Plan, build and scale agents that do real work
A chatbot talks. An agent searches, calculates, drafts and prepares — with a person who approves.
Why that first chatbot never changed your work
By now everyone has tried a chatbot. Most organisations stopped afterwards, because the conversation was pleasant but nothing about the work changed. That is not the language model's fault. It is because the chatbot could not reach anything: an agent that cannot search your catalogue, does not know your volume tiers and may not use your quoting system can only talk.
We build agents that can reach those systems, through MCP, with clear limits and an approval step. That is what actually moves the work: from looking things up and retyping them to checking and sending. It is also the only promise we make. Not that an agent takes over your trade, but that the handling around it disappears and that you can see afterwards who took which step.
What an agent is, and what it is not
The word agent is used for three very different things, and that difference decides whether anything about your work changes. So we keep to this split and, on any request, we first say which level something sits at. That avoids the disappointment that starts with a demonstration at level one and an expectation at level three.
The three levels rise in what they are allowed to touch, not in how clever they sound. An assistant that knows your whole archive sounds more impressive than an agent allowed to carry out a single action — but only the second one takes work out of your day. So with any supplier, watch what the system may do, not what it can say.
The model
It can understand and produce language, but knows nothing about your company. A good writer with no access to your archive.
The assistant
It knows your documents and answers questions. Useful, but it changes nothing about the work: you still have to build the quote yourself.
The agent
It may carry out actions in your systems: search, calculate, create, prepare. That removes part of the work instead of moving it somewhere else.
“A chatbot holds the conversation. An agent holds the conversation and does the work.”
One agreement on what an agent may request and may do
MCP is the standard for connecting an agent to a system safely. Not a bespoke integration per question, but one agreement on what an agent may request and may do. We run MCP servers for our PIM, the quoting module, the ERP and the mailbox, and we connect to systems you already have. Per connection we record what is readable and what has to pass a person.
That is not only a safety measure, it is also why it works. An agent that can reach your real prices and volume tiers through MCP gives an answer that holds; an agent that has to take it from a manual invents something that merely sounds plausible. With us the MCP servers sit inside Magic Orb — not a separate product you buy on the side, but the layer that makes the rest of Magic Studio reachable.
How we go about it
Explore
We follow your process and look for the place where most of the handling sits. Usually that is the quote. You get a short list of opportunities and an estimate of what each one saves — including the ones we advise against, with the reason. That second list is often the more useful of the two, because it spares you the project you would have had to call off yourself.
Demonstrate
We build one agent on your own data and run it on real requests from the past few weeks. You see what it does well and where it is off, before you invest. This is the step most suppliers skip, and it is the only step that gives you certainty about what you are buying rather than about what ought to be possible.
Build
The agent gets its connections through MCP, plus limits, logs and an approval step. Which steps it may take is written in a flow you can read — not in code you have to trust. What it may not do is in there just as firmly.
Keep it running
Monitoring, adjusting and extending. An agent is not a project but a colleague who needs the occasional refresher. Every month we look together at what it could do better and what changed in your range or your prices.
Three agents that are running today
The quoting agent
It reads the request, finds the item, picks the technique, calculates the volume tier and prepares the quote. Your inside sales team checks and sends; the looking up and the calculating are gone. At the Brinxx group that is roughly 380 quotes a month, across Brinxx, Promotionalz and Moxz together. Not a pilot but everyday practice, which is also why we know where it still gets things wrong.
The product data agent
It fills in missing copy, translates into the languages you need and flags anomalies in supplier data. The demo recording Magic PIM — Catalogus beheren met de AI-agent shows it on our own catalogue: 74,136 products, 61,693 variants and 398,678 translation lines in eight languages, maintained from eight supplier feeds that each have their own format and their own quirks.
The customer question agent
It answers questions about delivery times, print options and prices from your own data, and hands over to a person the moment it gets delicate — including mid-conversation on WhatsApp.
An agent is only as good as the data it can reach
So every AI project with us starts with the product data: are the variants right, are the print areas in there, is the copy multilingual? Skip that step and you build an agent that talks nonsense fluently — and you only notice when a customer asks. On our side that foundation is in place: 426,573 technique rules from eight supplier feeds, in eight languages, with 14,952 products ready for Magic Logo.
It is also why, in a first conversation, we would rather run a data scan than demonstrate a chat window. A chat window always looks good. A count of your variants, your print areas and your missing copy tells you what an agent can really do in your house, and in which order it has to be sorted out. That count takes a day.
What we do not do
There is a line, and it is not in the small print. We do not put an agent between you and your customer without an approval step, and we do not send quotes out automatically. What an agent may do is written in a flow you can read, and every step it takes lands in a log you can check yourself without calling us in.
We also do not let an agent change prices on its own, and we do not use your business data to train third-party models. That costs a little convenience. It is also why your inside sales team still trusts the agent after two weeks, and that weighs more than a demonstration in which everything seems to happen by itself.
Four recordings of agents on real data
What is written above is easier to watch than to read. For four of the agents running today there is a recording: no animation and no screen design, but a screen in which someone asks a question and the agent does the work. So you also see the moments where it asks for confirmation, and that is exactly the part a sales pitch leaves out.
The four below cover the quote, the two channels a request comes in through, and the question inside sales ask a supplier most often. The full series of twenty recordings is on the demo videos page.
Talk to an AI expert
We start with your requests from last week. Within an hour that shows where the handling sits and whether an agent changes anything about it.
Rather look first? The recordings of agents on real data are ready.