more resistant to both capability support and immediate demand.
For the Georgian firms, the next test was designed around that unresolved difference. Instead of asking owners to learn a general route into digital markets and decide later whether to use it, the study presented a specific reason to become able to sell online. A buyer, a conditional order, and a short deadline changed the question from “Would this be useful someday?” to “Can this firm make a transaction happen now?”