EVN's Eli Only Answers 'Why Is My Power Out?' That Is the Point
Picture a winter evening somewhere near Stara Zagora. The lights go out, the heating stops, and half the street reaches for the phone at the same moment. Until now that meant waiting on the line for an operator who was answering the same question for the two-hundredth time that night: is this a known outage, and when will it be fixed?
On 12 November Elektrorazpredelenie Yug, the EVN Group distribution company for south-eastern Bulgaria, announced that this first conversation is now handled by a voice assistant called Eli. According to the company, Eli asks for the caller's metering point number (ИТН), checks the cause of the outage, and if needed registers a fault report on the grid. It runs on BgGPT, the Bulgarian language model developed by INSAIT, whose team helped adapt it for customer service. Testing started in early 2026. Callers are told at the start that they are speaking to a digital assistant and can be transferred to an employee. The company says about 70% of outage calls are now handled entirely without human involvement.
That is a narrow product. In my view, the narrowness is exactly why it is worth looking at.
Meanwhile, the broad agents are struggling
A Bloomberg Opinion column published in Bulgarian on the same day collected the less flattering numbers. Uber's CFO told investors that out of roughly 320 million trips and orders in the week of Halloween, about 500 food deliveries and four rides were requested through an AI agent. Commonwealth Bank of Australia dropped a plan to cut 45 customer-service roles after its AI agents led to a sharp rise in calls. The column also cites a Gartner analysis that fewer than 10% of AI agents deployed by companies deliver solid return on investment.
These are general-purpose agents: assistants that are supposed to book, order, compare and resolve almost anything. Eli is the opposite. Put side by side, the difference is mostly about what the system knows before the conversation starts.
| Eli (as announced) | A general-purpose agent | |
|---|---|---|
| Who is calling about what | The ИТН identifies the exact site | Has to work it out from free text |
| Where the answer comes from | The distributor's own data about the grid | Whatever it can find or infer |
| What it may do | Register one type of report | Many actions across many systems |
| When it stops | Hands over to a person | Often unclear |
| How you measure it | Share of outage calls closed without a human | Hard to define |
The table is my reading of the public announcement, not a description of EVN's internal architecture, which has not been published.
Why the narrow version has a chance
What makes Eli worth copying has little to do with the language model.
Start with the key. Asking for the ИТН at the start turns a vague complaint ("we have no power on our street") into a lookup against a specific point in the network. This is where many customer-service bots lose the conversation: they spend it trying to understand who the customer is and what object the problem is about.
Then there is the fact that the answer already exists somewhere. When the outage is planned or already known, the distributor has that information. The assistant's job is to fetch it and say it clearly in Bulgarian, not to reason its way to an answer. That is the kind of task where a voice assistant is reliable, and where a wrong answer is easy to spot in testing.
And the single action it can take is cheap to get slightly wrong. A duplicate fault report costs a dispatcher a minute. An agent that books the wrong flight costs a customer money and the company trust.
Add the clear disclosure and the transfer to a person, and you get something that looks less like a revolution and more like a well-designed phone menu that finally understands speech. I mean that as praise.
What the 70% does and does not tell you
The figure comes from the company and covers outage calls only, not all customer service. "Handled without human involvement" means the call ended without an operator. It does not by itself tell us whether the caller got what they needed, or how many called back later. Those are the numbers I would want to see after the first big storm, when the line is under the most pressure.
None of that makes the result less interesting. It just means it is a measure of call handling, and should be read as one.
Finding your own "Eli" call
Most companies with a phone line or a busy inbox have at least one request type that looks like this. A practical way to find it:
- Take a week of calls or tickets and sort them by reason. Not by department, by what the customer actually wanted.
- Look for a reason where the customer can give an identifier. An order number, a contract number, a car registration, a customer card.
- Check whether the answer already sits in one system. Delivery status, appointment time, invoice amount, a known outage. If an employee has to make a judgement call, it is not your first candidate.
- Limit the assistant to one action. Register a report, book a slot, send a document. Everything else goes to a person.
- Decide in advance what you will measure. Share of calls closed without a human, and repeat calls within 24 hours for the same identifier. The second number keeps the first one honest.
If the top candidate on your list has peaks, like outages after a storm or deliveries before a holiday, that is usually where an assistant pays off first, because that is when people wait longest.
This is the kind of scoping we do in our work on AI voice assistants for business: picking the one call type where the data and the action are clear before anyone talks about the model.
The headline numbers in AI this week were about agents that can do everything and are used by almost no one. The more useful story was a Bulgarian distributor's assistant that does one thing and picks up the phone at three in the morning.
