AI agents (Hermes)

AI agents for business: how they differ from chatbots and how to deploy

9 min read Updated: Czytaj po polsku
AI agents for business: how they differ from chatbots and how to deploy

An AI agent is a system in which a language model decides for itself which actions to take to reach a goal. It calls tools (your CRM, calendar, order database), checks the results and repeats that loop until the job is done. A chatbot mainly answers questions, and classic automation runs a pre-planned sequence of steps. An agent earns its place where the next steps can’t be predicted, and it needs stronger safeguards than either.

What is an AI agent?

An AI agent is software in which a large language model (LLM) interprets a request and decides which actions to take to fulfil it, usually by calling tools. That’s how the n8n glossary defines it. Anthropic’s guide “Building effective agents” draws the line even more sharply. A workflow is a system where models and tools are orchestrated through predefined code paths. An agent is one where the model dynamically directs its own process and tool usage.

That distinction matters in practice. With a workflow you know exactly what happens after step 3. With an agent you only know its goal, which tools it can reach and which boundaries it must not cross. Designing an agent is largely about designing those boundaries.

AI agent vs chatbot vs automation

Feature Chatbot Automation (workflow) AI agent
Who decides the next step A script or model, within the conversation only The process designer, in advance The model, within set boundaries
Access to systems Usually none, or a read-only FAQ Fixed integrations in a set order A toolset it chooses from
Memory The current conversation Data passed between steps Conversation, past sessions, knowledge base
Predictability High Highest Lower, so it needs testing and limits
Cost and latency Low Low Higher, with several model calls per task
Main risk A wrong answer A broken integration A wrong or manipulated action in your systems
Example “What are your opening hours?” A form lead goes to the CRM and Slack “Move my booking to Friday and send me a confirmation”

In real life the lines blur: give a chatbot tools and it becomes an agent, and a workflow can contain an agentic step. What matters isn’t the label but who decides on actions and what happens if that decision is wrong.

How does an AI agent work?

An agent runs in a loop. It receives a message or a scheduled trigger, reads its instructions, weighs up the request and decides what to do: call a tool, search its knowledge base, hand off to another agent or ask a follow-up question. It repeats this until it has a final response. n8n’s agent documentation lists the parts you configure:

  • Model: the LLM that reasons and writes responses.
  • Instructions: the system prompt with role, tone and constraints.
  • Tools: workflows, custom code, integrations and MCP servers.
  • Skills: reusable bundles of instructions and tools for a specific task.
  • Knowledge base: files the agent can search, such as terms, price lists and procedures.
  • Memory: the current conversation plus context from earlier sessions.
  • Channels and schedules: Slack, Telegram and other places people reach it, plus recurring jobs.
  • Sub-agents: specialist agents the main agent can delegate work to.

What can a business use an AI agent for?

  • Customer service with system access. The agent checks an order or booking, reschedules within allowed limits and sends a confirmation. Refunds and complaints go to people. Autonomy: medium, with approval for anything touching payments.
  • A sales assistant in Slack. Ask it to “prep me for my meeting with Company X” and it pulls CRM history, recent emails and notes into a brief with open questions. Read-only, so the risk is low.
  • A scheduled research agent. Every week it scans chosen sources (tenders, competitor sites, regulatory updates in your sector) and sends a digest with links. Every claim needs a source you can verify.
  • An internal helpdesk. It answers staff questions about procedures, leave or tool access from current documents, and opens an IT ticket when needed.
  • Back office and documents. It reconciles orders, deliveries and invoices, explains mismatches and proposes corrections for a person to approve.
  • An operations agent. It watches key signals (site uptime, stock levels, unusual orders), describes the problem and proposes a fix instead of firing off yet another bare alert.

Plenty of what companies call an “agent” is really a well-designed workflow with one AI step. We cover those in AI automation for business: 15 practical examples.

When don’t you need an AI agent?

In June 2025 Gartner predicted that over 40% of agentic AI projects would be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. It also warned about “agent washing”: rebranding ordinary chatbots and assistants as agents. According to Reuters’ report, Gartner estimated that only about 130 of the thousands of “agentic AI” vendors are the real thing.

Anthropic’s advice is to find the simplest solution possible and only add complexity when it’s needed. Agentic systems often trade latency and cost for better task performance, and their autonomy brings higher costs and the risk of compounding errors. A simple test: if you can sketch the process as a flowchart with a few branches, build a workflow. Add an agent only where the number and order of steps genuinely can’t be predicted.

AI agent security: the risks that matter most

An agent with system access is a new member of staff, except one that can be manipulated with text. The OWASP Top 10 for LLM Applications 2025 includes four risks that are especially relevant to agents:

  • LLM01 Prompt Injection. Attacker-supplied content in an email, web page or document changes the model’s behaviour and overrides its instructions.
  • LLM02 Sensitive Information Disclosure. The agent reveals data it shouldn’t, such as another customer’s details.
  • LLM06 Excessive Agency. The agent has more permissions and autonomy than it needs, so a faulty or manipulated decision causes real damage.
  • LLM10 Unbounded Consumption. No usage limits, so a runaway loop of calls racks up costs or overloads a system.

