AI and automation in Uzbekistan: the adoption gap, and where it pays first

Uzbekistan has spent a decade wiring itself up. There are more mobile subscriptions than people, nine in ten people are online, and the state runs one of the more advanced e-government systems among emerging markets. On artificial intelligence, however, the country has barely started. That combination, strong digital rails and weak AI use, is unusual, and it says a good deal about where the opportunity lies for companies operating there.

This note sets out what the data show about AI use in Uzbekistan, how fast the state is moving, why local wages change the business case, and where automation tends to pay first. It ends with the constraints a project should be designed around.

An online country that has not yet adopted AI

Microsoft’s AI diffusion series estimates the share of working-age people who used a generative-AI product during each period. In the second quarter of 2026, Uzbekistan’s figure was 7.6%, up from 7.2% in the first quarter. The world average was 18.8%.

The obvious explanation would be access, and it does not hold. The World Bank puts internet use in Uzbekistan at 89.5% of the population in 2024, higher than the world average of 73.6%. Set the two series side by side and the difference is stark: for every 100 people online, about 8.5 use generative AI in Uzbekistan, against roughly 25 worldwide.

In other words, Uzbekistan converts internet access into AI use at about a third of the world rate. Some of the gap is measurement: Microsoft’s figures are built from its own telemetry, and use through local services or messaging bots may be under-counted. But the size of the difference, and the fact that Kazakhstan, with similar connectivity, is more than twice as high, suggests the gap is real. The likely reasons are familiar ones: language (most tools work best in English and Russian, less well in Uzbek), cost relative to local incomes, and the simple fact that most businesses have not yet built AI into the way they work.

For a company, an adoption gap of this size is both a warning and an opening. Customers and staff are not yet expecting AI-enabled service, which lowers the bar for those who move first.

The state is moving faster than business

On policy, Uzbekistan is well ahead of its private sector.

Two details stand out. First, the readiness ranking has improved by 25 places in two years, and the official target of the top 50 is now twelve places away. Second, compute is starting to arrive at home. DataVolt’s Tashkent project has up to $150 million of financing from the EBRD, DEG, Proparco and the OPEC Fund, and the Ministry of Digital Technologies told Euronews that its new government computing cluster reached 90% utilisation within its first month. For businesses that need to keep data in the country, local capacity matters, as the section on constraints explains.

The practical consequence is that public-sector demand, from AI-powered state services to the 2030 software target, is likely to grow faster than private demand for some time. Firms that build skills now will find both customers and partners.

Wages change the business case

In high-wage economies, much of the case for automation rests on hours saved. In Uzbekistan the same tool has to clear a different bar, because an hour of labour costs far less.

The National Statistics Committee’s preliminary figure for the average monthly wage in 2025 was about 6.38 million soums, around $507 at the year’s average official exchange rate. In Tashkent city it was about 10.75 million soums, and in banking and insurance, the best-paid sector, about 17.6 million.

Eight hours a month is not a high bar for a tool that drafts, summarises or answers routine questions. But the table points to where the case is strongest: in Tashkent, in finance and in other well-paid knowledge work, the payback on hours alone is quick. Elsewhere, the stronger arguments are rarely about headcount. They are about response time, error rates, working in three languages without three teams, and growing volume without hiring specialists who are scarce and expensive to keep.

That framing also fits the evidence on jobs. The International Labour Organization estimates that about a quarter of global employment is in occupations with some exposure to generative AI, with clerical work the most exposed, and concludes that transformation, not replacement, is the most likely outcome. Uzbekistan’s services sector already employs about half the workforce, according to the World Bank, so the scope for augmenting office work is large.

Where automation tends to pay first

The use cases that justify themselves earliest in the Uzbek market share three features: high volume, measurable output and a language burden.

None of these require a research laboratory. Most combine a commercial AI model with the company’s own data and its existing software, which is why AI projects in practice are often as much software and integration projects as AI ones.

Constraints to design around

Data localisation. Since April 2021, Article 27-1 of the Law on Personal Data has required personal data of Uzbek citizens processed with information technologies to be collected, systematised and stored on technical means physically located in Uzbekistan, with databases registered in a state register. For AI projects this shapes the architecture: keep the master customer data onshore, send foreign AI services only what they need (ideally anonymised or minimised), or run models on local infrastructure, which the new data centres make more practical.

The new AI law. Law ZRU-1115 of 21 January 2026 put a definition of AI into the Law on Informatization, set general rules for using AI in information resources and systems, and introduced administrative liability for unlawful processing of personal data using AI. It amends existing law rather than creating a standalone AI act, so the familiar rules on personal data, consumer protection and sector licensing continue to apply alongside it.

Language and data quality. Uzbek is a lower-resource language for most models, and Latin and Cyrillic scripts coexist. Anything customer-facing should be tested with real local messages before launch, and documents often need cleaning before they can be read reliably.

Skills and ownership. The most common failure is not technical. It is a pilot without an owner, a measure of success or a plan to scale.

What the global evidence says about getting value

McKinsey’s 2025 survey found that 88% of organisations regularly use AI in at least one business function, up from 78% a year earlier. Yet nearly two-thirds had not begun scaling it across the enterprise, and only 39% reported any impact on earnings at the enterprise level. The lesson is that experimenting is easy; value comes from redesigning a specific workflow around the tool and measuring the result.

For Uzbek companies starting later, this is an advantage. They can skip the scattered pilots and begin with one or two workflows that carry real volume.

A practical sequence

1. Choose one or two workflows with measurable volume, such as enquiries a month, documents processed or reports produced, and record today’s baseline for time, cost and errors.
2. Check data and compliance first: what personal data is involved, where it is stored, and what may leave the country.
3. Build into the tools people already use, whether that is the website, a messaging channel, the CRM or the accounting system, rather than adding another screen.
4. Measure after six to eight weeks against the baseline, keep a person responsible for quality, and only then extend to the next workflow.

GSI helps companies in Uzbekistan through each of these steps, from choosing the workflow to building and integrating the system: see our AI solutions and automation and software and app development services. For the wider economic context, see our analysis of Uzbekistan’s growth against the world and of ten quietly growing industries, which include IT services exports.

This article is for general information and does not constitute investment, legal or technology advice. Figures are taken from the sources listed below and may be revised.