What it is
A program that understands and writes text: ChatGPT, Gemini, Claude, DeepSeek. On its own it knows nothing about your site until it reads it through search or by link.
How we use it
We use them in bots through APIs: voice transcription, parsing bank statements and receipts, a site chat consultant with facts from the price list. The model answers only from what is in the prompt so it does not invent prices.
Where a language model helps a business
- Voice messages and call recordings need to become text you can search and quote.
- Receipt photos, complex phrases and cryptic bank statement lines need to be categorised without manual entry.
- Staff ask support the same questions, and some of them can be closed before a ticket is created.
- A user wants to ask about their own data in plain words: "what do I spend the most on?".
How we use it
- Voice transcription. The audio and an instruction go to the multimodal Gemini model in a single request, with no separate speech recognition step. Three modes: verbatim, cleaned up, summary. Before sending, FFmpeg checks the recording and compresses a heavy file, and the audio is deleted right after processing.
- Finance assistant. The model handles complex text, voice and receipt photos, and overnight it categorises unknown merchants in one request. Answers to questions about money are built from the bot's own data, and the bot gives no investment advice.
- IT help desk. An optional Gemini assistant as the first line of support, off by default.
We build AI bots under Telegram and VK bots.
Common problems
- A key runs out of quota. The free quota is limited, so keys work as a pool: an exhausted key goes on cooldown (60 minutes by default), the request goes through the next key, and the admin gets a notification.
- The model is renamed or withdrawn. On a 404 the bot switches to the backup model by itself. Keys and the model order are set in
.env, with no code changes. - The model makes things up. If a recording contains no speech, the bot says so instead of inventing text. Answers about money come only from the bot's own data.
- Data leaves for someone else's cloud. A receipt photo and the message text go to Google when a complex parse is needed. You should know that in advance; without keys only the built-in rules work.
- Filters and quality. Gemini safety filters may reject a recording on a sensitive topic, and noise or several voices at once lower quality.
- Repeats cost money. The same recording in the same mode comes back from the cache instantly and without an API call; the cache lives for 30 days.
When you do not need it
If simple rules solve the task (keywords, known merchants, fixed buttons), you do not need a model: rules are faster, cheaper and never hallucinate. In the finance assistant a category is first looked up in learned rules and keywords, and only the unclear cases go to the model. And if the data must not go to an external service, a cloud model does not fit.