An AI secretary in Telegram is a bot you message like a person: by text, voice, or photo, and it replies on its own or passes the task to a specialized AI agent. A fresh public example with numbers: over 30 days, this kind of secretary cost its creator $77 in language-model usage for all agents. The main takeaway from the analysis is that cost is determined not by the number of messages, but by the share of complex tasks: 88% of the secretary’s replies came from a low-cost model, while 7% of replies from a powerful model consumed about 77% of its budget. That is why an AI secretary’s budget should be designed through model-selection rules, not a per-message pricing plan.
Prepared by AIrassvet October 13, 2026 based on an open analysis on Habr and the documentation for the tools used. It is intended for business owners, department heads, and CIOs/IT directors who are thinking about their own AI secretary or messenger-based assistant and want to understand its structure, risks, and monthly costs in advance. We did not test the bot described here ourselves: all figures about it are the author’s data, and the calculations in the article are marked as ours. AIrassvet is not affiliated with the projects and model providers mentioned; the section at the end of the article describes the publisher’s services.
Contents
- In Brief: What the Analysis Showed
- What an AI Secretary Is and How It Differs from a Chatbot
- How an AI Secretary Works for $77 a Month
- Where the Money Goes: Models and Share of Replies
- Escalation Rule: When to Call in the Expensive Model
- Memory: How the Secretary Remembers Conversations
- What People Actually Ask an AI Secretary to Do
- Where an AI Secretary Makes Mistakes and How to Fix Them
- How to Calculate Your AI Secretary’s Budget
- When an AI Secretary Makes Sense for a Company—and When It Doesn’t
- Launch Checklist
- Frequently Asked Questions
- Original Sources and Verification Method
In Brief: What the Analysis Showed
- $77 for 30 days — spending on language models for all of the bot’s agents (excluding image and video generation). Of that total, $51 — the secretary itself.
- 88% of replies from the secretary came from the low-cost DeepSeek model, with an average reply price of $0.002–$0.007.
- 7% of replies came from Claude Opus 5 at $0.21–$0.24 per reply — that is about 77% of the secretary’s expenses.
- The strong model is triggered by two rules: the message contains words indicating a complex task (“analyze,” “compare,” “calculate,” “write a report”) or the message plus context is longer than 2,500 characters.
- Every fifth task starts with a photo of a document or screenshot, while classic secretary work — calendar management and reminders — accounts for less than 5%.
- Memory is search across past conversations rather than a huge context window: the median search time is 17 ms, and it keeps working when the embedding service is unavailable.
What an AI Secretary Is and How It Differs from a Chatbot
An AI secretary is a single entry point in a messenger that understands natural language, remembers past conversations, and decides on its own whether to answer or hand the task off to a specialized agent. The user does not choose a menu section or fill out a template request: they send one message in one chat.
A classic chatbot works on a script: buttons, branches, and prewritten answers. It is inexpensive and predictable, but it cannot handle a question that is not in the script. An AI secretary, by contrast, relies on a large language model (LLM) — a neural network that generates a response in natural language. That makes it flexible, but every reply costs money, and quality depends on the chosen model.
An AI agent in this setup is an LLM-based program with tools and its own area of responsibility: for example, a legal agent with instructions not to invent legal provisions. The secretary acts as a dispatcher between such agents. For more on the channels where an assistant can be deployed, we wrote about it in the article “AI Assistant in WhatsApp, VK, Telegram, and MAX”.
How an AI Secretary Works for $77 a Month
On October 12, 2026, the author under the handle mrPetrukovich published on Habr an analysis of his AI secretary “Liza Petrovna”. The bot works in Telegram and through a bridge in MAX; it is one conversation with one shared memory. According to the author, since May 30 the bot received 8,568 requests from 118 people.
