$3,600/yr in Slack vs $9/mo on Telegram: Same 4-Part Agent

By Lazar Milicevic · Published September 14, 2026 · 8 min read

Anthropic just gated @Claude in Slack behind Team and Enterprise tiers. Minimum floor for a 10-person team: $3,600 a year. I ship the identical four-part agent loop on Telegram for clients every week — the bill lands around $9/month total, not per user. Here's the architecture, the actual cost math, and the reason the price gap exists.

What Anthropic actually shipped in Slack

@Claude in Slack is a four-part loop. You tag the bot in a channel, Slack fires a webhook to Anthropic's backend, the bot reads thread context, calls Claude with a tool set attached (search, docs, connectors), then posts an async reply into the thread. That's the entire product surface.

The engineering is not the expensive part. The paywall is. To use it you need:

• Slack Business+ or Enterprise Grid — roughly $15/user/mo floor
• Claude Team or Enterprise on Anthropic's side — $25-30/user/mo floor
• Both billed annually, both per-seat

For a 10-person team that's a $3,600/year floor before you send a single message. Twenty-five people puts you north of $9,000. The compute cost of the actual Claude calls under normal SMB usage is a rounding error against that seat tax.

Anthropic didn't invent tagging a bot in a chat app. They branded it, tied it to Slack's enterprise SKU, and shipped a very clean install experience. That's a legitimate product — for the enterprise. It stops being legitimate the moment a five-person dental clinic thinks they need it.

The same four-part loop on Telegram

A dental practice I work with runs their entire front desk on Telegram, because that's where patients already text them. Nobody on staff has opened Slack in their life. So the agent went where the conversation was.

Same four parts, different runtime:

1. Telegram sends a webhook when a patient messages the bot
2. The bot pulls the last N turns of conversation from a local SQLite file
3. It calls Claude Haiku with tools attached — check the calendar, look up the patient, draft a reply
4. It posts the response back into the Telegram chat

Here's the minimum viable webhook handler. This is close to what actually runs in production, trimmed for readability:

from fastapi import FastAPI, Request
import httpx, sqlite3, os
from anthropic import Anthropic

app = FastAPI()
claude = Anthropic()
TG_TOKEN = os.environ["TELEGRAM_BOT_TOKEN"]

TOOLS = [
    {"name": "check_calendar", "description": "Find open slots",
     "input_schema": {"type": "object", "properties": {"date": {"type": "string"}}}},
    {"name": "lookup_patient", "description": "Fetch patient record by phone",
     "input_schema": {"type": "object", "properties": {"phone": {"type": "string"}}}},
]

@app.post("/webhook")
async def webhook(req: Request):
    update = await req.json()
    msg = update.get("message", {})
    chat_id = msg["chat"]["id"]
    text = msg.get("text", "")

    history = load_history(chat_id)  # from SQLite
    history.append({"role": "user", "content": text})

    resp = claude.messages.create(
        model="claude-haiku-4-5",
        max_tokens=1024,
        tools=TOOLS,
        system="You are the front desk assistant for Dr. Kim's dental office.",
        messages=history,
    )

    reply = handle_tool_loop(resp, history)  # runs tools, gets final text
    save_history(chat_id, history + [{"role": "assistant", "content": reply}])

    async with httpx.AsyncClient() as client:
        await client.post(
            f"https://api.telegram.org/bot{TG_TOKEN}/sendMessage",
            json={"chat_id": chat_id, "text": reply},
        )
    return {"ok": True}

That's the whole runtime. About 200 lines of Python with the tool loop and SQLite helpers expanded. It runs behind Caddy or Nginx on a $5/mo VPS. Telegram sets the webhook URL once with a POST to setWebhook — no polling, no cron, no queues.

What each part actually does

• Webhook in: Telegram Bot API pushes updates to your HTTPS endpoint the moment a message arrives. Zero latency vs. polling.
• Context load: SQLite stores last ~20 turns per chat_id. Rolling window keeps token cost predictable.
• Model call: Claude Haiku 4.5 handles routine intent + tool selection at pennies per thousand messages. Escalate to Sonnet only when the tool loop needs multi-step reasoning.
• Message out: A single sendMessage POST back to the Bot API. Async, no delivery infra to run.

