Every week a new vendor pitches you an "AI agent" that will run your business. Most of those pitches are wrong about what an agent actually is. An AI agent is software that observes a situation, decides what to do, takes an action, and records the result without someone standing over it. That is a meaningful step up from a chatbot. It is also narrower than a human employee. This post explains what that means for a small or mid-sized business: where agents pay off, where they do not, and how to decide whether to build one.
What an AI agent actually does
A chatbot answers a question when you ask it one. An agent keeps working on its own. It waits for a trigger: a new email, a scheduled time, or a change in a spreadsheet. When the trigger fires, the agent runs a loop. Read the situation. Decide what to do. Call a tool. Check whether the action worked. Try again or escalate to a human if it did not. Most agents that do real work are specialized. One handles customer service. Another drafts sales follow-ups. A third compiles a weekly revenue report. The value is not in the intelligence of any one agent. It is in removing a specific recurring task from your team's calendar.
How an agent gets from a trigger to a result
The pattern looks like this. A deterministic system, typically an n8n workflow, receives the trigger and handles everything rule-based: database lookups, simple formatting, yes/no decisions. Only when the situation is ambiguous does the workflow call a language model. This n8n-first pattern keeps costs predictable and keeps the agent reliable. In practice, roughly 80 percent of real business requests are handled by rules, not by the model. The model only gets invoked for the remaining 20 percent, where a rule cannot decide. After the agent acts, every step is logged: what the input was, what the model said, what the action was, and how long it took. The work is auditable the same way your bookkeeping is.
Where agents actually pay off
The break-even point is a task that happens at least weekly, takes a predictable shape each time, and costs 15 minutes or more of human attention per instance. In practice that means email triage, meeting notes and action-item extraction, invoice reminders, lead qualification from form submissions, and internal research summaries. A simple sanity check: if you can write down the decision rules on one page, an agent can do it. If the task requires real judgement about a client relationship, a contract, or a staffing decision, keep the human in charge and use the agent for the preparation work around it. The use cases section lists eight patterns that consistently produce measurable savings, from a digital employee covering internal coordination to an operations automation layer keeping CRM data clean.
What it actually costs
Most "AI agent" articles go vague here. The honest answer has two parts. The platform costs (model calls, infrastructure, and logs) land in the tens of euros per month for a single agent handling a few hundred tasks, not hundreds or thousands. That low floor comes from caching repeated answers and routing requests through layered fallbacks so that the expensive frontier model is only called when a cheap local model cannot answer. The second cost is the design work up front. A production-grade agent takes two to six weeks to specify, build, and tune against your actual data. That is the part where vendors undersell and where the quality of the result is decided. Per-agent daily spending caps should be non-negotiable. Ask any vendor about them before signing anything.
Where agents go wrong
Two failure modes account for most of the trouble. The first is the agent making something up. Language models can produce confident, plausible-sounding sentences that are factually wrong. The phenomenon is called hallucination, and in a customer-facing role it can reach your client before anyone notices. The fix is to constrain the agent. Give it only the tools and data it needs. Require a source citation for any factual claim. Route anything outside scope to a human. The second failure mode is the agent running on autopilot without anyone checking its output. The practical answer is human-in-the-loop review for anything that touches money, contracts, or customer communication. For low-stakes work like internal summaries, full autonomy is fine. Match the oversight to how much a mistake can cost.
How to start small
Pick one task. Write down its trigger, its decision rules, and the acceptable output format. Run the process by hand for a week to confirm the shape is stable. Then build the smallest possible agent that handles it, measure the time saved, and expand only after the first one has a month of clean logs. The related post on 5 tasks to automate this week lists concrete starting points. If you are weighing whether to build in-house or commission a studio, the build vs buy breakdown has the trade-offs. When you have picked a specific use case, book an intro call and we will walk you through it.
Frequently asked questions
Q: Is an AI agent the same as ChatGPT? A: No. ChatGPT is a general-purpose chat interface that waits for your prompt. An AI agent is software that runs on a trigger, follows a defined workflow, uses tools, and completes a task without constant prompting. ChatGPT is a component an agent can use; it is not the agent itself.
Q: How much does a basic business agent cost to run? A: A well-designed single-purpose agent handling a few hundred tasks a month typically costs under 50 euros in platform fees when a cache and local-model fallback are in place. The larger cost is the initial design and integration work, which depends on how clean your data sources already are.
Q: Can an agent replace an employee? A: It can replace a specific repeatable task an employee spends time on. It cannot replace the judgement, relationship work, or accountability of the person. The better framing is that an agent removes a chunk of boring work so the employee has time for work only a human can do.
Q: What happens if the agent makes a mistake? A: Every action is logged, so you can trace what input produced what output. For anything touching money, contracts, or customers, a human reviews the output before it ships. For low-stakes internal work, the agent runs autonomously and logs are spot-checked weekly.
Q: Do we need a technical team to deploy one? A: To build from scratch, yes. Someone has to understand workflows, APIs, and language model limits. To adopt a platform-delivered agent, no. The studio handles the implementation and your team owns the running workflow.
---
*Written by the Leap Laboratory team. Updated April 2026. Figures above reflect workloads we run on this site's own agents and on early Agent Leap client deployments.*