Chatbot
answers- Understands the question
- Writes a friendly reply
- Asks for the item number and address
- ···
AI agents
Software that does not reply. It finishes the job.
A chatbot hands you a text. An agent hands you a result: read, checked, booked, sent, logged. Here is how one is built, which tasks actually hold up today and what you need to settle before you start.
GoalAnswer the quote request in full
- 01PLANRead request, pull price, build quote, send
- 02MAILRequest read: 500 units, item A-2041
- 03ERPVolume price for 500 units found
- 04RULEMargin 18% sits above the 12% approval floor
- 05PDFQuote AN-2026-0412 created
- 06CRMCase linked to customer 8821
- 07SENDReply out after 2 min 14 s
- 08STOPSpecial terms requested, handed to sales
An example run. The last step is the important one: an agent that cannot hand over is not an agent, it is a liability.
01 The difference
A chatbot answers. An agent works.
Same email, two systems. The difference is not the quality of the prose, it is the question of who did the work in the end.
Inbound
“Can you send me a quote for 500 units?”
AI agent
finishes- Reads the email and the attachment
- Pulls the volume price from the ERP
- Checks the margin against the approval floor
- Creates the quote as a PDF
- Opens the case in the CRM
- Sends the reply
- Writes the log
02 Anatomy
Seven parts no agent runs without.
Agents rarely fail at the language model. They fail at one of these seven parts. Click through them and you will see early where your own plan will get stuck.
01
Goal
What counts as done?
An agent needs a result you can tick off, not a topic. “Handle customer enquiries” is not a goal. “Answer every delivery-time enquiry within 15 minutes with the confirmed date” is one.
How you notice it is missing
Nobody on the team can say in one sentence when the agent is finished.
03 Autonomy levels
Autonomy is a dial, not a switch.
Almost every failed agent project started at the wrong level. Drag the dial and watch what moves with it: the work, the risk and what you need to have settled first.
L1
Draft
The agent finishes it, you release it.
Who does the work
The agent completes the task in full, but the result waits for your click. This is where almost every good project starts.
- Example
- The quote sits finished in the drafts folder, prices pulled from the ERP.
- Prerequisite
- A place where drafts wait visibly, and somebody who looks at them daily.
- The risk
- Contained. The most expensive outcome is a draft nobody reads.
We almost always start at L1 and only move up once four weeks of operating numbers justify it.
04 Use cases
Tasks that hold up today.
Not futurology. These are the cuts that run reliably in small and mid-sized companies, because they have a clear goal, digital tools and a handover point.
Inbox triage
Read enquiries, classify them, answer or pass them on with context to the right team.
Quotes
Understand the request, calculate the line items, build the document, send it, open the case.
Incoming invoices
Read documents, match them against order and delivery note, post them, report discrepancies.
Scheduling
Check availability, propose slots, confirm, send the invitation and the prep material.
Master data
Reconcile information across systems, find duplicates, report gaps.
Monitoring
Follow sources, filter what matters, summarise it briefly, route it to the right person.
05 Fit check
Does your task suit an agent?
Six questions, answered honestly. The scoring runs in your browser, we never see any of it.
06 Failure modes
Why agent projects fail.
Four patterns we see again and again. None of them is about the language model.
Scoped too big
“The agent runs our sales” fails reliably. The cut is so wide that nobody can say when the agent has done its job well.
What helps
One task, one trigger, one result. You widen it once the first number holds.
No escalation path
Without a clean handover point the agent keeps guessing when in doubt. Automation turns into quiet chaos that only surfaces at the customer.
What helps
Define the stop conditions first, not last. And put a named owner behind them.
Nothing measured
If nobody defines what “done” means, neither quality nor savings can be evidenced. The project ends in opinions.
What helps
Count from day one: cases closed, escalation rate, rework, hours saved.
Permissions too wide
An agent with full write access to the ERP is not a productivity project, it is a security incident on a delay.
What helps
Smallest possible permission per tool, hard limits in code and a kill switch everyone on the team knows.
07 Rules
What the EU AI Act asks of an agent.
An agent is not automatically high-risk AI. But it triggers obligations faster than a chatbot, because it acts and reaches the outside world. This is what affects most companies.
- Art. 4
AI literacy
In force since 2 February 2025Anyone deploying AI must make sure the people working with it understand it. For agents that means something concrete: the team knows what the agent may do, how an error looks and how to stop it.
