Million Hours in Georgia’s Public Sector
Georgia’s state-owned sector employed 334,700 people in 2025, equal to 24.1% of total employment. If AI saved each worker only two hours a week and the gain held for 48 working weeks, the sector would free 32.1 million hours a year. That equals roughly 4.0 million eight-hour workdays or 16,735 full-time work-year equivalents.
| Key finding Even one hour saved per employee per week would release 16.1 million hours annually. Two hours produces 32.1 million; four hours produces 64.3 million. |
These figures are neither a forecast nor a redundancy plan. They are mechanical measures of capacity: time that could move from drafting, minute-taking, search, summarisation, translation and data entry to citizen service, complex cases, oversight and policy analysis.
Why two hours a week is a plausible benchmark
A 2025 UK government experiment involving about 20,000 civil servants across multiple departments found that Microsoft 365 Copilot users saved an average of 26 minutes a day. Across a five-day week, that is about two hours and ten minutes. More than 70% of users reported spending less time searching for information and completing mundane tasks, leaving more time for strategic work.
The central Georgian scenario therefore uses two hours a week, while the model presents a range. One hour is cautious; two hours is close to the international public-sector trial; four hours requires deep workflow integration rather than access to a stand-alone chatbot.
Structured data: model inputs
| Indicator | Value | Status / formula | Source |
|---|---|---|---|
| Public-sector employment | 334,700 | 2025; based on official data | Geostat |
| Share of total employment | 24.1% | 334.7 / 1,389.7 | Geostat; independent check |
| Total employment | 1,389,700 | 2025; official | Geostat |
| Working weeks per year | 48 | Scenario assumption | BTUAI Research Team |
| Working day | 8 hours | Scenario assumption | BTUAI Research Team |
| Full-time work-year | 1,920 hours | 40 hours × 48 weeks | BTUAI Research Team |
Note: the state-owned sector is broader than central administration and includes parts of public education, health and other state organisations.
Three scenarios for annual time released
| Scenario | Saving per employee | Annual hours | Workdays | FTE work-years |
|---|---|---|---|---|
| Cautious | 1 hour/week | 16.07m | 2.01m | 8,368 |
| Base | 2 hours/week | 32.13m | 4.02m | 16,735 |
| High | 4 hours/week | 64.26m | 8.03m | 33,470 |
Formula: 334,700 workers × weekly hours saved × 48 weeks. Workdays = hours / 8; FTE work-years = hours / 1,920. The result is a capacity equivalent, not a headcount estimate.
Applying the UK trial’s 26-minute daily benchmark mechanically would produce about 34.8 million hours, 4.35 million workdays or 18,130 FTE work-years. But a foreign trial cannot be transferred automatically: process digitisation, data quality, Georgian-language performance and security constraints will differ.
Where time is likely to be saved first
The quickest gains are likely where staff already work with digital text and repeatable rules: first drafts of correspondence, meeting minutes, long-document summaries, regulatory search, standard responses, translation drafts, application classification, extraction from tables and routing citizen requests to the correct unit.
High potential does not justify automating high-stakes decisions. Decisions involving benefits, taxes, licences, policing, healthcare or legal rights need human authority and review. AI may prepare a case file or flag inconsistency, but citizens need explanation, appeal and a path to human reconsideration.
Saved time is not automatically a budget saving
If AI helps an employee finish a task two hours faster, payroll does not fall immediately. The software has licence and infrastructure costs, and outputs still require verification. A financial gain appears only when freed capacity reduces overtime, avoids future headcount growth, lowers outsourcing or allows more services with the same resources.
Agencies must therefore measure more than time: cycle time, backlog, first-contact resolution, error and rework rates, citizen satisfaction and time redirected to complex work. Faster text generation is not useful if it merely produces errors faster.
How Georgia should test the result
The practical route is a controlled three-to-six-month pilot across several types of agencies. Select high-volume, low-risk tasks; record baseline time and quality; establish a comparison group; deploy AI in an approved environment; and measure time, accuracy, rework and user outcomes together.
| Metric | Baseline | After AI | Success test |
|---|---|---|---|
| Document preparation | Minutes/case | Minutes/case | Time falls; quality holds |
| Error and rework | % of cases | % of cases | Stable or lower |
| Backlog | Number of cases | Number of cases | Sustained decline |
| Citizen outcome | Deadline/satisfaction | Deadline/satisfaction | Improves |
| Freed capacity | Hours/employee | Hours/employee | Moves to higher-value work |
The pilot requires data classification, an approved secure environment, audit logs, human accountability and random quality audits.
A credible national target
For year one, a credible target is not a promise of four hours but one reliably demonstrated hour a week. Even that releases 16.1 million hours-more than two million workdays. If pilots preserve quality and repeat the two-hour result, potential rises to 32.1 million hours.
The public value of AI should ultimately be measured in shorter waits, fewer errors and more human attention for difficult cases-not in headcount reduction. Hours saved are only an intermediate metric; the final result must be a better-functioning state.
Sources
Geostat – Employment and Unemployment
https://www.geostat.ge/en/modules/categories/683/Employment-Unemployment
Gnomon Wise – The Role of the Public Sector in Employment, 2 July 2026
https://gnomonwise.org/en/publications/analytics/328
UK Government – Microsoft 365 Copilot Experiment: Cross-Government Findings Report
https://www.gov.uk/government/publications/microsoft-365-copilot-experiment-cross-government-findings-report/microsoft-365-copilot-experiment-cross-government-findings-report-html