An AI service subscription rarely exceeds 10–15% of the real cost of a project. The rest is data, integration, oversight, and reworking processes. Those who calculate the effect of automation based on the vendor’s invoice should recalculate.
Money is flowing into the market but the effect lags behind
Russian businesses invested 257 billion rubles in artificial intelligence in 2025, with average annual expense growth of about 50 billion rubles. The technology is used by 71% of large companies, but only 26% of those who allocate AI in their budgets have a formalized implementation strategy.
In June 2026, participants at AI Sochi Camp unambiguously formulated the criterion by which a project should be evaluated: implementation has taken place only if either new money has appeared or costs have been reduced. Everything else is an experiment at the budget’s expense.
The gap between expenses and results can be measured. By the end of 2025, only 37% of AI initiatives had delivered the expected business value, and 29% reach industrial scale.
The MIT NANDA report titled The GenAI Divide, based on a sample of 300 implementations, shows a harsher figure: 95% of generative AI pilots produce no measurable effect on profit. More than 45% of projects with AI agents will likely be shut down before the end of 2027 – due to rising costs, blurred value, and weak risk controls.
What doesn’t make it into the calculation
A typical mistake in a business case is counting the license and tokens. The real cost structure is broader:
- Data. Labeling, cleaning, building storage and retrieval pipelines. Without this, the model responds based on data that cannot be trusted.
- Integration. Connecting to CRM, ERP, billing, and document management systems. This is development, not configuration.
- Infrastructure. The cost of deploying leading large language models in Russia in the summer of 2026 averaged 38.8 million rubles, versus 13.7 million rubles a year earlier.
- Human oversight. Checking output, analyzing errors, accountability for decisions. This does not disappear after the AI launch.
- Fine-tuning and monitoring. Models and prompts degrade as processes change.
- Risk management. Leaks, personal data, hallucinations in customer communications, dependence on a single vendor.
A joint study by Cloud.ru and the HSE Graduate School of Business provides an estimate of the investment required: launching AI agent-based projects at a large company costs between 10 million and 50 million rubles. Capital expenditures (CAPEX) include infrastructure, platforms, and integration, while operating expenditures (OPEX) cover maintenance, retraining, monitoring, cloud services, and personnel.
Payback periods vary considerably. Simple applications built on off-the-shelf solutions can pay for themselves within 12 months, while projects requiring customization typically take 12-24 months. Developing proprietary platforms and models takes more than two years to recoup the investment, while standalone infrastructure projects may require five to seven years.
Why costs rise instead of fall
The explanation is straightforward: automation does not necessarily replace an existing process. Instead, it adds another layer. If the original workflow remains unchanged, the company ends up paying twice: once for the employees and again for the AI model.
Common reasons for cost overruns include:
- Pilot projects are never integrated into day-to-day operations and remain demonstrations;
- No baseline performance metrics are established, making it impossible to measure the impact of automation;
- Companies choose to automate the most visible processes rather than those that generate the highest costs;
- Costs fluctuate: inference expenses (the cost of running a trained neural network) increase with usage, while savings remain fixed;
- Responsibility for model errors is not clearly assigned, and the cost of implementing safeguards eats into potential savings.
MIT researchers explicitly point out that the problem lies not in the quality of AI models but in the design of the processes into which they are integrated. The same principle was discussed at the AI Sochi Camp: the projected 30-60% reduction in operating costs is achievable when workflows are redesigned around AI agents, not when an AI model is simply added to an existing process.
In the case studies presented at the event, cost savings only emerged after the cycle was restructured: the cost of producing an editorial piece fell from 15,000 to approximately 2,500 rubles, while a single employee compiled a BI dataset containing 133,000 sales records in one day – a task that would previously have taken a team months to complete. These results were made possible by having data readily available, rather than by the neural network itself. Extrapolating the savings achieved in a single digitized process to build a business case for company-wide implementation is a recipe for cost overruns.
How to verify that the technology delivers savings
Before expanding a contract, test the technology’s cost-effectiveness in practice:
- Calculate total cost of ownership (TCO), not just the subscription fee. Include licenses, inference costs, data, integration, training, monitoring and support over a 24–36-month period.
- Establish baseline metrics. Record the process’s operating costs and duration, rework rates, and the number of full-time equivalents (FTEs) assigned to it.
- Measure savings in monetary terms per unit of output, not simply in staff hours saved. Time savings will not show up on the profit and loss statement unless they translate into lower staffing costs or higher output.
- Assess variable costs. In particular, determine how inference costs would change if the workload doubled.
- Quantify the cost of errors. Establish the financial impact of a single incorrect output in a customer-facing or regulatory workflow.
- Set a clear exit threshold. Define in advance the conditions under which the pilot will be terminated.
Risks and uncertainties
At the Sochi Camp, the window for gaining a competitive advantage through implementation was estimated at 12–18 months. Late adopters would eventually gain access to the same technology, but at a higher cost. Even within this window, however, three factors could begin to undermine the economics of these projects as early as 2027.
The first is computing costs. Rising hardware prices and growing volumes of model queries could erode savings and make them increasingly difficult to sustain.
The second is regulation. Requirements governing data processing and liability for automated decisions increase compliance costs.
The third is personnel costs. Competition for engineers and rising payroll expenses could simply shift costs from one budget line to another.
Conclusions
Spending on AI will continue to grow faster than its measurable business impact, as the market remains in a phase where investment is outpacing process transformation. In 2027, the divide between winners and losers is expected to widen, with many companies to abandon projects that fail to deliver a return, while the benefits will accrue primarily to those that have calculated their total cost of ownership and redesigned their workflows, rather than simply purchased a subscription.
By Andrey Kreer, Founder and CEO of the insurance aggregator INFULL