An AI agent is a language model (GPT, Claude) connected to your tools: it reads data, takes actions in your CRM, e-mail or messengers and decides the next step until the task is done. Agents pay off first in lead handling, customer support and document work. A pilot for one process typically takes 20-200 hours of development; model usage is usually a small monthly cost.
What is an AI agent
A chatbot answers questions. An AI agent also acts. It works in a loop: the model reads the task, chooses a tool (search the knowledge base, read the CRM, send a message, create a document), sees the result and decides the next step, until the task is done or a person needs to approve it.
Chatbot vs automation vs AI agent
| Chatbot | Classic automation (n8n, Zapier) | AI agent | |
|---|---|---|---|
| Understands free text | Yes | No | Yes |
| Takes actions in your systems | Rarely | Yes | Yes |
| Handles unexpected cases | Partly | No | Yes, within set limits |
| Best for | FAQ and simple answers | Fixed, predictable steps | Multi-step tasks that need judgement |
Where agents pay off first
- Lead handling: replies within a minute, asks qualifying questions, writes the lead to the CRM and notifies sales
- Customer support: answers from your documentation, checks order status, escalates complex cases with a summary
- Inbox and documents: sorts e-mails, extracts data from invoices and contracts, prepares drafts
- Internal assistant: answers staff questions from policies, wiki and CRM
- Reporting: collects data from several systems and writes a weekly summary
- Content operations: drafts product descriptions, posts and replies for human review
What is inside: the agent platform (harness)
The model alone only turns text into text. Everything that makes it an agent is the code around it, often called the harness:
- The agent loop: send the task to the model, run the tool it asks for, return the result, repeat
- Tools: connections to your CRM, e-mail, messengers, databases and APIs
- Context and memory: which documents, history and rules the model sees
- Permissions: what the agent may do on its own and what needs human approval
- Logging and monitoring: every step, its cost and its quality are recorded
When this layer is built once and reused, each new agent is mostly configuration: a role, data sources and a set of tools. Some vendors call this an agent factory. For you it means a faster launch, a lower price and agents that behave consistently.
Typical cost
| Project | Typical effort | At $25-40/h |
|---|---|---|
| Simple agent in n8n (one process, 2-3 tools) | 20-60 h | $500-2,400 |
| Support or sales agent with RAG and CRM integration | 60-200 h | $1,500-8,000 |
| Multi-agent system or agent inside your product | 200-600+ h | $5,000-24,000+ |
Model usage (OpenAI, Anthropic and others) is billed separately; for most small and medium businesses it is tens to a few hundred dollars per month.
How to launch an agent safely
- Start with one process and measure the baseline: time per task, response time, errors
- Give the agent only the permissions it needs
- At first, require human approval for payments, deletions and messages to key clients
- Log every action and review a sample every week
- Expand autonomy only after the agent proves itself on real cases
Amsoft
Amsoft builds AI agents for lead handling, support, document processing and internal assistants on GPT, Claude and open-source models, integrated with your CRM, messengers and n8n. We build on a reusable agent framework, so a pilot for one process is usually ready in 1-3 weeks. Rates start from $25/hour; estimates are free.