AI in Bulgaria's Audit Office: Why 4x Became 2x, and Why That Is Good News

Exactly a year ago, on 10 November 2025, INSAIT and Bulgaria's National Audit Office showed a pilot AI assistant for auditing public procurement. The promise then was that checking a single tender could drop from four days to one.

On 5 November 2026 the Audit Office announced that the system, now called ODIS ZOP, is going into real use. Its president, Dimitar Glavchev, quoted a different number: a check that currently takes anywhere from two days to more than a week will be cut in half. Auditor training starts this week, and AI will be used for procurement checks in the audits of the 2027 programme. The story also reached regional business media, including SeeNews on 9 November.

From "four times faster" to "twice as fast" in one year sounds like a retreat. We think it is the opposite, and that this is the most useful part of the story for any company currently listening to a vendor's promises.

What is actually going live

The facts, from the Audit Office announcement and BTA:

• ODIS ZOP is the first module of a broader project. It checks whether public contracts were awarded lawfully.
• It runs on BgGPT, the Bulgarian language model, and pulls the data for each procedure from the central e-procurement system (CAIS EOP): parameters, requirements, templates, contracts, bidders.
• Changes to the Public Procurement Act and its implementing rules are built into the system.
• For each audit question, the assistant proposes an answer with reasoning and a reference the auditor can verify.
• It looks for deviations from the legal framework, but only the auditor decides whether a deviation is a violation.
• Audit documents, checklists and findings are kept in an environment isolated from the internet.
• Planned next steps include processing large document sets, financial audits and support for audit planning.

At the pilot stage, INSAIT added that the system runs on GPUs at Sofia Tech Park and the BRAIN++ AI factory, so the data stays in Bulgaria.

Why the smaller number is the better sign

From here on, this is our interpretation. The Audit Office does not explain the difference between the two figures, so we are not claiming to know the exact reason.

Pilots are almost always measured on selected cases. Real work brings scanned documents, tenders split into many lots, contracts with amendments. It also brings the time a person needs to verify each suggestion the machine makes.

That last part matters most. ODIS is designed so that every suggestion comes with a source reference and an auditor confirms it. That costs time, and the cost cannot disappear if you want the conclusions to hold up. A 50% saving measured with verification included is a more useful number than a 4x saving measured without it.

So when a vendor promises "four times faster", the useful question is not whether the number is true. It is whether it was measured after a person checked the output, and on which cases.

The design worth copying

Most companies will never audit public tenders. Many, though, want AI to check something against rules: contracts against internal policy, invoices against purchase orders, bids against a technical specification. ODIS is a well-documented example of how to do that where mistakes are expensive.

The first choice is the split of roles. The machine extracts, compares and flags. The person qualifies. The word "violation" never comes from the model. In a private company, that is the difference between "the system flagged 12 contracts for review" and "the system rejected 12 contracts".

The second is traceability. A suggestion without a reference to a specific document and clause is an opinion. A suggestion with one can be checked in seconds. If your AI assistant cannot show where an answer came from, checking it will take as long as doing the work by hand.

The third is that someone maintains the rules. The Audit Office says explicitly that legal changes are built into the system. In a company, that means a named person responsible for keeping the policy in the model identical to the policy on paper. Without that owner, the system goes stale quietly.

The fourth is the data split. Inputs are public, while audit working papers stay isolated. Most companies face the reverse problem: they want to feed internal documents to a model and have not decided where those documents are allowed to go.

What it means for contracting authorities and bidders

ODIS audits contracting authorities: municipalities, hospitals, state-owned companies and other bodies that run tenders. For them, the practical point is direct. When a system compares document versions, reads scanned files and walks through every requirement, mismatches between the decision, the notice, the tender documents and the contract become easier to find. If you run tenders, it makes sense to do the same check before publishing rather than waiting for the audit.

For companies that bid, the effect is indirect and harder to predict. Our view is that more systematic audits will make contracting authorities stricter about formal requirements, and that usually reaches bidders too.

What to watch

Two things will show whether ODIS works as announced. First, whether the 50% saving holds after the first year of real audits in 2027. Second, whether INSAIT takes the model to other institutions, as its executive director Borislav Petrov said many of them have asked. If that happens, part of the compute will likely run through the AI factory we wrote about yesterday.

If you are planning AI that checks documents against rules in your own company, the hard part is not choosing a model. It is deciding who signs off and making every answer traceable to its source. That is what we work on in AI Governance.