The AI assistant on Telegram in the Hanse Studio package costs from 3,000 PLN setup plus 800 PLN monthly retainer. It is a productized service: a personalized bot with the client’s brand voice, integrations with Gmail, Calendar and one CRM, 4 to 8 custom commands, monthly improvement included in the retainer. In this article we describe exactly what the client receives, how it works under the hood, which 4 use cases it most often automates, and why voice extraction from the client’s emails is a critical element.
What the client gets in the package
The AI Assistant on Telegram package is not generic ChatGPT with plugins. It is a dedicated bot, built for a specific company, with a persona and integrations. The client receives 5 components in the 3,000 PLN setup.
- Telegram bot with personalized persona. Voice extraction from 30 to 50 client emails before the build. Result: JSON voice profile (tone, greetings, sign-offs, vocabulary, sentence rhythm), embedded in the bot’s system prompt. The bot sounds like the company owner, not like ChatGPT.
- Integrations with Gmail, Calendar and one CRM. Standard options: Pipedrive, HubSpot or Salesforce. Each integration has its own MCP server (Model Context Protocol) exposing tools: search_emails, draft_reply, list_events, create_event, update_deal. The client sees these tools as commands or uses them naturally in chat.
- 4 to 8 custom commands per client. Selection from a library of 30+ standard commands (summarize inbox over the last 24h, send an offer to X, schedule a meeting with Y, daily standup, weekly report, monthly closure) plus the option to design custom commands for the client’s specific workflow.
- Monthly improvement in the retainer. Up to 2 hours per month for tweaks, new commands, bug fixes, system prompt adjustment. After each month a short review call with the client, list of changes, deployment in the following week.
- Uptime monitoring and audit log. The bot runs 24/7 on a Mac Mini or Hetzner. Health check every 5 minutes, alert to Hanse Studio on Telegram if the bot is down >10 minutes. Audit log of every action retained for 6 months (GDPR compliance).
The package does not include the client’s production hosting (Mac Mini or VPS purchased separately), API licenses (Anthropic claude.ai approximately 50 to 200 PLN per month depending on usage), or integrations with non-standard CRMs (beyond Pipedrive, HubSpot, Salesforce a custom integration costs an additional 1,500 to 3,000 PLN). The scope boundary is clear in the pre-signature offer.
Pricing: from 3,000 setup + 800/mo
Setup from 3,000 PLN is a build in 2 to 3 weeks: kickoff call (1 hour), voice extraction from 30 to 50 client emails (Hanse Studio anonymizes plus analyzes, the client does not need to sort), integrations setup (Gmail OAuth, Calendar, CRM API key), command definition together with the client (selection from library plus 1 to 2 custom), testing in sandbox with the client, deployment to production. Invoice issued in two installments: 50 percent advance after signing the contract, 50 percent after go-live.
Retainer 800 PLN monthly after setup covers bot hosting (Hanse Studio shared Mac Mini or client’s own VPS, zero or minimal cost), uptime monitoring with alerts, ad-hoc tweaks up to 2 hours per month, monthly performance review (a 15-minute call once a month plus a PDF with metrics). Retainer invoiced monthly in advance, 30-day notice period.
Why a retainer and not a one-time fix-price? The bot needs continuous improvement. In the first 2 to 3 months the client notices edge cases (bot misinterpreted, missing command) that need addressing. Without a retainer the bot would become neglected after 6 months. A retainer of 800 PLN/mo is low enough to fit an SMB budget but covers the real cost of maintaining quality.
For comparison: ChatGPT Team costs 25 USD/user/mo (100 PLN at 4 users equals 400 PLN), generic AI assistant SaaS such as Synthflow or Voiceflow starts at 100 to 300 USD/mo (400 to 1,200 PLN). Hanse Studio sits in the middle of this scale with two key differences: a dedicated client persona and monthly improvement included.
Architecture: how it works under the hood
The technical stack is deliberately opinionated. We do not experiment with every new framework in client production. The architecture is proven across 3+ Hanse Studio deployments since 2025.
- Claude Agent SDK as orchestrator. Hanse Studio recommends Claude from Anthropic due to reasoning quality, predictable tool use and API stability. Alternatives (OpenAI Assistants API, LangChain) were tested, but Agent SDK gives the cleanest production code for Hanse Studio packages.
- Telegram Bot API. The bot receives messages directly from Telegram. Communication secured with TLS, restricted to Telegram IPs, rate limit 30 requests/min per user.
- Modular integrations. Separate module per integration (Gmail, Calendar, CRM, Stripe). Claude selects the right tool from the list for each action. Adding a new integration does not require rewriting the existing ones.
- Lightweight per-user state store. Conversation history, command preferences, last 30 days of interactions. Lightweight, no separate database required. Daily backup to encrypted S3 or Hetzner Storage Box.
