Have an AI agent built: when it pays off (and not)

What an AI agent is, in plain terms

A chatbot answers one question and stops. An AI agent goes further: it gets a goal, can call real tools and keeps taking steps until the goal is met. The engine is a language model (an LLM from Anthropic or OpenAI, for example). Around it you build three things: tools the agent can act with, memory so it holds context, and a loop that lets the agent choose, execute and evaluate.

A concrete example. You put an agent on your shared inbox. For each email it decides the intent (support, sales, invoice, complaint), pulls out the customer name and order number, looks up the order in your system and hands the email to the right person with a priority. That is not a single reply, it is a series of actions with real effects in your software.

The power sits in that acting. So does the risk: an agent allowed to act can also make mistakes that carry through. That is why you build in boundaries from day one, not afterwards.

What you can have an AI agent do

The strongest use cases are tasks vague enough that a fixed script cannot handle them, yet bounded enough that you can check the result. Think of work where a human currently reads text, looks something up and makes a decision.

• Classify incoming email or tickets, enrich them with data from your CRM and route them to the right person.
• Draft quotes or reports from data across several systems, with a human who approves.
• Do research: gather sources, summarize and keep the reference so you can verify it.
• Build an AI copilot in your product that helps users with tasks inside your own software, using your data and rules.
• Workflow automation with AI where several steps that used to be manual get stitched together.

When you should not have an AI agent built

This is the honest side. An agent is not always the best choice, and sometimes the most expensive way to solve a simple problem.

Is the task fixed and predictable? Then a plain script or off-the-shelf software is faster, cheaper and more reliable. An agent that runs the same fixed steps every time only adds cost, latency and unpredictability. Using an LLM for something an `if` statement can handle is waste.

Does every outcome have to be exactly right, with no room for a human check? Think of posting payments or finalizing legal text. A language model gives good odds, not guarantees. In that case always put a human in between, or pick a deterministic solution.

Do you not yet have clean data or clear processes? Fix that first. An agent on top of messy data mostly produces fast mess. In those cases you are better off with custom software or with improving business processes with AI on a smaller, concrete piece.

What does having an AI agent built cost

An honest answer: it depends on how many systems the agent must touch and how strict the requirements are. A bounded agent on one process, with a few tools and a human approval step, is a manageable project. An agent that reaches deep into your business software and may act autonomously is not.

The largest costs rarely sit in the model itself. They sit in the integrations (connecting to your CRM, helpdesk, accounting), in testing edge cases and in the guardrails that stop the agent doing something dumb with real data. Count on iterations: your first version is never the last.

For a realistic picture, our piece on the cost of custom software helps. The breakdown is similar. And start small: building an MVP of one agent on one process tells you within weeks whether the business case holds, before you invest big.

How we build an AI agent

We work in short iterations. First a kickoff where we check whether an agent is even the right solution, or whether a simple piece of software serves you better. Then scope and plan: which process, which tools, which boundaries, and where the human check sits.

Next we build a working first version on one bounded process, with logging so you see exactly what the agent does and why. We test on real edge cases, not on the pretty examples. If it works, we expand. If it does not, you know that early and cheaply.

Data and privacy are in the design from the start. In the Netherlands and the EU you fall under the GDPR and the AI Act; the Dutch Data Protection Authority is clear on what is allowed with personal data. If you want to connect the agent to existing systems, also read AI integration in business software. Unsure about a supplier, then look at how to choose a software agency.