One billion lari is a dramatic number, but it is equal to roughly 2.8% of Georgia’s GEL 36.285 billion consolidated budget for 2026. The question is therefore no longer whether a billion-scale effect is imaginable. The real question is what counts as a saving and what kind of institutional system would be required to produce it.
A chatbot that summarises documents will not save a billion. A public-finance system that connects procurement prices and contracts, tax risks, benefit records, medical claims, infrastructure conditions and administrative workflows could find what fragmented government often misses: inflated prices, suspicious bidding patterns, duplicate or incorrect payments, unaddressed tax risks, delayed maintenance and thousands of hours of repetitive work.
According to BTU researchers, Georgia could reach a billion-lari scale only over several years and through the combined effect of several reforms. Some gains would be cash savings, some prevented losses, some additional lawful revenue and some productive capacity. Combining all four into one headline number would exaggerate the result. Measuring them separately could produce a genuine fiscal reform.
“Savings” is not one thing
Public discussion tends to place every positive outcome under the label of savings. For a budget, however, four effects must be separated.
Direct savings occur when the state buys the same output for less. Prevented losses occur when an incorrect, excessive or fraudulent payment is stopped. Revenue gains arise when better risk analysis improves the collection of taxes that are already legally due. Productivity gains occur when employees complete the same work faster.
The fourth category is the most easily overstated. If a civil servant processes a document in one hour instead of four, time has been saved, but cash does not automatically return to the Treasury. A fiscal effect appears only if the time reduces overtime, replaces outsourced work, avoids additional hiring or increases service capacity without additional cost.
Responsible reporting should therefore show cash savings, prevented losses, additional revenue and released capacity separately.
Procurement offers the largest immediate opportunity
Contracts worth approximately GEL 6.4 billion were signed through Georgia’s public procurement system in 2025. At this scale, a small percentage becomes a large number. A 3% effect from better market research, demand aggregation, price comparison, anomaly detection and contract monitoring would equal about GEL 192 million a year. A 5% effect would equal GEL 320 million, and 7% would equal GEL 448 million.
These are scenarios, not guaranteed forecasts. The World Bank reports that Bangladesh’s electronic procurement reform reduced procurement costs by about 7% compared with paper-based processes. That result came from broad digital reform, not AI alone. Georgia already has a strong electronic procurement foundation; the next step is to use the resulting data more intelligently.
AI could compare prices for identical items across agencies, identify specifications that appear tailored to one supplier, detect suspicious patterns among connected bidders, estimate the delivery risk of abnormally low offers and indicate where consolidated purchasing may be preferable. Human officials should retain the decision, but audit attention can move from random checks toward the highest-risk transactions.
Better tax administration is the second source
Georgia’s 2026 consolidated budget projects GEL 27.04 billion in tax revenues. On this base, even a small improvement in compliance matters: 0.5% is equivalent to GEL 135.2 million, 1% to GEL 270.4 million and 2% to GEL 540.8 million.
This does not require higher tax rates or more intrusive audits of every business. Better analytics should reduce unnecessary scrutiny of compliant taxpayers and direct investigators toward declarations where data reveal a genuine inconsistency. International administrations use data matching and machine learning to identify false declarations, abnormal transactions, VAT fraud and undeclared income.
False positives are the principal risk. An unusual but legitimate business can appear suspicious to a model. An algorithmic flag must therefore determine review priority, not become a fine or a legal conclusion. Evidence and a human decision must follow.
Health, social programmes and infrastructure
Georgia’s 2026 allocation for the Ministry of Health and Social Protection exceeds GEL 9.6686 billion. Much of this spending represents essential rights and services, so the objective cannot be to cut benefits. The opportunity is to identify incorrect claims, duplicate applications, unusual medical billing, mismatches between programmes and unjustified price variation.
Protecting or using even 1% of this broad allocation more effectively would represent GEL 96.7 million; 2% would represent GEL 193.4 million. These figures require exceptional caution. They are scenarios, not a claim that these amounts are currently being lost. The relevant expenditure base must be established programme by programme through audits and pilots.
In infrastructure, the benefit of AI often lies in timely maintenance rather than cheaper contracting. Detecting damage to a road, bridge, water network or public building early can be far less costly than emergency reconstruction. More than GEL 8 billion is planned for infrastructure development in the 2026 consolidated budget. An effect equivalent to 0.5% of that scale is GEL 40 million; 1% is GEL 80 million. Each case would still need evidence that predictive maintenance reduced an identifiable capital or operating cost.
Is one billion realistic?
To illustrate the scale, the BTU research team calculated three scenarios. They are not an audit or an official forecast, and overlaps between categories mean that simple addition must be treated carefully.
