IOV LABSSearch the news for "AI call center" and you get thousands of seats, enterprise contracts, headlines about agents replaced. Search for what the same technology does for a hair salon, a clinic, or a five-person shop, and you get almost nothing. That is the odd gap this study fills: the segment with the most automation leverage, Korean small and medium businesses, is the one whose economics nobody has bothered to compute.
The saving is arithmetic, not a pitch
We make no new experimental data. We take two numbers that are already public and multiply them. The first is what a Korean customer-service agent actually costs an employer, not the salary but the fully-loaded figure with the employer's share of social insurance, severance, and benefits, which pushes it to about 1.2 to 1.25 times gross. A regular agent lands near 37.5M KRW a year, and the salary is only about eighty percent of that. The second is the deflection rate, the share of inquiries the AI closes with no human. From there, saved headcount is inquiries deflected times agents, and the annual saving follows. A three-agent team at a conservative 55% deflection saves about 62M KRW a year; a five-agent team saves over 100M.
One uncertain number, set low on purpose
The entire result rides linearly on the deflection rate, so we did not flatter it. Structured intents, the reservations, hours, and price questions that dominate small-business chat, benchmark at 65 to 80 percent. But there is a warning in the field that a 45 percent deflection can mean only 14 percent true resolution, because handing a conversation off is not the same as solving it. So we set the base case at 55 percent, below the structured band, and report a conservative 40 and an optimistic 70 alongside it. Even after subtracting a plausible subscription fee, the return is 17 to 52 times.
For the smallest shops, it is time, not headcount
A one-person business has no agent to remove; automation gives the owner back roughly 29 hours a month otherwise spent answering the same questions. The after-hours inquiries a closed shop would have missed, converted into bookings, are revenue rather than cost, so we leave them out of the model entirely.
This is a model, not a measurement, and it names its own interest: the author runs a product in this category. The honest move was to build the whole thing from third-party benchmarks, set the one uncertain input low, and publish the arithmetic so anyone can change it.
What would make it real
Not a large study, a single measurement. One real SMB customer, one month, four numbers: total inquiries, the share the AI closed alone, the share actually resolved without a repeat, and the human hand-off rate. Those four replace the single 55% assumption and turn the whole model into evidence. We plan to publish that follow-up.