The real bill, line by line

I get this question every time I post about a Telegram deployment: "what does it actually cost." Here's the itemized monthly bill for the dental clinic setup, running roughly 200 patient messages per day:

Compare to the Slack path for the same 10-person team:

~35x price delta for the same four-part loop. The Slack side buys you SSO, audit logs, SOC 2, and enterprise admin. If you're a 200-person company with a security team and a compliance officer, that math is fine. If you're a five-person clinic, agency, or e-commerce shop, you're paying $3,591/year for features you'll never open.

Why Slack is the wrong runtime for small business

Slack is B2B corporate infrastructure. Small business runs on consumer messaging. Your clients are not sitting in your Slack workspace waiting for updates — they're texting you on WhatsApp, Telegram, iMessage, or straight SMS. Meta reports over 200 million businesses use WhatsApp Business globally. For a solo operator, a local service business, or a lean e-commerce team, that's where the conversation actually is.

The corollary is uncomfortable for a lot of AI vendors: if your agent only exists inside a workplace tool, you've built for the top 5% of the market. The other 95% — local commerce, one-person practices, small agencies, non-Western SMB entirely — talks to customers on channels Slack doesn't touch.

That's not a Slack criticism. Slack is excellent at what it does. It's a channel-fit criticism. You don't put a receptionist agent inside an internal team chat when patients are texting the front desk phone number.

The pattern is not Telegram-specific either

The four-part loop is portable. Every messaging platform with a webhook API becomes an agent runtime with the same recipe:

• Telegram Bot API — free, instant webhook, no approval
• WhatsApp Business Cloud API — free tier from Meta, ~$0.005-0.08 per conversation depending on category and country, requires a Meta Business verification
• Discord — bot user + interaction webhook, free, great for community ops
• Viber Business — webhook-based, common across Eastern Europe and SE Asia
• SMS via Twilio — ~$0.0079/msg US, universal reach, no app install
• Instagram/Messenger Graph API — for DMs on Meta properties

The code changes are surface-level. Auth headers, payload shape, and the sendMessage endpoint. The core — intent → tools → reply — is identical. I've ported the same base agent between Telegram and WhatsApp in an afternoon. The hard part is the tool integrations (calendar, CRM, order system), and those don't change when you swap the runtime.

The mental model that matters

• Don't ask which AI tool your team should adopt. Ask where the conversation already happens.
• Meet the human on their existing channel. Adoption goes to zero the moment you make them install something new.
• The runtime is a delivery mechanism. The business logic lives in your tools, your prompts, and your data.

When Slack's price actually makes sense

I want to be fair here because "Telegram is cheaper" is a shallow read. There are real cases where the Slack + Claude Team bill is the right check to write:

• You have 50+ employees and internal ops chat lives in Slack already
• You need SSO/SAML for compliance (HIPAA, SOC 2, ISO 27001 audits)
• You need audit logs of every AI interaction retained for legal
• Enterprise admin controls — deprovisioning, DLP, retention policies — matter to your risk team
• Your users would riot if they had to open a new app

If three or more of those apply, buy the Slack integration. It's a good product for that buyer.

If none of them apply, you're the buyer this post is for.

Where bizflowai.io fits in

The dental clinic bot above is a representative build, not a one-off. At bizflowai.io I ship this exact four-part agent — webhook in, context load, Claude tool call, message out — as a productized service for solopreneurs and small teams on Telegram, WhatsApp Business, and SMS. Same architecture, tuned to the client's actual channel and their real tools (calendar, invoicing, CRM, inventory), typically live within a week and running on infrastructure the client owns. The point isn't to compete with Slack; it's to make sure small businesses don't pay Slack's price to get functionality their customers won't ever see.

The takeaway

Anthropic and Slack will win enterprise headlines with @Claude because they both have the marketing budget to make sure of it. That's fine. Enterprise headlines aren't where AI adoption in small business is actually going.

The real growth over the next two years is agents that live on WhatsApp Business, Telegram, and SMS — because that's where local commerce, service businesses, and the entire non-Western SMB market talk to customers. If you're building an agent and it only works in Slack, you're building for 5% of the market and paying 35x the infrastructure cost to do it.

Same four parts. Different runtime. Pick the one your customers already have open.

Want more like this?

I publish practical AI automation, GenAI engineering, and faceless content workflows on YouTube every week.

Subscribe to bizflowai.io on YouTube — never miss a new tutorial.

Planning an AI automation project or need a second opinion on your architecture?

Connect with me on LinkedIn — Lazar Milicevic, GenAI Engineer & bizflowai.io Founder.

Visit bizflowai.io for our services, case studies, and AI consulting.