The safeguards we put into every deployment:

  1. Least privilege. A dedicated service account for the agent, read-only by default. Every write tool is added deliberately.
  2. Human approval for irreversible actions. In n8n, human-in-the-loop for tools holds back an email, a deletion or a payment until someone approves it in Slack or chat.
  3. Input and output filters. The Guardrails node detects jailbreak attempts and off-topic requests and can replace personal data and secrets with placeholders.
  4. External content is data, not instructions. A customer email or a fetched web page must never be able to rewrite the agent’s instructions. High-risk tools shouldn’t be available in the same conversations where the agent reads untrusted content.
  5. Isolated code execution. If the agent runs code or commands, it does so in a sandbox such as a Docker container, never on your production server.
  6. Limits. A maximum number of steps per task, a token budget with your model provider and a request cap per user.
  7. Logging and testing. Every tool call is logged. Anthropic recommends extensive testing in sandboxed environments, with appropriate guardrails, before an agent goes in front of people.

AI agents and the law: the AI Act and GDPR

If you operate in the EU, Article 50(1) of the AI Act has applied since 2 August 2026. Systems intended to interact directly with people must be designed so that people know they’re dealing with AI, unless that’s obvious to a reasonably well-informed, observant user. In practice, an agent on your website or WhatsApp should introduce itself as an AI assistant.

GDPR applies to an agent like any other processing. You need a lawful basis, data minimisation, data processing agreements with your model and platform providers, and clarity on where data is processed. Watch out for Article 22 GDPR too: people have the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. Leave things like refusing trade credit or rejecting a complaint to a human.

How to roll out an AI agent, step by step

  1. One process, one success metric. For example: “the agent resolves a set share of order-status questions on its own, and hands the rest to the team with full context”. Without a metric you can’t tell whether the pilot worked.
  2. Map tools and permissions. List the systems the agent needs, which actions are reads and which are writes, and which need approval.
  3. Workflow first, agent second. Build the fixed parts (verifying the customer, fetching data) as ordinary steps, and leave the agent only what genuinely needs judgement.
  4. A knowledge base with an owner. Up-to-date documents and a named person who keeps them current. An agent quoting last year’s prices is worse than no agent.
  5. A test set. A few dozen real cases, including awkward, unusual and malicious ones (attempts to extract another customer’s data). Run it after every model or prompt change.
  6. A pilot with a human in the loop. A limited group of users, approval on every write action, and a daily review of conversations.
  7. Increase autonomy gradually. Remove approval only from actions where the agent has been reliably right for a sustained period. Monitoring stays on for good.

What should you build your agent on?

Approach Best for Watch out for
n8n: the AI Agent node or Agent Builder Agents embedded in business processes, with many integrations, human approval and self-hosting Agent Builder is in Preview, and complex conversation logic needs careful design
Hermes Agent (Nous Research, MIT licence) An assistant with memory across sessions, skills it builds from experience, scheduling and one gateway to 20+ platforms (including Telegram, Slack, WhatsApp and Discord) It’s an autonomous agent running on a server, so permissions and sandboxing need careful setup
Custom code (a model provider’s SDK) An agent that’s part of your own product, with unusual requirements The most build and maintenance work, and you implement every safeguard yourself

According to its documentation, Hermes Agent ships with 60+ built-in tools, MCP server support, sub-agents for parallel work and six terminal backends (local, Docker, SSH, Daytona, Modal and Vercel Sandbox). It works with Nous Portal, OpenRouter, OpenAI or any compatible endpoint. We go into detail in our Hermes Agent guide for businesses. Often the best results come from combining the two: the agent talks and decides, while actions in your systems run through tested workflows in n8n.

Thinking about an AI agent for your business but unsure where to start or how to keep it safe? See how we build AI agents for businesses, or get in touch. We’ll start by analysing one process and running a pilot with a human in the loop.

Frequently asked questions

How is an AI agent different from a chatbot?

A chatbot mainly answers questions in a conversation. An AI agent decides which actions to take and carries them out with tools inside your systems, such as checking an order, changing a booking or opening a ticket. Give a chatbot tools and let it choose its own actions, and it effectively becomes an agent.

Are AI agents safe for businesses?

They can be, if they run with least privilege, irreversible actions need human approval and inputs and outputs pass through filters. OWASP lists prompt injection, sensitive information disclosure and excessive agency among the biggest risks. You also need cost limits, logging and testing before go-live.

Do I have to tell customers they are talking to an AI agent?

In the EU, in most cases, yes. Since 2 August 2026, Article 50(1) of the AI Act requires AI systems that interact directly with people to make clear they are AI, unless that is obvious. The simplest approach is for the agent to introduce itself as an AI assistant at the start of the conversation.

When is regular automation better than an agent?

When the process can be drawn as a flowchart with a few branches and the steps are known in advance. A workflow is cheaper, faster and more predictable. Add an agent only where the number and order of steps genuinely cannot be predicted, such as unusual customer requests.

What should I build a business AI agent on?

Most often n8n (the AI Agent node or Agent Builder), when the agent needs to work inside business processes with many integrations. Nous Research's Hermes Agent suits a standalone assistant with memory that people reach through messaging apps. Choose custom code when the agent is part of your own product.

Sources

  1. Anthropic — Building effective agents
  2. n8n Docs — Key concept glossary
  3. n8n Docs — Build and manage agents
  4. n8n Docs — AI Agent node
  5. n8n Docs — Human-in-the-loop for tools
  6. n8n Docs — Guardrails node
  7. OWASP — Top 10 for LLM Applications 2025
  8. Gartner — Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 Jun 2025)
  9. Reuters via The Star — Over 40% of agentic AI projects will be scrapped by 2027, Gartner says
  10. EU AI Act — Article 50: Transparency obligations
  11. GDPR — Article 22: Automated individual decision-making
  12. Hermes Agent — Documentation (Nous Research)

Need help with this?

An AI agent is more than a chatbot: it plans steps, uses tools, remembers context and works in your channels. We deploy agents built on Nous Research’s Hermes Agent and n8n.