3 min read

How I stopped confusing chatbots with AI agents and fixed our support workflow

How I stopped confusing chatbots with AI agents and fixed our support workflow

A ticket that sat for 12 hours

Last month, a client wrote to our Telegram bot asking to reset their API key. The bot replied with a link to the docs. The client clicked, got lost, and waited 12 hours for a human to step in.

That bot was a chatbot. It could answer, but it could not act.

This is the difference I want to walk through. Not a theory. A real pain I fixed with AgentForge.

What I had before: two fake solutions

I tried a rule-based automation first. If the message contained the word “reset”, send the docs link. If it said “stuck”, escalate to a person. It worked for the first week. Then a user wrote “I need to rotate my key” and the bot sent them a FAQ about billing.

The rule system could not understand. It could match patterns, but pattern matching is not understanding.

Then I switched to a simple chatbot with a large language model. It felt smarter. It could explain the reset process in three paragraphs. The user still had to click, copy, paste, find the right page, confirm the action in two sub-menus, and wait for a confirmation email. Many gave up.

The chatbot could talk. It could not touch the system.

The real problem: conversation without action

The pain was not bad answers. The pain was that every answer forced the user to do the work inside a different tool. Our support flow had a gap between “I understand what you need” and “I have done what you need.”

A user who has to leave Telegram, open a dashboard, log in, find the right setting, click confirm, and return is not a supported user. That user is a frustrated user. And frustrated users churn.

The cost of doing nothing was visible in our retention numbers. Accounts that asked for a key reset had a 40% higher churn rate within 30 days. The problem was not the reset itself. The problem was the friction between the request and the action.

What I actually did

I set up an AI agent with AgentForge. I installed the Hermes agent on our own infrastructure, configured its tools through the web UI, and deployed it with Docker Compose. The agent lives in Telegram, where our users already are.

Here is what changed. The agent can read a message like “rotate my API key for project delta” and do three things in sequence: verify the user identity against our internal directory, call our backend API to invalidate the old key and issue a new one, and post the new key directly in the chat thread.

There is no link. No dashboard visit. No confirmation email. The agent executes the action.

An AI agent does not just answer the question. It completes the task.

I gave the agent four tools: identity verification, user profile lookup, API key management, and notification dispatch. Each tool is a small API call wrapped in a function that the agent can call autonomously. The agent decides when to call each tool based on the user’s intent, not on a fixed rule.

What changed after the switch

The same client who waited 12 hours now gets a new key in under 20 seconds. The agent handles 70% of our support requests end-to-end without a human in the loop. Our support team went from triaging password resets and key rotations to handling only the exceptions that require judgment.

The churn rate for users who interact with the agent dropped by half. The average time-to-resolution for technical requests went from hours to seconds. More importantly, the agents run on our own hardware. No data leaves our network. We control the deployment, the model, and the tool definitions.

I still use chatbots for quick FAQ lookups. I still use rule-based scripts for automated backups. But when a user needs something done, I route them to the agent. That one change fixed a persistent support bottleneck and stopped a leak in our retention.

AgentForge

AI agents for real work. If you want to see how this works in practice, visit our product site and blog.

Share
Serhii

Serhii