Electricity Distribution South, part of EVN Bulgaria, announced its Eli voice assistant on November 12, 2026. The service uses INSAIT's BgGPT for calls about loss of electricity supply. Also covered by Bulgarian technology outlet Kaldata, the deployment is a concrete example of a Bulgarian language model entering customer service.
What the service does
The utility describes checking the reason for an interruption and registering a report where necessary. A caller provides an ITN, the identifier of the electricity measurement point. Callers are told that they are speaking to a digital assistant, and transfer to an employee is available. This is a customer-service interface, rather than an autonomous electricity repair or grid-control system.
INSAIT confirms its involvement and the use of BgGPT. The institute associates an automation figure of around 70% of no-power calls with initial testing. EVN describes testing that began in early 2026 and round-the-clock operation. These are statements from the project participants. The reviewed announcements do not provide enough public methodology to reproduce that percentage independently.
A language model is one component
The public card for one BgGPT release describes a text language model for Bulgarian and English. It does not establish which specific version is deployed in Eli. Neither announcement supplies a complete speech-stack specification, hardware configuration or hosting design. Applying the specifications of an arbitrarily selected public checkpoint to this deployment would therefore be unjustified.
Our editorial interpretation is that this type of service has several distinct correctness boundaries. Speech must first be understood. The identifier must then associate the conversation with the right property. Finally, an answer or a recorded report must correspond to the actual information in the utility's service systems. A confident spoken answer does not, by itself, validate any of those connections.
Consider a customer with two properties. Accurate transcription alone does not establish that the service checked the property without power. This illustrates a verification requirement; it is not an allegation of an observed fault in Eli. Such scenarios explain why a specific property identifier and access to human assistance matter operationally.
Another useful distinction is between generating a response and completing a transaction. A service report needs an unambiguous destination and an observable outcome. If the conversation ends without that outcome, the apparent fluency of the assistant is a weak measure of value. We are describing a general evaluation principle, not claiming access to EVN's internal implementation.
Interpreting the automation rate
Around 70% of calls handled without a person does not automatically mean 70% less waiting or 70% lower cost. Those conclusions would require additional measurements. The period, sample size, call mix, definition of completion and repeat contacts about the same issue all matter. The reviewed public material does not provide sufficient detail to calculate these outcomes.
A useful future evaluation could separately track correctly registered reports, responses to uncertain information and transfers to employees. A short automated conversation is valuable when it resolves the question or directs it appropriately. These are proposed assessment criteria, rather than findings from an SIExpo test. They also prevent a throughput measure from being mistaken for a complete assessment of customer experience.
The significance is the deployment of local language technology in a defined service process. The boundary is equally important: the assistant supplies information and receives a report, while restoring electricity remains a physical operational task. SIExpo has not placed test calls to Eli or measured its performance. The official announcements establish the deployment and its stated scope, while independent performance validation remains outside the evidence reviewed here.
The cover is a conceptual AI illustration, not a photograph of the actual deployment.
Editorial verification: 2026-11-14T17:12:57.436197+00:00.
