aceArt · AI Agency

Have AI agents built

A chatbot answers. An AI agent gets something done. It takes on a task, works out the steps it needs, calls systems to carry them out and keeps going until it is finished or stuck. That is the interesting part, and it is also the part where you have to be careful.

Get expert advice

What separates an agent from a chatbot

The difference is not the model, it is what the model is allowed to do. A chatbot gets a question and returns text. An agent gets a goal and tools to reach it: a search, a database lookup, an interface into the CRM, the creation of a ticket. It decides for itself which tool to use and when, looks at the result and carries on.

From that follows the difference that matters: a chatbot can answer wrongly, an agent can act wrongly. So the interesting question in an agent project is never what it can all do, it is what it must never be allowed to do. More on the term in the glossary.

How to tell that an agent is the right answer

  • The task has several steps, and their order depends on what comes out in between.
  • Several systems are involved, and today somebody switches between them by hand.
  • Part of the work is a matter of judgement and cannot be written down as rules.
  • The task comes up often enough for the build to pay off.

If the sequence is always the same, you do not need an agent, you need plain automation. That is cheaper, faster and holds no surprises. More on that under process automation.

Where agents genuinely carry their weight

Groundwork in support

An incoming ticket is read, sorted into a category, enriched with the customer history and given a draft reply. A person approves it. The time saved is not in the writing, it is in gathering the information, which used to start from scratch every single time.

Research with a result

Pulling information out of several sources and putting it into a usable shape. Watching competitors, going through tenders, comparing data sheets. The agent delivers a compilation with its sources, so it stays possible to check where something came from.

Data maintenance

Finding duplicates, reconciling master data, filling in incomplete records and putting the doubtful cases up for review. Unspectacular, but often the case with the best ratio of effort to effect, because nobody else does it willingly.

Internal assistance

An agent that does not only answer from documents but also prepares the obvious next action: suggesting the appointment, pre-filling the form, opening the case. The usual growth path out of an existing chatbot.

Staying in control

Autonomy is not an end in itself. It is a setting you choose per task. An agent that sends out quotes needs different limits from one that sorts product data. So we build the brakes in from the start, rather than fitting them after something has happened.

  • Approval by a person wherever an action reaches the outside world or is hard to take back.
  • Write access only where it is needed. Reading is cheap, changing is not.
  • A log of every step, so it stays possible to see why the agent did something.
  • Upper limits on steps and cost, so a loop cannot produce a bill overnight.
  • A defined behaviour for the case where the agent gets stuck. Stopping and reporting beats guessing.

The term for this is Human-in-the-Loop, and it is often mistaken for a temporary measure. We see it as a permanent state for anything that touches money, law or customer relationships.

How we build agents

We start small, and we do that on purpose. The first version usually handles one step of the task and puts everything else up for review. Only once that step runs reliably does the agent get the next one. This takes a little longer than the grand design, but it has the advantage that you actually notice the point where things tip over.

Technically we work independently of any one model and connect the tools over open interfaces where that is possible, for instance over the Model Context Protocol. The reason is unromantic: models get replaced faster than projects run, and anyone who ties themselves to a single provider builds the switch twice later on.

Frequently asked questions about AI agents

What is an AI agent, and what makes it different from a chatbot?
A chatbot answers questions with text. An agent gets a goal and tools, and decides for itself which steps are needed to get there. It can call systems, fetch data and open cases. The practical difference: a chatbot can answer wrongly, an agent can act wrongly.
Where is an agent genuinely worth it?
Wherever several steps run across several systems and the order depends on what comes out in between. If the sequence is always the same, classic automation is enough, which costs less and behaves predictably. In the first conversation we say which of the two categories your case belongs to.
How do we stay in control?
Through four things: approvals by people at every point that reaches the outside world; permissions that go no further than necessary; a complete log of every step; and hard upper limits on steps and cost. On top of that we define what happens when the agent gets stuck, because stopping and reporting beats guessing.
What is the difference to Agentic AI?
Agentic AI is the umbrella term for the same field. It is used mainly in marketing and describes systems that plan and carry out steps of their own accord. We say AI agent on these pages, and it means the same thing.
What does it cost to develop an agent?
That depends on how many systems get connected and how strict the approvals are. The connections are usually more work than the agent itself. On top of that come running costs, because an agent sends several requests to the model for one task and is therefore more expensive to operate than a chatbot. We work that out with you beforehand.
What happens if the agent makes a mistake?
That is exactly what the limits are for. Actions with an outward effect go through an approval, and everything else is logged and can be traced back. When we pick the first use cases, we look for ones where a mistake is annoying rather than dangerous.
Can agents access our internal systems?
Yes, that is usually the point. CRM, ERP, ticket system, file storage and calendar are the common ones. Access runs over the interfaces you already have and through a separate technical user whose permissions we cut narrowly. What the agent does not need, it does not get.
Do we need our own infrastructure for this?
Not necessarily. An agent can run at a provider or in your own environment, depending on which data it sees. For sensitive data we recommend your own environment, even though that means more work to operate.

Related topics

AI chatbots & RAG

The usual way in, and often where agents come from in the first place. To chatbot development.

AI automation

For when the sequence is always the same and needs no judgement. To automation.

AI integration

The connection to the systems you already run. To AI integration.

Ready for your KI project?

Tell us about your project. We'll get back to you promptly with an honest assessment.

Contact us
HOME › AI-AGENTS