How is an agent different from an automation?

An automation runs a fixed sequence: the same steps in the same order every time. An agent gets a goal and decides which of its tools to use in which order. That is the right choice exactly when the cases differ, and the wrong one when they do not. For the second kind you want an automation, not an agent.

How is an agent different from a chatbot?

A chatbot answers in a chat window and has no hands. Whatever it writes, somebody copies into the system it should have gone into. An agent has tools: it reads the enquiry, looks up the customer, pulls the price and writes the draft into the ERP itself. The chat window is optional.

Does the agent make decisions that commit us?

Only the ones you hand it. For every agent we set what it may do itself and where it hands over. Anything binding that leaves the company, a quote for instance, is checked by a person first. Every step stays logged and traceable.

What happens when it does not know something?

It hands over rather than guessing. An agent that invents something plausible when in doubt costs more checking time than it saves. The boundary where it hands over is part of the design, not a fault discovered later.

Does it run against our existing systems?

Yes. An agent with no connection to your ERP, CRM, mailbox or ticketing system is a chat window you retype things into. Your systems stay as they are. The agent comes on top, and the integration is the part of the work that decides whether a case holds up.

Does our data stay with us?

Yes. Processing in German data centres under a data processing agreement under Article 28 GDPR, or directly inside your own infrastructure, and contractually and technically separated from model training. For every agent we record the EU AI Act risk class before it goes live.

All services

AI agents that work inside your ERP, CRM and mailbox

They take on the work that repeats: putting quotes together, answering enquiries, creating orders, directly inside your systems.

Last reviewed: September 2026

The work an agent is for

Every business has work that repeats but is never quite the same. A quote: each enquiry asks for a different mix, the answer draws on three systems, and the person who does it well is the one you can least spare. An enquiry about a delivery: the answer is in the ERP, but somebody has to read the email, find the order and write back. A new order: the data is in the confirmation, and a person keys it in.

You notice it on a Friday afternoon, when the quotes go out late because the one person who prices them was in the workshop all day. You notice it when an enquiry waits for somebody to return from leave. The cost lands on your most experienced staff, because the work needs judgement and they are the ones who have it.

Agent, chatbot, automation

The three words get used interchangeably, and buying the wrong one is the most common mistake we see.

A chatbot answers questions in a window. It has no hands. Whatever it writes, somebody copies somewhere else.

An automation runs a fixed sequence: the same steps in the same order. It is reliable exactly where the cases are alike, and a list of exceptions that never ends where they are not. That is process automation.

An agent gets a goal, a set of tools and a boundary. It reads the enquiry, looks up the customer in the CRM, pulls prices from the ERP, drafts the quote and stops at the boundary: a person sends it.

The test is short. If you can write the rule down, you want an automation. If every case needs a person to look at three things and decide, you want an agent. If nobody needs to act, only to know, you want a knowledge system.

How we build one

Cut the case. One job, one boundary. Two jobs in one agent are two agents getting in each other’s way.
Write the boundary before the tools. What the agent may do itself, what it prepares for a person, what it must never do. The head of the department signs that page, not the IT department.
Connect the tools. ERP, CRM, mailbox, ticketing, file storage. Only the ones the job needs, each with a credential issued to your organisation that carries the rights of the person the agent works for, and no more.
Build the test set. Real cases from the past months with known correct outcomes. Every change is run against them before it reaches anyone.
Run it in the shadow first. The agent drafts, a person does the job as before, and the two are compared. When the department is satisfied, the agent’s draft becomes the starting point.
Log everything, keep the sign-off. Every step traceable. Anything binding that leaves the company is checked by a person first.

At the end you hold an agent running inside your systems, a written description of its job, its tools and its boundary, the EU AI Act risk class with purpose, data sources and limits documented, and the test set that tells you whether the next change made it better or worse.

A worked example: the quote

An enquiry arrives by email at a sheet-metal fabricator in North Rhine-Westphalia, with a drawing attached. The agent reads the email and identifies the customer in the CRM, or flags a new one. It extracts quantities, material and finish from the text and the drawing. It finds the last three comparable quotes, pulls current prices from the ERP and drafts the quote lines, with a note on what it was unsure about: the tolerance was not stated, the material is one you have not quoted for this customer before.

Sales opens the draft in the ERP, corrects two lines, and sends. The agent never sends. If the drawing cannot be read, or the customer is on credit hold, it hands over with the reason instead of guessing.

What you need to bring, and when it is not worth it

You need frequency. If the case comes up a few times a month, the build costs more than the handling. You need systems the agent can reach: an interface, an export, or at least a user interface, see integration into your systems. You need somebody in the department who owns the agent; one without an owner drifts. And you need past cases to show us, because without them there is no test set, and without a test set nobody can say whether it works.

It is not worth it when the rule can be written down, because then an automation is cheaper and more predictable. It is not worth it when the decisive information lives only in people’s heads. It is not worth it when a module in your ERP already does most of it and nobody switched it on.

How it differs from the neighbouring services

An agent decides per case within a boundary. Process automation runs the same route every time and is the better tool where the cases are alike. A knowledge system answers from documents and does not act; it is often one of the tools an agent uses. A voice agent is an agent whose channel is the phone, with the questions that channel brings. Integration is the connection all of them stand on.

The next step

If you know the case, the free introductory call is where we test it against the four questions above. If you do not, the assessment finds it before anything is built.

Common questions

How is an agent different from an automation?: An automation runs a fixed sequence: the same steps in the same order every time. An agent gets a goal and decides which of its tools to use in which order. That is the right choice exactly when the cases differ, and the wrong one when they do not. For the second kind you want an automation, not an agent.
How is an agent different from a chatbot?: A chatbot answers in a chat window and has no hands. Whatever it writes, somebody copies into the system it should have gone into. An agent has tools: it reads the enquiry, looks up the customer, pulls the price and writes the draft into the ERP itself. The chat window is optional.
Does the agent make decisions that commit us?: Only the ones you hand it. For every agent we set what it may do itself and where it hands over. Anything binding that leaves the company, a quote for instance, is checked by a person first. Every step stays logged and traceable.
What happens when it does not know something?: It hands over rather than guessing. An agent that invents something plausible when in doubt costs more checking time than it saves. The boundary where it hands over is part of the design, not a fault discovered later.
Does it run against our existing systems?: Yes. An agent with no connection to your ERP, CRM, mailbox or ticketing system is a chat window you retype things into. Your systems stay as they are. The agent comes on top, and the integration is the part of the work that decides whether a case holds up.
Does our data stay with us?: Yes. Processing in German data centres under a data processing agreement under Article 28 GDPR, or directly inside your own infrastructure, and contractually and technically separated from model training. For every agent we record the EU AI Act risk class before it goes live.

Check your potential, free