Main conclusion
BTU researchers conclude that the central AI challenge for Georgian businesses is no longer access to the technology. The challenge is proving that completed work becomes faster, cheaper or better. A company may own many AI licences and generate more text and analysis while achieving no financial gain if review, rework, poor data or additional supervision absorb the saved time.
No nationally representative public study currently measures AI return across Georgian companies through a common set of process-level indicators covering time, full cost, error and additional revenue. It is therefore too early to give a general yes-or-no answer on whether Georgian companies obtain real gains. The result has to be demonstrated process by process against a baseline and, where possible, a control group.
BTU researchers propose a strict definition: AI counts as productivity only when full cycle time or full cost falls at unchanged quality, or when the same resources produce more accepted work, contribution margin or service quality. Adoption, prompts and generated output are implementation indicators rather than proof of impact.
Why tool counts mislead
During the first stage of the AI boom, companies often measure what is easiest to count: licences, active users, prompts and generated pages. Those metrics describe activity. They do not establish productivity.
An employee may produce a first draft 30 minutes faster while a manager spends 40 additional minutes checking facts and rewriting it. A support team may respond more quickly but create more repeat contacts. Developers may generate more code without completing projects faster once review and rework are included. The correct unit is the accepted business outcome: a resolved request, processed document, paid invoice, retained customer, completed project or additional contribution margin.
What the international evidence shows
International evidence is deliberately mixed. Outcomes depend on the task, the worker’s experience, the data environment and whether review and rework are counted.
| Evidence | Observed result | Implication for Georgia |
|---|---|---|
| NBER and the Quarterly Journal of Economics | Among 5,179 customer-support agents, productivity increased by 14% on average and by 34% for novice and lower-skilled workers | AI can spread best practice in standardised high-volume processes |
| OECD survey of SMEs | Generative AI was used by the respondent or a colleague in 31% of SMEs across seven countries; 65.1% of users reported improved employee performance, but the magnitude was not measured | Reported improvement is not the same as verified financial productivity |
| METR experiment with experienced developers | Across 246 real tasks completed by 16 developers, early-2025 AI tools increased completion time by 19%, while participants believed they were 20% faster | Perception can move in the opposite direction from measured performance |
| METR update in 2026 | Newer tools showed signs of possible speed gains, but selection and measurement problems prevented a confident estimate | Fast-changing tools require repeated measurement |
The evidence does not say that AI always accelerates or always slows work. It shows that productivity is a property of the task and organisational design. International averages should not be imported directly into a Georgian company’s business case.
The Georgian measurement gap
Geostat publishes indicators on enterprise use of software, the internet of things and artificial intelligence. This is important for tracking diffusion, but adoption and return are different questions. Public data does not reveal how much a company reduced the full cycle time of an operation, lowered error costs or created additional margin.
BTU researchers identify three risks that are particularly relevant in Georgia. In small teams, saved time is easily absorbed by unrelated work. Georgian-language and local-context output may require additional review. Fragmented data can leave the overall process unchanged even when the AI response itself is fast.
Four tests of a real result
| Test | Correct measure | What should not be confused with a result |
|---|---|---|
| Time | Full cycle time from the first step to final acceptance | Generation speed |
| Cost | Cost per completed case including software, integration, data, training, review and rework | Monthly subscription price alone |
| Error | First-pass accuracy and full correction time | A fluent or confident answer |
| Revenue | Additional contribution margin, conversion or customer retention attributable to the process | Number of leads or messages generated |
Quality is a guardrail across all four tests. Time saved is not real if customer satisfaction, compliance or decision accuracy deteriorates. Revenue growth should be evaluated through margin rather than turnover alone.
A ninety day productivity test
- Use the first four weeks to establish the baseline for time, volume, first-pass accuracy, rework, quality and full cost.
- Select one clearly defined process and specify where AI may act and where human judgment remains mandatory.
- Run a six-week pilot with comparable cases and, where possible, a control group or alternating AI and non-AI workflow.
- Count data search, prompting, review, correction and supervisory time as part of the process.
- Use the final two weeks for verification and a decision to scale, redesign or stop.
A successful demonstration is not production evidence. The pilot should be bounded, measurable, reversible and linked to a specific model, workflow, data set and accountable owner. A material system change requires remeasurement.
An illustrative Georgian calculation
Geostat reported average monthly remuneration of GEL 2,471.5 in the business sector in the second quarter of 2026. Using 176 hours per month gives a simple wage proxy of GEL 14.04 per hour. It is not a fully loaded employer cost, but it provides a transparent starting point.
Assume a service process with 20 employees who each save one hour a day over 220 working days. The theoretical annual time value is GEL 61,788. If the company converts only 60% of that capacity into productive work, monetisable labor value is GEL 37,073. A further GEL 6,000 is saved through lower error and rework, producing total annual benefit of GEL 43,073.
If the full annual cost of licences, integration, training and human oversight is GEL 26,000, BTU researchers calculate a net benefit of GEL 17,073, a return on investment of approximately 65.7% and payback in about 7.2 months. Under those assumptions, the system creates a real economic result.
| Measure | Productive scenario | False productivity scenario |
|---|---|---|
| Theoretical annual time value | GEL 61,788 | GEL 61,788 |
| Time converted into productive capacity | 60% | 20% |
| Monetisable labor benefit | GEL 37,073 | GEL 12,358 |
| Error and rework savings | GEL 6,000 | GEL 2,000 |
| Full annual AI cost | GEL 26,000 | GEL 26,000 |
| Net result | GEL 17,073 | Minus GEL 11,643 |
| Return on investment | About 65.7% | About minus 44.8% |
Both scenarios claim the same one-hour daily saving. The difference is whether the company uses the capacity, reduces actual error and counts full cost. The calculation is illustrative rather than a finding about a specific firm.
Where genuine gains are most likely
The strongest early opportunities in Georgia are frequent, repeatable processes with digital inputs and verifiable outputs: first-line customer support, document classification, sales preparation, accounting reconciliation, internal knowledge retrieval, reporting drafts and standard communication.
Benefits can be especially large for new or less experienced employees because AI can expose them to the organisation’s best existing examples. Gains may be smaller or negative for experts if the tool produces shallow suggestions that require repeated rejection or correction.
Where the productivity illusion appears
- Generation time is counted while checking and rework are ignored;
- Intermediate actions increase without an increase in completed cases;
- AI creates more analysis than decision-makers can use;
- Saved time is not converted into additional service, sales, avoided hiring or lower overtime;
- Quality losses appear later through complaints, returns, audit findings or employee overload;
- A new model is introduced while productivity assumptions from the previous system remain unchanged.
What management should require
BTU researchers recommend a one-page result statement for every AI project. It should contain the baseline, pilot period, control comparison, quality threshold, full cost and monetised benefit. If the owner cannot name the completed unit of business output, productivity measurement is not ready.
Management should also separate employee convenience, operational productivity and financial return. Convenience may be valuable for retention and work quality, but it cannot automatically substitute for evidence on output and economics.
Final assessment
Some Georgian companies are almost certainly obtaining real value from AI, particularly in repetitive and text-intensive processes. Public evidence does not yet establish how widespread or large that value is, or how often benefits exceed full costs.
The key business question in 2026 is not whether a company uses AI. It is how much accepted work it completes per hour and per lari at unchanged quality, how many errors it avoids and how much additional margin it creates. Those measures will separate the AI boom from real productivity.