An AI agent is software that receives a task, decides how to complete it, and uses tools to finish it. Unlike a chatbot that answers questions or automation that follows a fixed flowchart, an AI agent chooses which steps to take based on the situation.
We build agents on top of automation platforms like n8n, with connections to CRMs, the WhatsApp Cloud API, Google Sheets, and internal systems. The agent receives a general instruction – for example, “handle inquiries coming through WhatsApp” – and then decides for each inquiry whether to answer directly, escalate to a person, update a CRM record, or request more information.
How is an AI agent different from regular automation?
Regular automation follows a flowchart someone built in advance: if X happens, do Y. An AI agent receives a goal and chooses how to reach it, using tools we defined for it.
In regular automation with n8n or Make, we define every condition and branch. If a customer writes “I want to cancel my order”, the automation checks for the word “cancel” and triggers a specific flow. If they write “can you stop my order?”, and the word “stop” isn’t in the list we defined, the automation won’t recognize the intent.
An AI agent understands both inquiries as a cancellation request, even if the phrasing differs. It decides whether it has enough information to cancel, checks the policy we defined, and if needed – asks a clarifying question or escalates to a person. This decision wasn’t written in advance in a flowchart; it’s made based on context.
The following table compares the two approaches:
| Criterion | Regular Automation | AI Agent |
|---|---|---|
| How behavior is defined | Detailed flowchart with all conditions | General goal and available tools |
| How new cases are handled | Need to edit the flow and add conditions | Agent decides if existing tools are sufficient |
| When it fits | Repeating processes with fixed structure | Processes requiring judgment or handling high variance |
| Maintenance cost | Low if the process doesn’t change | Higher – need to check decisions and define policy |
We use regular automation for stable processes – moving leads from a form to a CRM row, sending a weekly report, updating status when an order is received. We build an AI agent when the process requires understanding intent, choosing between options, or handling situations we haven’t seen before.
What can an AI agent do in a business?
An AI agent in business can handle customer inquiries, manage workflows requiring judgment, and operate systems based on context. It works within infrastructure we build – a workflow in n8n, connection to a CRM like HubSpot or Pipedrive, interface to the WhatsApp Cloud API, and access to internal knowledge sources.
A practical example: an agent handling support inquiries on WhatsApp. When a message arrives, the agent reads it, identifies the topic, checks if there’s an open CRM ticket for this customer, and decides whether to answer directly based on a knowledge base, ask a clarifying question, or escalate to a person. If it answers, it also updates the CRM ticket with the response it gave. If it escalates, it writes a brief summary for whoever receives the inquiry.
All these decisions are made based on general instructions we defined – “don’t handle billing complaints, always escalate to a person”, “if the customer asks about order status, check the system and answer directly” – but the agent decides for each inquiry separately which instruction is relevant and how to apply it.
Other agents we build: an agent managing a document collection process (sends reminders, checks if the received document is correct, updates status), an agent filtering leads before they enter the CRM (asks clarifying questions, decides if the lead qualifies, fills initial fields), and an agent generating reports on request (understands which data to pull, generates the report, and sends it in the appropriate format).
The common denominator: the agent works within a system we control. It doesn’t “do whatever it wants” – it chooses from tools we defined, with policies we wrote, and with a log that records every decision. We see what it did, why it decided that way, and can change the instructions if the behavior isn’t appropriate.
How do you build an AI agent that works?
Building an AI agent starts with defining the task and tools, continues with building the workflow that connects them, and ends with monitoring that ensures the agent behaves correctly.
We build the agent in n8n, where we define:
General instructions: what the agent should achieve, what its boundaries are, and what policy it should operate by. This is written in natural language, not code – “you handle support inquiries. If the inquiry relates to billing or cancellation, always escalate to a person. If the inquiry relates to product information, answer based on the knowledge base. If you’re not sure, ask one clarifying question, and if still unclear – escalate to a person.”