- Art. 50
Outbound transparency
In force since 2 August 2026Where your agent communicates with people, they must be able to tell that AI is involved. This deadline was expressly not postponed.
- Art. 50(2)
Marking generated content
Applies from 2 December 2026A transition period for machine-readable marking of AI-generated content runs until 2 December 2026. Relevant as soon as your agent produces text, images or audio that gets published.
- Annex III
High-risk obligations
Applies from 2 December 2027If your agent has a hand in decisions about job applications, creditworthiness or access to services, the full obligations apply: risk management, data quality, documentation, human oversight under Art. 14 and logging under Art. 12.
- GDPR Art. 22
Automated decisions
In force since 25 May 2018Independent of the AI Act: decisions with legal effect must not be made by automation alone. That is exactly what part 06 is for, not as a convenience feature but as evidence.
This is an orientation, not legal advice. The official text of the regulation governs.
08 How we work
How we build a first agent.
Not a pilot that fizzles out. Four steps, and after each one you keep something, even if we stop there.
- 01
Watch the process
3 to 5 daysWe watch the process once, together with the people who run it. Afterwards it is on paper: which step takes how long, where decisions are made and what the agent must not touch.
What you have afterwardsA process map, an honest feasibility call and a cost range.
- 02
Build the smallest cut
2 to 3 weeksOne agent, one task, level L1. Tools with the smallest possible permissions, guardrails in code, logging from the start.
What you have afterwardsA running agent in your environment, delivering drafts for release.
- 03
Measure
4 weeks of operationFour weeks of real operation on real cases. We count closed cases, escalations, rework and hours saved, instead of debating impressions.
What you have afterwardsNumbers that turn the decision about the next level into more than a hunch.
- 04
Raise or extend
ongoingOnly once the numbers hold does it continue: raise the autonomy level or add the next task. Never both at once.
What you have afterwardsA system that grows with your experience, not with your nerve.
09 FAQ
Plainly answered.
What is the difference between an AI agent and a chatbot?
A chatbot answers questions, an agent completes tasks. It is given a goal, plans the steps, uses tools such as mail, ERP or the calendar and works until the goal is met or a human is needed. The difference shows in the outcome: a text on one side, a closed case on the other.
Do I need my own servers or my own model?
Usually not. For a first agent a good model behind an interface is enough, as long as the data allows it. If data must not leave the building it can run locally, which costs more preparation and a little quality. What you really need is access to your systems and somebody who decides exceptions.
What does an AI agent cost?
It comes down to two things: how many systems have to be connected and how cleanly they can be reached. Running it is usually the smallest item. That is why we start with a short look at the process and give a range before anyone writes a quote. The AI ROI calculator shows how to work the other side of the equation.
How long until an agent is live?
For a cleanly cut task it is two to three weeks to the first run at level L1, then four weeks of operation before we decide on more autonomy. When it takes longer it is almost always for one reason: access to the systems is not sorted.
What happens when the agent makes a mistake?
That is why we start at the level where a human releases the result. Add hard limits in code, a complete log and a kill switch everyone on the team knows. An error should surface, be reversible and be explainable from the log. Autonomy rises only once that holds for weeks.
Is an AI agent high-risk AI under the EU AI Act?
Not automatically. What counts is what it helps decide. An agent that writes quotes or confirms appointments generally falls outside. An agent that pre-sorts job applications or has a hand in creditworthiness does not. The Art. 50 transparency duty, by contrast, applies today as soon as people are writing to it.
Can several agents work together?
Technically yes, and for internal data work it makes sense. In practice we advise against starting there. Several agents multiply not just the work but the sources of error, and afterwards nobody can say which step produced the wrong number. Once one agent measurably holds, the second is worth it.
→ Contact
Describe one task. You get an honest answer.
Send us a process that costs your team hours every week. We will tell you whether an agent is worth it, which level we would start at and what range it sits in. If doing it by hand is cheaper, we say that too.
Describe your task →Read on
AI agents in the companyThe background article: the term, suitable tasks, the usual stumbling blocks.AI for automationWhat a process looks like from trigger to result, with examples.EU AI Act risk checkCheck for free which risk class your system falls into.AI ROI calculatorWork out when automating a task starts to pay.