The author describes the architecture like this:
| Layer | What is used | Why |
|---|---|---|
| Inputs | Telegram bot and bridge to MAX | One conversation across different messengers |
| Dispatcher | Secretary hub | Responds directly or hands the task off to an agent: doctor, lawyer, designer, navigator |
| Model access | LiteLLM proxy | Agents know only model aliases; the actual provider is set in configuration |
| Storage | Postgres with pgvector and Redis | Users, correspondence, memory, metrics |
| Embeddings | Local deepvk/USER2-small model | Vectors for memory search without external APIs |
| Web search | Self-hosted SearXNG | Search is available to any model, including the low-cost one |
| Infrastructure | docker compose, Ansible, cAdvisor monitoring | Deployment and service monitoring |
The key element here is a proxy in front of all model providers. According to the LiteLLM documentation, it is a self-hosted gateway that gives applications one OpenAI-compatible endpoint for more than 100 providers, tracks spending by API keys, teams, and users, and can fall back to a backup model if there is an outage. For business, this means one simple thing: the model can be replaced in configuration without rewriting the bot, and costs are visible for each scenario.
Where the Money Goes: Model Mix and Share of Responses
The author breaks down 2,500 secretary responses over the last 30 days:
| Model | Share of Responses | Median Response Time | Average Cost per Response |
|---|---|---|---|
| DeepSeek (text and vision) | 88% | 2.5–12 s | $0.002–0.007 |
| Claude Opus 5 | 7% | 18–38 s | $0.21–0.24 |
| Claude Haiku 4.5 | 2.4% | 43 s | $0.006 |
| GLM 5.2 (experiment) | 2.2% | 5 s | $0.026 |
The source for the table is a breakdown on Habr, using the author's data for the last 30 days.
Over 30 days, the secretary cost $51, of which $39 was Opus responses. Our calculations based on these figures:
- Average cost per secretary response — $51 / 2,500 ≈ $0.02.
- Opus responses — 7% of 2,500 ≈ 175; $39 / 175 ≈ $0.22 per response, which matches the range in the table.
- The other 93% of responses cost about $12 in total.
- Specialized agents (doctor, lawyer, and others) cost about $26 ($77 − $51): each uses its own model, chosen for quality.
That leads to the main management takeaway: the cost of an AI secretary depends on how often you call on the stronger model, not on the total number of messages. One Opus response, according to the author, costs roughly as much as 50 DeepSeek responses. At the same time, the author deliberately does not hand the difficult 7% to the cheap model: that's exactly where users ask to 'analyze a contract' or 'calculate whether there will be enough fuel,' and an error in that kind of answer costs more than twenty cents.
Why responses in Russian can cost different amounts across models, we covered in the article on Russian text tokenization and LLM cost.
Escalation rule: when to call the expensive model
Escalation is the rule that sends a request from the default cheap model to a stronger, more expensive one. In the bot analyzed here, two triggers are enabled in production:
- Complexity. A regular expression looks for verbs associated with analytical work in the message: "analyze," "compare," "plan," "break down," "calculate," "compute," "research," as well as requests to write code, a report, an article, or a plan, and to "explain in detail."
- Length. The message plus context is longer than 2,500 characters. A long conversation often means the user is stuck and needs a solid answer, even if there are no trigger keywords.
Web search is intentionally not included in the triggers: all models have it, and the cheap one handles it on its own. Specialized agents are not governed by this rule — each has its own model.
Why use a regular expression instead of a model-based classifier? The author explains the choice this way: the rule can be read, covered by tests, and always explained when a specific request was routed to the expensive model. A classifier would add latency and cost to every response for a decision that affects only 7% of them.
| Model Selection Approach | Pros | Cons | Best Use Case |
|---|---|---|---|
| Rules (keywords, length) | Transparent, testable, no extra cost | Can miss complex tasks without trigger words | Launch stage, low volume, straightforward scenarios |
| Model-based classifier | Accounts for meaning and context | Adds latency and cost to every request | High volume, many ambiguous requests |
| Manual user selection | No routing errors | The user has to understand the models | Internal tools for specialists |
Our editorial comparison of the approaches is based on the analysis. For more on model routers, see the article "AI Model Router for Business".