- System prompt with brand voice, persona and tools list. The system prompt is ~2000 tokens, contains: brand voice JSON (from voice extraction), persona description (who the bot is, what boundaries it has), available tools per integration, examples of typical interactions. Updated monthly in the retainer.
- Audit log of every action. Each action logged with timestamp, user ID, action name, parameters and result. GDPR-compliant 6-month retention. The client can request log export (right of access) or deletion (right to be forgotten).
The full architecture is technically described in the Claude Agent SDK documentation. For the client one sentence is enough: the bot is hosted in the EU (Hetzner Frankfurt by default), uses Claude from Anthropic, integrates via proven APIs.
4 use cases the bot automates
An assistant on Telegram works best in 4 typical scenarios. Most Hanse Studio clients use bots for a combination of 2 to 3 of them.
Use case 1: Email triage
Every 15 minutes the bot reviews the client’s Gmail inbox. It classifies incoming messages into 3 categories: urgent (client action needed within 24h), normal (to read within a week), spam/noise (newsletters, automated notifications). For the top 5 urgent it drafts a reply in the client’s voice (using the brand voice JSON). Drafts land in a queue on Telegram, the client one-tap accept/edit/reject. After acceptance, the bot sends from the client’s Gmail. Average saving for an SMB: 1.5 hours per day x 22 business days x 120 PLN/h = 3,960 PLN per month.
Use case 2: Calendar management
The client writes to the bot: “schedule Karol next week, 30 minutes, video”. The bot checks the client’s Calendar for free slots (matching working hours, 15-minute buffer after each meeting), proposes 3 time options, sends an email to Karol with a confirmation link. After Karol confirms, the bot creates an event in Calendar and sends a confirmation to both parties. Saving: no more chaotic “when works for you” exchanges across 5 emails. On average 30 to 45 minutes per week for a typical SMB with 5 external meetings per week.
Use case 3: CRM auto-update
After each call (client note on Telegram after the conversation) or email exchange (the bot detects a new thread), the client’s CRM auto-updates: adds an activity log to the deal, updates the stage if applicable (e.g. “after 2 demo meetings” automatically moves from Discovery to Proposal stage), tags based on keywords in emails. The client does not need to remember to enter into the CRM manually. Saving: 5 to 10 hours per month of sales ops time for an SMB with 30+ active deals.
Use case 4: Daily standup
Every morning at 7:30 the bot sends the client 1 message on Telegram with a summary: the 3 most important incoming items from the inbox (urgent triage), meetings for today from Calendar (times, agenda, contact), 2 to 3 open tasks with the highest priority (from CRM or Todoist), optionally market updates (the client defines whether they want them). The client opens Telegram, reads for 30 seconds, knows what is ahead. Saving: no more scrolling through 3 apps in the morning. On average 15 to 20 minutes per day x 22 days = 5 to 7 hours per month.
Persona: why voice extraction is critical
Generic AI sounds like AI. The client distrusts it, end customers (to whom the assistant sometimes writes on the owner’s behalf) sense it, the bot becomes ineffective. Voice extraction solves this problem at the system prompt level.
Process: the client sends 30 to 50 emails representative of different contexts (sales, customer service, internal communication, follow-ups). Hanse Studio anonymizes (we remove recipient PII), analyzes. Result: a JSON with structure: tone (professional, casual, technical), greetings (list of typical openings), sign-offs, vocabulary (characteristic phrases, industry jargon), sentence rhythm (short sentences or long ones, paragraphs 1-2 or 3-5 sentences).
The JSON voice profile is embedded in the bot’s system prompt. Every message the bot generates (email draft, response, notification) goes through this filter. Result: the bot sounds like the company owner, not like Claude/ChatGPT. End customers receiving replies from the bot do not know it is not the owner. This is not deception (the client accepts every draft), but elimination of the “AI uncanny valley” which is a common reason why bots are ignored.
Voice extraction is not a one-time event. In the retainer we can refresh quarterly (if the owner’s style evolves, add new contexts, e.g. a new product with new marketing language). The voice profile is the client’s property, exportable to other tools (e.g. as a brand voice in marketing automation outside Hanse Studio).
Security and GDPR compliance
The bot operates with client data (emails, contacts, deals), which often contain PII (Personally Identifiable Information). Compliance is non-negotiable. Hanse Studio implements 6 layers of security per deployment.
- User allowlist. The bot only receives messages from pre-approved Telegram user IDs. NEVER public access. An access attempt from another ID returns “Access denied” and logs the incident.
- DPA with Anthropic and Telegram. Hanse Studio has a standard DPA (Data Processing Agreement) with Anthropic (Claude API) and operates on Telegram Bot API in accordance with their ToS. The client receives a copy of the DPA as an attachment to the contract.
