AI at Finam: Surviving the Move from GPT-3.5 to a Corporate Pipeline

Many people still sincerely believe that bringing artificial intelligence into a large company looks like buying a ChatGPT Plus subscription for every department. In reality, Finam found that the enthusiasm of lone individuals quickly crashes into harsh corporate reality. When your Flutter prototype runs on one developer's machine, it feels like magic and a technological breakthrough. When the same technology has to be handed to a thousand employees of a financial organization, the magic turns into an endless string of problems with security, query costs, and answer quality. Finam's story is instructive precisely because the team was not afraid to go from a "toy" to an industrial solution, stepping on every possible rake along the way.

It all began with a classic experiment on GPT-3.5. At the time, it seemed enough to simply give people access to the API and productivity would soar.

But the financial sector is not a sandbox. The very first serious review showed that uncontrolled use of external neural networks carries the risk of leaking confidential data, along with unpredictable costs. You cannot simply send client data to the OpenAI cloud and hope for the best.

The team had to urgently build an internal platform to act as a "layer" between the employee and the model. That layer took over access control, logging of every request, and strict token limits, so the AI budget would not run out in the first week of the month.

The next stage was the fight against "hallucinations." Neural networks are excellent storytellers, but in finance a lie costs too much. To make the AI tell the truth, Finam adopted RAG (Retrieval-Augmented Generation).

Now the model does not simply invent an answer out of thin air: it first searches the company's database of internal regulations and documents, and only then forms an answer based on the facts it found. This turned an abstract chatbot into an expert who knows the fine points of the company's internal processes. In parallel, experiments with local models began.

Running Llama or other open-source solutions on the company's own hardware answered security requirements: the most sensitive data must not leave the company's internal perimeter at all.

Scaling to 1000+ people forced a complete rethink of the architecture. It turned out that keeping an AI service stable under heavy load is an engineering discipline of its own. The team had to build complex monitoring systems that track not only server availability but also answer quality in real time. Finam concluded that technology is only 20% of success. The other 80% is painstaking work on infrastructure, teaching employees to phrase requests properly, and constant adaptation to changing regulatory requirements. The irony is that the smarter AI gets, the more qualified engineers it takes to support it.

Today Finam's platform is not just a chat but a full hub that hosts external APIs, local models, and specialized agents for different departments. It is a path of compromises between response speed, the price of a single token, and data security. The company has shown clearly that the era of "just chatbots" is over, giving way to the era of corporate AI ecosystems. Those who try to simply "bolt on an API" today will face the same scaling challenges that Finam has already overcome. The only question is whether others will have the patience to turn a fashionable trend into a working tool.

Key takeaway: Enterprise AI adoption is 80% infrastructure building and only 20% choosing the model itself. Are you ready to pay double for the security of your data in the form of engineering hours?