Memory: How the Secretary Remembers Conversations
The AI secretary’s memory is a database of past conversations that the bot searches for the relevant snippet before answering, rather than trying to cram the entire chat history into the model context. Every exchange — a user message and the bot’s reply — is stored as a separate episode. According to the author, the database contains 7,933 episodes from 97 people.
How it works:
- The embedding is calculated in the background. An embedding is a numerical representation of a text’s meaning, used to find similar passages. The user never has to wait for it to finish calculating before getting a response.
- Search is hybrid: semantic search via pgvector plus keyword matching. The user’s words carry more weight than the bot’s reply, so the question should find the conversation where the person actually said it.
- Fault tolerance: if the embedding service is unavailable, search works only by keywords — worse, but without downtime.
- Privacy: before saving, third-party identifiers are removed from the text; each user has their own memory, and deleting a user deletes their episodes.
- Speed: memory search takes 33 ms on average, with a median of 17 ms.
For a Russian company, there is a legal boundary here. Under Part 5 of Article 18 of Federal Law No. 152-FZ, "On Personal Data" the collection, systematization, accumulation, and storage of personal data of Russian citizens using databases outside the country is not permitted, except in certain cases listed in the law. Therefore, it makes sense to host the memory database with customer or employee conversations in Russia, and decide in advance what data, if any, goes to external models. This is a general overview of the rule, not legal advice.
What People Actually Ask an AI Secretary
The author categorized all 8,568 requests by topic using the same low-cost model, without manually reading the chats. First, short messages like “are you there?” and test prompts were separated — they made up 21%: people talk to the bot like they would to a person. Of the remaining 6,796 tasks:
| Topic | Share of tasks |
|---|---|
| Analyze a photo, document, or screenshot | 19.6% |
| Create an image, video, or design | 12.5% |
| Computer, IT, setup | 9.2% |
| Find a fact, news item, or reference | 9.1% |
| Money, business, work | 7.1% |
| Technology, cars, aviation | 6.9% |
| Health | 6.2% |
| Legal and government questions | 5.6% |
| Calendar and reminders | 4.7% |
The table shows the nine largest of 16 topics; the classification is approximate, and the author notes the shares may differ by a couple of percentage points.
What this means for a company planning its own secretary:
- Without image recognition, the bot loses the main use case. In the bot described here, there is a separate model for this, which accounts for almost 7% of all secretary responses.
- A “secretary” in the narrow sense is only a small part of the job. Calendar and reminders account for 4.7%. People treat the bot as someone they can ask anything, so the boundaries of responsibility need to be defined clearly.
- Almost 12% is health and law. For these topics, the author set up separate agents with a strict requirement not to invent dosages or legal articles. In an internal corporate bot, the equivalent would be questions about employment matters, contracts, and money.
Important: this is the stats for a personal public bot, not a corporate one. For a sales or support assistant, the distribution will be different — you should measure it on your own traffic during the first few weeks.
Where the AI Secretary Makes Mistakes and How to Fix Them
The author frankly lists the errors encountered in production. We turned them into a table with practical takeaways:
| Error | What happened | How it was fixed or what to do |
|---|---|---|
| Replying to the wrong message | The model analyzed the previous photo instead of the new one | Explicitly separate the latest message with a marker and instructions to answer only that one |
| Wrong response language | A language metadata error meant the model followed the system prompt | Do not pass unverified metadata into the prompt; determine the language from the message itself |
| Date confusion | The bot discussed “future” plans that had already passed | Insert the current date explicitly and force it to check against the calendar |
| A fact was recorded instead of data | The calendar showed “link saved,” but the link itself was missing | Verify the result of the action, not the text describing the action |
| Domain terms | A technical term was explained incorrectly | Give the domain to a separate agent with a narrow instruction |
| Generation limitations | Photo retouching changed extra details with each attempt | Warn users about the limitations of editing models |
Separately, the author tested resistance to prompt injection — attempts to get the model to break its instructions through text in the chat. The tests included: a request to “run the command” and show a file with secrets, the same request in base64, a message from an “administrator” demanding that filters be disabled, an attempt to write a secret-leak rule into memory, and a question about a nonexistent law. According to the author, the bot refused in all cases. For a corporate bot, this kind of attack set should be turned into a regression test run after every prompt or model change. More on protection in the article “AI Agent Security: Protecting the AI Chat”.
