Bulgaria's National Audit Office expects to halve the time its auditors spend reviewing a single public-procurement procedure. The current review takes anywhere from two days to more than a week depending on complexity. A week after that number surfaced, EVN's distribution arm announced that about 70% of power-outage calls to its new voice assistant Eli can be handled without a human operator. Two claims, two very different operational contexts, one underlying model: BgGPT, built by INSAIT at Sofia University.
Those numbers deserve attention. They also deserve scrutiny.
On November 5, INSAIT and the National Audit Office (BNAO) announced ODIS PPA, an AI assistant that retrieves publicly available procurement data from Bulgaria's central e-procurement system: procedure details, requirements, templates, contracts, bidder information. It processes documents, proposes answers to audit questions, provides explanations with source references, and flags possible inconsistencies against the legal framework. The system is planned for use under BNAO's 2027 Annual Audit Programme.
BNAO was explicit about one boundary: the AI does not decide whether a discrepancy constitutes a violation. Audit analysis, findings, and conclusions remain the auditors' responsibility. That distinction matters more than it sounds.
On November 12, EVN Bulgaria's Electrodistribution South launched Eli, a Bulgarian-language digital voice assistant for customers reporting power outages. Eli can check for known causes, register a fault report, request the customer's measuring-point number, and transfer to a human operator when needed. It runs 24/7, and callers are told before the conversation starts that they are speaking with a digital assistant. According to 3e-news.net and Mediapool, EVN reports that roughly 70% of outage-related calls can be handled entirely by AI. The reporting does not clarify whether "handled" equals "resolved," nor does it disclose the measurement period, call volume, or calculation method.
A local-language model moving from research demos into procurement audits and utility call centers is genuinely significant. BgGPT is also a public project: INSAIT has offered it as a free public chat since 2024. But model capability alone explains very little about operational success.
The BNAO deployment gets this right in at least one dimension: human accountability is designed in from the start. The AI proposes; auditors decide. Source references are surfaced so findings can be verified. That is not a limitation of the technology. It is a governance choice, and a sound one for any process where the output carries legal weight.
The EVN case is harder to evaluate. A 70% containment rate during outage calls sounds impressive, especially during severe weather when call volumes surge. But containment and resolution are different things. Did the caller get accurate information? Was the fault report correctly logged? Did the customer call back ten minutes later and reach a human anyway? Without those metrics, the number tells us something about call routing efficiency and very little about service quality. Anyone who has run a contact center knows the difference.
Bulgaria's Institute of Public Administration published a 2026 AI Readiness Index based on 5,475 public employees across central, regional, and municipal levels. The finding: the administration remains at a "capacity-building" stage. Barriers include limited access to approved tools, unclear internal rules, and weak knowledge-sharing.
That context reframes the BNAO announcement considerably. Having a capable model is one thing. Having the organizational infrastructure to deploy it safely across dozens of audit teams, with consistent training, clear escalation paths, and reliable data pipelines into the e-procurement system, is something else entirely. The gap between a working pilot and a production system that survives contact with real institutional complexity (staff turnover, inconsistent data quality, shifting legal frameworks) is where most AI projects quietly stall.
This is not a criticism of BgGPT or INSAIT. The model quality may be excellent. The point is that model quality is a necessary condition, not a sufficient one.
The pattern here is familiar to anyone working on AI operationalization in regulated environments. A capable model gets connected to a real workflow. Early numbers look promising. The press release goes out. And then the slower, unglamorous work begins: defining what "handled" actually means, building logging and traceability so outcomes can be audited, designing fallback paths for the cases the model gets wrong, and establishing ownership so someone specific is accountable on a Tuesday in November when the system behaves unexpectedly.
Since August 2, 2026, the EU AI Act's transparency rules apply: an AI system that talks to people, a voice assistant included, must tell them they are dealing with AI. EVN already does this. The heavier duties, logging and human oversight, apply to high-risk systems, and the Digital Omnibus moved them to December 2, 2027. That is close enough to design for now, and the decisions that will count are made in the integration layer: routing logic, escalation triggers, audit trails, policy enforcement. The model sits inside that layer. It does not replace it.
At KONDEVS, this is precisely the work we do: building the integration and governance architecture that lets AI operate safely inside enterprise workflows. Routing, observability, fallback design, accountability. The control plane that makes the difference between a promising demo and a system you can defend in an audit.
Half the review time. Seventy percent of calls contained. Both numbers are expectations or early reports, not independently verified production metrics. They may prove accurate. They may prove optimistic. The real measure will come in twelve months, when BNAO publishes its first audit results using ODIS and EVN has a full winter of storm-season call data.
What Bulgaria is doing, quietly, is putting a sovereign language model into two workflows where failure has real consequences: public money and electricity supply. The ambition is worth respecting. The execution will be worth watching closely, because the gap between a working model and a governed, observable, accountable production system is where the actual engineering happens.
