2people IT CTO explained where AI in business really pays off and where companies lose money

Companies are increasingly asking to «implement AI», but economic impact does not appear in every case. 2people IT CTO Ilya Trokhin breaks down which tasks truly pay off, and where neural networks still remain an expensive experiment with no clear return.

Why requests fail

According to the author’s observations, most implementation requests begin not with a problem, but with a desire to «do what everyone else is doing». At the very first review, it turns out that the company has no formalized process, no clear metrics, no proper data structure, and not even a single way of performing the task within the team. In that situation, AI does not become a magical layer on top: it does not fix operational disorder, it only accelerates its replication and increases the cost of mistakes.

AI does not fix chaos. It scales it.

Hence the main filter before any pilot: first, you need to calculate exactly where the business is losing money, time, or quality. If employees act differently every time, the task flow is small, and the result cannot be measured, there is almost nothing to automate there. In that scenario, the project turns either into an image-building initiative or into an expensive hypothesis test with no chance of reaching payback quickly. The author clearly leads to the idea that order in processes must come first, and only then a neural layer on top.

Where the effect is visible quickly

The most predictable cases are those with a large volume of repetitive operations, clear input, and a noticeable share of manual work. It is precisely in these areas that AI delivers measurable results fastest: it cuts processing time, reduces the number of errors, and removes routine load from the team. This is not about fully replacing specialists, but about scenarios where the model takes over the initial or draft work, while a person steps in for exceptions and quality control.

• triage of incoming requests and emails
• extracting data from contracts, forms, and other documents
• answering standard support questions and routing requests
• initial processing of applications and HR requests
• preparing draft documentation, test cases, and prototypes in IT

The article separately analyzes a document-processing case: previously, an operator manually checked a passport image and entered the fields into the system, which took time and led to input errors. After OCR and automatic data extraction were introduced, the system began reading the full name, document number, and date of birth on its own, while a person stepped in only for disputed cases.

A similar logic works in support as well: the assistant closes standard requests, speeds up the first response, and passes along only non-standard cases where an employee is truly needed.

Another effective area is internal IT routine. The author writes that the team uses AI as an accelerator for preparing draft documentation, generating test cases, assisting with code review, and rapid prototyping. Such tools still cannot fully replace a developer, but they reduce hours spent on repetitive tasks and help launch MVPs faster. For commercial projects, this means shorter timelines and less pressure on the budget, especially in the early stages.

How to calculate payback

The author suggests not starting with choosing a model or vendor, but with a short operational audit and an assessment of whether the process is ready for implementation. First, you need to understand the cost of the current process; then the volume of operations per month; after that, identify the bottleneck and check whether the company has enough data for automation. Only when these four points are confirmed does AI become not a fashionable add-on, but a tool with clear business logic and predictable impact.

Where the volume is small, the process is unstable, or the company is trying to replace a person completely, implementation often stalls. Economic impact appears only in cases where the solution either reduces specific costs, shortens the cycle, or lowers the number of errors. If this cannot be expressed in money, timelines, or SLA, the project was most likely launched too early. Under such conditions, conventional automation without neural networks sometimes brings more value because it costs less, is implemented faster, and does not require complex data preparation.

What this means

For businesses, this is yet another reminder that AI works better not as a showcase of innovation, but as a practical tool for large-scale, repeatable, and costly processes. The fastest wins today lie in documents, support, internal assistants, and the routine tasks of IT teams, not in attempts to «automate everything at once».