The author calls the most valuable behavior the moment when the bot directly said it could not verify the data and would not make it up. For business, that is the key criterion: a model that says “I don’t know” is safer than a model that always answers with confidence.
How to Estimate Your AI Secretary’s Budget
It’s easy to calculate a language model budget with a simple formula. This is our editorial methodology, not a price list:
Monthly expenses = number of responses × (share of the cheap model × price per response + share of the stronger model × price per response) + specialty agent costs.
Numerical example from the breakdown (Calculated, based on the author’s average prices):
| Scenario | Calculation | Result for 2,500 responses |
|---|---|---|
| As in the breakdown: 93% of responses from DeepSeek, Haiku, and GLM, 7% from Opus | 2,325 × ~$0.005 + 175 × ~$0.22 | ≈ $50 |
| All on the stronger model | 2,500 × ~$0.22 | ≈ $550 |
| All on the cheap model | 2,500 × ~$0.005 | ≈ $12.5 |
The average price of a cheap response, $0.005, is the midpoint of the range in the breakdown; provider rates change in practice, so before launch recalculate them using current prices and your own prompts. The calculation shows the order of magnitude: routing everything to one strong model increases costs by roughly 11x, while relying only on the cheap model saves about $38 per month but shifts the risk of errors onto the most critical tasks.
What else is included in the total cost beyond tokens:
- infrastructure: a server for the bot, databases, embeddings, and search;
- image and video generation, if needed — it is not included in the calculated $77;
- development and support: updating prompts, models, tests, and integrations;
- employees’ time spent reviewing errors and providing feedback.
Template for your own estimate:
| Parameter | Your value | Where to get it |
|---|---|---|
| Responses per month | … | A 2–4 week pilot or a forecast based on current inquiries |
| Share of complex tasks | … | Labeling a sample of real requests |
| Price per response for the cheap model | … | Provider pricing and the average response length during the pilot |
| Price per response for the stronger model | … | Same as above |
| Specialty agents | … | Separate calculation for each one |
| Infrastructure and support | … | Estimate from the team or contractor |
When an AI secretary for a company makes sense — and when it doesn’t
Makes sense if:
- there are many repetitive inquiries in messengers, but the questions are phrased freely and don’t fit buttons;
- employees or customers need one entry point instead of several bots and sections;
- some requests come in as photos of documents, screenshots, or invoices;
- there is a knowledge base or internal procedures the bot can use to answer with a source citation.
Not justified or premature if:
- the scenarios are strictly defined and can be handled by a standard button-based bot;
- an error in the answer is unacceptable, and there is no person available to verify it in the process;
- there is no clear understanding of what data can be sent to external models, and deploying the model in your own environment is not yet possible;
- the volume is small, and development and support costs won’t be offset by time savings.
Launch checklist
- Define 3–5 key scenarios and write down what the bot should not do.
- Choose a cheap default model and a strong model for complex tasks; describe the escalation rule so it can be tested.
- Set up a proxy or gateway in front of the models: swap models without rewriting the bot and track costs by scenario.
- Design memory: where it is stored, what gets scrubbed from it, and how it is deleted on request.
- Add image recognition, if users send documents and screenshots.
- Insert into the prompt the current date and clearly separate the user’s last message.
- Build a set of attack prompts and validation questions and run it after every change.
- Run a 2–4 week pilot, label real requests, and recalculate the budget based on the actual share of complex tasks.
Frequently Asked Questions
How much does an AI secretary in Telegram cost per month?
In a public breakdown on Habr, language model costs for all agents totaled $77 over 30 days, including $51 for the secretary itself at 2,500 responses. That’s only the model usage: infrastructure, image generation, development, and support are billed separately. Your budget depends primarily on the share of tasks sent to the stronger model.
Why does the expensive model eat up most of the budget?