Available tools: which actions the agent can perform. Each tool is a node in n8n that the agent can trigger: sending a WhatsApp message, searching for a CRM ticket, updating a field, reading a document from Google Drive, querying an internal knowledge base. We define each tool with a clear description of what it does and when to use it.
Knowledge source: where the agent knows what to answer from. This can be a repository of documents we upload, a table of questions and answers in Google Sheets, or a direct query to an internal system. This information needs to be current – an agent working with a price list from a year ago will give wrong answers.
Approval process: whether the agent performs actions directly or sends them to a queue for human approval. Initially, we typically define that any action changing state – updating CRM, sending a message to a customer, cancelling an order – goes through approval. After we see the agent deciding correctly, we release it to direct action under certain conditions.
After the workflow is built, we test it with real cases. We don’t release an agent to production after three tests – we run it on dozens of inquiries, see where it decides correctly and where it errs, and fix the instructions or tools.
Monitoring is part of the system, not an add-on. Every agent decision is logged: what it received, which tools it triggered, what it decided, and why. We get an alert if the agent gets stuck, if it escalates too many inquiries to a person (a sign the instructions aren’t clear enough), or if it triggers a tool at unusual frequency. This information helps us improve the agent over time.
More information on building agents and security considerations is available on our AI agents solution page.
How much does it cost to set up an AI agent?
The cost of setting up an AI agent consists of three components: the cost of the platform you build on, the cost of the model powering the agent, and the cost of building and maintenance.
The automation platform – n8n or Make – costs according to plan. n8n offers a self-hosted version where you only pay for infrastructure (a server or Kubernetes cluster), and a cloud version starting from a basic plan and scaling by number of executions. Make works on payment per number of operations – every action the agent performs (API call, field update, message send) is counted.
The model itself – typically Claude or GPT through API – is measured by tokens: how much text was sent to the model and how much it returned. An agent handling short inquiries consumes less, an agent reading long documents or generating reports consumes more. Prices vary between models, but typically we’re talking about tens of cents per conversation.
Building cost depends on complexity: how many tools the agent needs, how many systems it connects to, and how much policy needs defining. A simple agent answering FAQs from a fixed knowledge base takes less time to build than an agent managing a multi-step process with approvals and updates in multiple systems.
Ongoing maintenance includes updating the knowledge base, adjusting instructions when policy changes, and handling cases the agent can’t handle. An agent working in a stable domain requires less maintenance than an agent working in a domain that changes frequently.
Bottom line: an AI agent costs more than regular automation, but less than a full-time position. It pays off when it handles high volume of tasks requiring judgment, or when it frees up time of people doing higher-value work. A detailed guide to calculating automation ROI is available in our article on ROI calculation.
When is an AI agent not suitable?
An AI agent is not suitable when the process requires absolute accuracy, when the cost doesn’t pay off relative to volume, or when the risk of error is too high.
Processes requiring absolute accuracy – payroll calculations, regulatory reporting, customer billing – are not suitable for an agent making decisions based on a language model. A language model can make mistakes, and it doesn’t understand mathematics the way code does. In these processes, regular automation with explicit logic is the right choice.
Low volume – if the process happens five times a month, the cost of building and maintaining an agent doesn’t pay off. In that case, a simple form or even manual handling can be more efficient.
High risk of error – if the agent can approve a payment, delete data, or send sensitive information, the risk of error is too high. In such cases, you can build an agent that recommends an action but doesn’t execute it, or require human approval for every critical action – but then part of the agent’s advantage is lost.
Processes that change constantly – if policy changes every week, or if the process isn’t well-defined yet, an AI agent will require frequent updates and make many mistakes. In that case, it’s better to start with regular automation that’s easier to change, and move to an agent only after the process stabilizes.
In summary: an AI agent fits processes with high volume, variance in inputs, and need for understanding intent or context. It doesn’t fit processes requiring absolute accuracy, low volume, or high risk of error. The right choice depends on the specific process, not the technology.
Want to know which automation is right for your business? A free 30-minute diagnostic call, no commitment, we'll map out your processes together and honestly tell you if and how automation can help.