The conservative scenario combines a 3% procurement effect, 0.5% additional tax compliance, protection of 1% of the health and social allocation, a 0.5% infrastructure effect and productivity equivalent to 2% of the GEL 3.4 billion public payroll. The mathematical total is approximately GEL 532 million.
The central scenario uses 5%, 1%, 2%, 1% and 3%, respectively. It produces approximately GEL 966 million – effectively the scale of one billion. A high scenario using 7%, 2%, 3%, 2% and 5% reaches approximately GEL 1.61 billion, but would require a much deeper institutional transformation and strict independent verification.
The crucial caveat is that productivity value is not cash and tax revenue is not expenditure savings. A defensible conclusion is that an AI-enabled fiscal reform could create an annual fiscal and productivity effect of around one billion lari or more after several years. The amount that remains as cash in the Treasury would be smaller and must be reported separately.
Why one giant system would fail
A central “super-AI” built around a billion-lari promise would be a high-risk project. Agencies hold different data, exercise different legal powers and make different decisions. Procurement risk, medical billing and benefit eligibility cannot rely on the same model or the same standard of intervention.
A more realistic architecture is a common data and governance standard supporting several specialised systems. Each needs an explicit mandate: which data it may read, which signal it may produce, which action it may take and where human intervention is compulsory. An AI agent may review a contract, compare prices and draft a risk explanation. It should not stop a payment or remove a benefit solely on the basis of an opaque score.
This reflects a principle developed in BTU’s research materials on agentic management: technical capability is not authority to act. In the public sector, that principle must be reinforced by appeal rights, privacy protection, algorithmic transparency, independent audit and a complete record of consequential decisions.
A possible role for BTU
BTU’s role would not be to make government decisions. It could provide an independent research and testing environment that brings together economists, data scientists, cybersecurity specialists, lawyers and public-administration researchers.
One project could be a research laboratory for anonymised procurement data, examining price variation, competition, contract modifications and delivery risks. A second could focus on AI analysis of Georgian public documents. Administrative texts, laws, budget programmes and procurement specifications must be understood correctly in Georgian; simply translating them through a foreign-language model is not enough.
BTU’s work on teaching Georgian to AI and building a Georgian knowledge bank is therefore not only a cultural initiative. It is also infrastructure for fiscal technology. If a model cannot distinguish Georgian legal terminology, institutional phrasing and economic context, it cannot compare documents reliably and may generate costly errors.
BTU could also develop a Fiscal AI Evaluation Protocol. Before a pilot begins, it would define the baseline and then independently measure accuracy, false positives, cash savings, prevented losses, additional revenue, working time, citizen impact and appealed decisions. That architecture would turn technological enthusiasm into an accountable result.
How Georgia should begin
The correct starting point is not the purchase of generative AI licences for every agency. Georgia should select three to five areas where data quality is adequate, expenditure is material, decisions are repeated and results can be measured within six to twelve months.
Procurement pilots could compare a limited number of high-volume categories with market prices. Tax systems could determine audit priority only. Health pilots could detect billing anomalies without replacing clinical judgment. Infrastructure pilots could predict maintenance needs for selected roads or bridges. Administrative assistants could accelerate document search, comparison and drafting.
Every pilot needs a control group or historical baseline. Failures should be published alongside successes. A system that saves GEL 10 million but unfairly delays thousands of citizens is not efficient. A model that creates more risk alerts than investigators can process merely builds a longer queue.
Conclusion
AI can create a billion-lari-scale effect for Georgia’s budget, but not by installing software and declaring victory. The result begins with quality data, a narrow mandate, a responsible institution, final human control and agreed measures of success.
Georgia has a particular opportunity. Its budget is large enough for small percentage improvements to create hundreds of millions of lari, while the country is small enough for a successful pilot to scale relatively quickly. The danger is to announce the billion first and later combine saved hours, prevented payments and additional taxes into one inflated number.
The stronger objective is more demanding: the state should know where each lari was saved, why it was saved, who was affected and whether the same result could have been achieved without AI. Only then does artificial intelligence become a budget instrument rather than an expensive technological promise.
Sources
- Ministry of Finance of Georgia and Parliament of Georgia – 2026 state and consolidated budget parameters.
- State Procurement Agency of Georgia – 2025 procurement and activity data.
- OECD – Governing with Artificial Intelligence, 2025.
- World Bank – evidence on electronic government procurement.
- UK Government – AI Playbook, Algorithmic Transparency Recording Standard and public-sector fraud-risk tools.
- BTU research materials – agentic management, mandate, accountability and Georgian knowledge infrastructure.