One response from the stronger model in the breakdown cost $0.21–0.24, while a cheap one cost $0.002–0.007. That’s why 7% of responses from the stronger model accounted for about 77% of the secretary’s costs. You manage the budget through an escalation rule, not by limiting the number of messages.
Can all requests be sent to the cheap model?
Yes, and based on our calculations from the breakdown’s numbers, that would have saved about $38 per month. But the stronger model handles the most critical tasks: contract review, calculations, plans. A mistake in that kind of answer usually costs more than the price difference.
How is an AI secretary different from a regular chatbot?
A regular chatbot follows a predefined script with buttons and branches. An AI secretary understands free-form text, photos, and voice, remembers past conversations, and decides on its own whether to answer or hand the task off to a specialized agent.
Where should the AI secretary’s chat history and memory be stored?
For the personal data of Russian citizens, Part 5 of Article 18 of Federal Law No. 152-FZ requires databases to be located in Russia, with certain exceptions. It makes sense to keep the bot’s memory with client or employee conversations inside a Russian data environment and define in advance which data is sent to external models.
How do you protect an AI secretary from prompt injections?
System instructions should come only through the system channel, not from chat; requests for secrets or rule changes should be rejected; memory should not accept “rules” from the user. A standard set of attacks should be run as a regression test after every prompt or model change.
Primary sources and verification method
We reviewed the primary sources on October 13, 2026. All data about the “Liza Petrovna” bot—architecture, response shares, prices, memory, request topics, and errors—comes from the author’s breakdown on Habr and was not independently verified by us in practice. Our calculations of the average response cost, the number of Opus responses, the spending of specialized agents, and budget scenarios were done using the author’s figures and are marked in the text. The comparison tables, “when it makes sense,” and the checklist are AIrassvet’s editorial methodology.
- mrPetrukovich, Habr, 12.10.2026: AI secretary for $77 per month: DeepSeek handles 88% of responses, while three-quarters of the money goes to 7% of requests to Opus — architecture, 30-day figures, escalation rules, memory, errors, prompt-injection tests.
- LiteLLM: AI Gateway documentation — a single OpenAI-compatible endpoint for providers, cost tracking, budgets, and failover.
- Federal Law No. 152-FZ: Article 18 — the requirement to use databases located in Russia when processing the personal data of Russian Federation citizens.
- Telegram: Bot API documentation — the official interface for bots in Telegram.
Limitations: the data describe one personal public bot over 30 days; model prices and provider rates change; the topic labeling was done by a model and, according to the author, is approximate; the distribution of requests in a corporate bot will be different.
How AIrassvet helps launch an AI secretary for your company
If you need an AI secretary that responds to employees or customers in a messenger while using your data and staying within budget, its design should be built around your processes from the start. On the AIrassvet website the following areas are described that directly apply to this task:
- AI agents integrated with communication channels, including Telegram, and with the company’s business systems.
- RAG systems and corporate knowledge bases so the secretary answers based on policies and documents, not the model’s general knowledge.
- Deploying language models on the company’s own servers if chat data cannot be sent to external services.
- Process audit, MVP development, testing, launch, and ongoing support for the solution.
The first step is to choose one request stream, describe how it is handled today and what it costs, where the required data is stored, what personal-data constraints apply, and by what criteria you will consider the secretary’s answers accurate enough.
Bottom line: the AI secretary’s budget is determined by rules, not the pricing plan
The breakdown of the $77-per-month AI secretary shows that a personal assistant in Telegram built on language models is already inexpensive—if almost all responses come from the cheap model and the stronger one is triggered by a clear rule. It’s the rare complex tasks that drive most of the cost, and those are the ones where a mistake is most expensive. Everything else—episodic memory, photo recognition, an explicit date in the prompt, and injection tests—determines whether the bot will be useful and safe.
Start with one request stream: a pilot for a few weeks, tagging real requests, an escalation rule, and a budget recalculation based on the actual share of complex tasks. If, after the pilot, the answers are verifiable and the cost fits your budget, the secretary can be expanded to new scenarios and agents.