What the study actually examined
A study published in the journal World on June 18, 2026 looks beyond AI adoption rates. It examines how different forms of AI use relate to three business outcomes: labor productivity, business-process efficiency, and lower operating costs.
The authors collected 228 valid questionnaires from Croatian small and medium-sized enterprises. Services accounted for 68% of the sample, industry for 22%, IT for 6%, and agriculture for 4%. Respondents were owners, managers, or people involved in business decisions. The researchers used multiple linear regression.
The practical conclusion for managers is straightforward: economic value is associated primarily with where AI is embedded in work. Buying another assistant or increasing the number of experiments does not automatically improve an end-to-end process.
Where the relationship was stronger
The researchers grouped AI use into operational applications, support for managerial decision-making, and growth in AI use over time.
Operational use—automation of repeatable work and data analysis—showed the strongest relationship with business outcomes. Decision support was also significant, but weaker. Simply expanding the number of AI applications did not have a statistically significant relationship with business-process efficiency.
The regression models explained 37.4% of the variation in reported labor productivity and 34.2% of the variation in operating-cost reduction. For process efficiency, explanatory power was much lower at 15.9%. This does not mean that AI “raises productivity by 37.4%.” R² describes the model’s explanatory power, not the savings any individual company should expect.
This distinction matters when building an investment case. The study identifies positive associations, but it does not provide a universal ROI percentage and cannot replace measurement within a company’s own workflow.
Why a process may still fail to speed up
AI may prepare a response draft, classify a request, or extract fields from a document in seconds. Yet the total cycle time will not improve if the result then waits in a manual approval queue, is copied between disconnected systems, or must be redone because the source data is unreliable.
The unit of measurement should therefore be the whole process rather than model latency:
- time from task arrival to completion;
- share of outputs accepted without rework;
- number of manual touches and system switches;
- cost per correctly completed operation;
- frequency of errors and exceptions;
- employee workload before and after deployment.
If the model saves five minutes while approval adds a day, the demo may look impressive but the economics do not.
A practical implementation architecture
For an SME, the first economically testable deployment rarely requires an autonomous “digital department.” One repeatable process with a clear input and outcome is enough.
A workable architecture can have five layers:
1. Data source: CRM, service desk, email, ERP, document database, or file storage.
2. Preparation: field normalization, duplicate removal, required-field validation, and access control.
3. AI function: classification, extraction, knowledge retrieval, response drafting, or prediction.
4. Business rules: permitted actions, confidence thresholds, exception routing, and audit logs.
5. Human and system of record: approval of material actions and writing the result back to the operational system.
A local model or protected private environment is appropriate when prompts contain trade secrets, personal data, contracts, internal pricing, or technical documentation. For a de-identified pilot with variable demand, an API or cloud service may be faster and less expensive. Infrastructure should be selected after evaluating data sensitivity and workload, not by choosing the largest model first.
Data required for a credible pilot
Before implementation, the team needs a small but representative set of real tasks. A customer-service use case might start with 200–500 historical requests that include the correct category and final response. A document workflow needs different templates, scans, exceptions, and known failure cases. A RAG system needs approved documents with owners, revision dates, and access rules.
The minimum control set should include:
- reference examples for quality evaluation;
- a list of critical failure types;
- an owner for the business metric;
- logs of model, prompt, and source versions;
- rules for storing inputs and outputs;
- a fallback to manual processing.
Without these elements, a team can compare impressions but not economic outcomes.
How to calculate economics without fooling yourself
The basic formula is simple:
**Monthly impact = hours saved × fully loaded hourly cost + avoided losses − operating expenses.**
Operating expenses include more than tokens or GPU rental. Integration, monitoring, knowledge-base maintenance, quality review, exception handling, security, and process-owner time all belong in the calculation.
Consider a modelled example. Eight employees save 30 minutes per working day. Over 20 working days, that is 80 hours per month. At a fully loaded cost of RUB 1,000 per hour, the gross impact is RUB 80,000. If monthly operation costs RUB 60,000, the net impact is RUB 20,000. A RUB 300,000 pilot would take about 15 months to recover. If measured savings reach 120 hours instead, payback falls to roughly five months.
These figures are not market prices or a performance promise. They illustrate sensitivity. Before buying permanent infrastructure, validate three variables: actual time saved, acceptance rate, and full operating cost.
The Russian market context
Russian market data suggests that infrastructure choices are increasingly driven by the economics of experimentation. A cloud-market study reported by ComNews says 24% of surveyed companies already run AI workloads in the cloud, while another 22% are piloting them or plan to start within a year. At the same time, 41% obtain only computing infrastructure from providers and implement AI themselves.
There is an operational downside: 57% manage multi-cloud environments manually, while only 10% report mature cloud-cost management. Cloud can lower the entry cost of a pilot, but it does not eliminate FinOps, consumption limits, or budget ownership. A local environment provides stronger data control but shifts hardware, maintenance, and upgrade costs to the company.
The Croatian SME survey and the Russian cloud study should not be combined statistically because they use different samples and methods. Together, they point to a practical sequence: start with a measurable workflow, then choose the long-term infrastructure.
A four-week pilot plan
- Week 1: select one process, record baseline metrics, and collect representative examples.
- Week 2: connect the data and apply AI to one stage while retaining human approval.
- Week 3: run the old and new workflows in parallel and measure errors, time, and cost per operation.
- Week 4: calculate the full impact, document exceptions, and decide whether to scale, redesign, or stop.
The deliverable should not merely be a chatbot. It should be a before-and-after table covering quality, time, cost, and risk. If the benefit is visible only in a presentation, the project is still an exploration rather than a deployment.
The management takeaway
Do not set a KPI to “deploy more AI.” Select a repeatable operation, assign an owner for the outcome, and measure the cost of correctly completed work. The value of the 228-company study lies in this qualification: operational application is more closely associated with economic results than simply expanding the toolset.