- Audit log 6 months. Every action (tool call, message sent, file accessed) logged in an append-only log. GDPR-compliant retention: 6 months default, the client can extend to 12 or 24 for regulated industries.
- Right to be forgotten. The client can request deletion of the entire bot state (conversation history, voice profile, audit logs) within 30 days. We implement a hard delete without backup retention.
- PII redaction in storage. Conversation history contains content (required for the bot), but the long-term audit log redacts phone numbers, NIP and PESEL before archiving.
- EU server option. Hetzner Frankfurt is the default for EU clients. Hanse Studio’s Mac Mini (Szczecin) is an option only for PL/EU clients who explicitly agree. Anthropic Claude API is hosted in the US but operates on data-in-transit only (no training on client data per Anthropic policy).
For clients in regulated industries (medicine, law, finance) we add additional compliance layers: ISO 27001 audit trail format, encrypted at rest for all state, 2FA on the bot admin. Additional cost: 1,000 to 3,000 PLN setup and 200 PLN monthly added to the retainer.
Onboarding flow: 2 weeks to live
The standard timeline from contract signature to go-live is 2 weeks. Longer for clients with exotic CRMs or compliance requirements. Shorter (1 week) for very simple scopes (only email triage, no CRM integration).
Week 1 (days 1 to 5): Kickoff call of 1 hour (Tuesday after signing), the client sends 30 to 50 emails on Thursday (paragraphs 5 and 6 explain voice extraction). In the background: Hanse Studio configures Gmail OAuth, Calendar API, CRM API. Voice extraction completed in 1 to 2 days, JSON profile sent to the client for review. The client approves the sound (“yes, that is exactly how I write”) or requests an adjustment.
Week 2 (days 6 to 10): Command definition on the previous Thursday (client selects 4 to 8 from the library of 30+ commands). Bot build Monday-Wednesday. Testing with the client in sandbox on Thursday (the client writes to the test bot, 5 to 10 typical questions, checks responses). Deployment to production on Friday afternoon, intensive monitoring through the weekend. Monday of week 3 official go-live, the client starts daily use.
The first 30 days post-go-live are active monitoring from Hanse Studio’s side. 1 to 2 times a week we check the audit log, identify edge cases (poor triage classification, awkward draft phrasing), adjust the system prompt. After 30 days the bot usually stabilizes and moves to standard retainer mode (monthly improvement).
FAQ
What if AI gives a wrong answer?
The bot has an escalation trigger: in case of low confidence (non-standard situations, queries outside scope, doubts) the bot sends a message “I need your judgment” instead of guessing. The client confirms/rejects in the Telegram queue. Sensitive actions (sending an email to an end customer, updating a deal in the CRM above a value threshold) ALWAYS require explicit accept from the client. The bot never acts autonomously on sensitive actions without consent.
Can I have multiple users in my company?
Yes, every user in the allowlist. The standard 3,000 PLN package covers up to 3 users (e.g. CEO, sales director, ops manager). Above 3 users we add 500 PLN setup per user. Role-based commands: some commands available only to the CEO (e.g. financial reports), others to the team (e.g. calendar booking).
What if I cancel after 1 month?
The retainer has a 30-day notice period. After notice in a given month, the retainer is payable for the following month, then the bot is shut down. The 3,000 PLN setup fee is non-refundable after go-live (the client received a complete build). The JSON voice profile, system prompt, configuration documentation stay with the client (export within 7 days after termination). The client can self-host the bot or hire another dev studio for continuation.
Is it compatible with WhatsApp?
Yes, WhatsApp Business API is an additional option, plus 500 PLN setup added to the package. Telegram remains the primary channel due to: better Bot API (Telegram is more developer-mature), no monthly costs per message (WhatsApp Business charges per conversation), better rich messaging performance (inline buttons, quick replies). Clients who use WhatsApp with end customers sometimes want a bot on WhatsApp as well (parallel channel). It is feasible, but we recommend starting with Telegram.
Can I have the bot in English (or German) instead of Polish?
Yes, the bot is multilingual. Voice extraction works in any language from which the emails come. DACH (DE) and international (EN) clients have the same package from 3,000 PLN setup plus 800 PLN/mo. Mixed languages (e.g. CEO writes in Polish to PL clients and in German to DE clients) are supported: the bot detects language per context and responds in the appropriate one.
An AI assistant on Telegram at Hanse Studio is a dedicated bot with the client’s brand voice, Gmail/Calendar/CRM integrations, 4 to 8 custom commands and monthly improvement in the 800 PLN retainer. Setup from 3,000 PLN, go-live in 2 weeks. Before the assistant it is worth going through an AI audit to verify that the assistant is the right choice (vs the Automation package). For the full AI stack in the company see process automation. Ready for a conversation? Check the AI packages or book a 30-minute process call.