Artificial intelligence may become the foundation of the next technological cycle, while some investments in it may turn out to be unprofitable. There is no contradiction here. The usefulness of a technology, the profitability of its provider and the return on a particular implementation are three different things.
Confusing them is what makes the debate about an AI bubble of little use to a company executive. Rising technology stock prices do not guarantee the success of a corporate project. Falling prices do not remove the need to process documents faster, find information, serve customers and reduce errors.
In my view, Ukrainian businesses should build their assessment of AI around results they can verify within their own organisations. Stock market expectations change quickly. The cost of an operation, the duration of a process and the quality of the work provide a much more reliable basis for a decision.
AI infrastructure: investment and supplier risk
There are reasons for concern about AI infrastructure. Data centres require substantial upfront investment, energy, cooling and regular hardware upgrades. A company must recover these costs by selling services. If demand, prices or capacity utilisation fall short of expectations, the economics of the project deteriorate.
Even financial indicators, however, need careful reading. For example, in the third quarter of 2026, Meta generated $32.79 billion in operating cash flow, spent $34.91 billion on capital expenditure including principal payments on finance leases, and reported negative free cash flow of $2.12 billion. These are company-wide figures; they do not isolate spending on AI. The company generated substantial cash but invested even more. This raises a question about future investment returns; on its own, it does not establish insolvency. Meta: third-quarter 2026 results.
This has practical implications for corporate customers. When choosing an AI supplier, businesses need to assess its ability to maintain the service, the predictability of its pricing and the possibility of migrating the solution. Making a critical process dependent on a single provider creates a risk even when its technology is currently the best.
Cheap tokens and the full cost of implementation
The rapid decline in inference costs is a separate issue. According to the Stanford AI Index, the cost of using a model that achieved GPT-3.5-level performance on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 in October 2024. That is a price reduction of more than 280-fold for a specified level of quality on a particular test. Stanford AI Index 2025: inference costs, page 41.
For the owner of computing infrastructure, this trend may put pressure on prices. For a software developer, it may make previously unaffordable applications viable. The consequences can be opposite for different market participants.
Cheap tokens do not automatically mean cheap automation. A corporate solution also requires spending on data preparation, integration with existing systems, access controls, quality checks, employee training and support. Model errors have a cost too: they need to be identified, corrected and accounted for in subsequent work.
That is why I find it useful to calculate the cost of a correctly completed business operation. For a customer enquiry, this means all costs up to the point when the customer receives the correct answer. For a document, it means the costs of entering it correctly into the system, checking it and handing it over to the responsible employee. This calculation reveals what the price of a single model query leaves out.
How to measure a pilot's results
Imagine a system that recognises an incoming document, identifies its type and proposes an approval workflow. It may perform convincingly in a demonstration. Daily operations will introduce poor scans, conflicting details, unusual wording and documents with restricted access. If an employee has to check every field and determine the workflow again, the expected savings may disappear.
Before launching a pilot, it is therefore necessary to record the baseline: how long an operation takes, how much it costs and how often errors occur. After implementation, the measurements should be repeated on a comparable set of tasks. Cases the system hands over to a person, and the time spent correcting its output, should be accounted for separately.
Even verified time savings still need to be converted into an economic benefit. An hour of an employee's time freed up does not automatically become a cost reduction. It may allow more enquiries to be handled, shorter queues, faster service or work that previously lacked resources. Management must define that outcome in advance.
Where Ukrainian developers create value
For Ukrainian developers, this creates an opportunity to deliver value through domain knowledge. Connecting a language model is only part of the solution. Customers also need accurate data, integration with their business systems, clear access rights, exception handling and accountability for support.
In contract approval, for example, value lies in the entire process: from receiving the document and identifying its version to verifying authority and recording the decision. High-quality text analysis helps at individual stages. Overall reliability comes from the system architecture and the way work is organised.
This also affects the approach to the team. Justifying a project solely through future headcount reductions is risky: some work may shift to verification, exception handling and quality control. In its research on generative AI, the ILO emphasises the difference between automating individual tasks and replacing occupations; job transformation is the more likely outcome for most exposed occupations. ILO: how generative AI may affect different occupations.
I support a phased approach: a limited process, a defined budget, measurable quality criteria and a decision on scaling after the results have been assessed. This model allows a business to change suppliers, reconsider its architecture or stop an unsuccessful initiative before the entire budget is lost.
The investment excitement around AI may subside. Some companies may overestimate demand, and some products may fail to find customers. For Ukraine's IT industry, the ability to demonstrate the value of the work delivered will remain a lasting foundation.
Companies implementing AI must be ready to show customers how much a completed operation costs, how its quality is verified and who is responsible for an error. Those are the answers on which a long-term business can be built.
Related work at the Association: Software Ukraine AI Commission.