Digital colleagues for work that needs a decision of its own.
An agent reads an enquiry, pulls the right information out of your documents, replies and closes the case. When it can't get further, it hands over to a human. Connected to email, chat, phone or your own systems.
At 9:40 in the evening an email comes in asking about a delivery date. The agent checks the inventory system, replies with the real date, files the case in the CRM and flags it for the next morning. The customer has their answer in two minutes instead of the following afternoon.
When the task is too varied for an automation and too frequent for a person.
An automation follows a fixed path: A comes in, B happens. An agent makes sense when the cases differ, the answer has to be pieced together and the whole thing still follows the same pattern a hundred times a month. It decides within the limits you set and hands over the moment it goes beyond them.
The same kind of question keeps coming in, but never in the same words.
The answer exists somewhere: in documents, in a system, in old emails.
Enquiries arrive in the evening, at weekends or faster than anyone can work through them.
A human should only see the cases that genuinely need a decision.
The agents I build most often.
USE CASESAnswer enquiries
The agent knows your products, prices and processes, replies in your tone and hands over the moment it needs to get personal.
Qualify enquiries
It asks the follow-up questions that would otherwise cost three emails and passes the case on fully prepared.
Take calls
A voice agent picks up when nobody can, notes the request, books appointments and sends the summary to your inbox.
Knowledge from your documents
Contracts, manuals, price lists: the team asks in chat and gets the answer with a source instead of a folder of PDFs.
Gather things independently
Check tenders, watch competitors, reconcile data sources and deliver the result as a short report.
Process cases
Read documents, check data, enter it into the system and only put the exceptions up for approval.
Your case isn't listed? What matters isn't the industry, it's whether the answer can be looked up somewhere. Describe the case briefly and you'll get an honest read.
What an agent delivers, on paper.
THE NUMBERSRun it with your numbers: enquiries × handling time × hourly rate, times the share the agent takes on. Not counted: the revenue from enquiries that would otherwise have gone unanswered.
The same enquiry, before and after
| STEP | BY HAND | WITH AN AGENT |
|---|---|---|
| Arrival | Sits until someone checks the inbox | Picked up instantly, around the clock |
| Understanding | Read the mail, find the context | Request recognised and classified |
| Finding the answer | Look through documents and systems | From the knowledge base and system, with a source |
| Replying | Write it yourself | In the agreed tone, directly or after approval |
| Edge case | Sometimes falls through the cracks | Goes to a human with a summary |
| Total per case | 6 min, office hours only | Under 1 min, any time |
How an agent you can trust gets built.
HOW IT RUNSSet the limits
We settle what the agent may decide, what it only prepares and when it hands over. Those limits matter more than any technology choice.
Build the knowledge
Your documents, prices and typical cases become the knowledge base. Whatever is missing shows up here and gets added instead of invented.
Run it in the shadows
At first the agent only answers internally. You read along and correct, and only once the quality holds does it face the outside world.
Go live and sharpen
After launch we look at the real cases. Weekly fine-tuning until handovers to humans are rare and warranted.
Models are replaceable, your data isn't.
I build agents so the language model stays swappable, because something changes in that field every few months. The knowledge base, the rules and the integrations stay with you. Where things are sensitive, the model and the data run on European servers or on your own.
How you can tell it's working
IN PRACTICE“My expectations were exceeded. I'm especially impressed by how great the booking tool turned out – and my customers love the website.”
The best agent is the one that knows when to hand over
The most common failure with agents isn't the wrong answer, it's the overconfident one. That's why I build the handover to a human first, and only then what the agent is allowed to do alone.
Talk through your case →The price is set before I start working.
First internal read-along, then real replies.
Knowledge base, prompts and code live in your accounts.
Any case can be handed to a human at any point.
What does an AI agent cost?
Fixed price, agreed upfront.
ALL PRICES EXCL. VATOne agent, one channel
from €1,900One clearly defined job, for example enquiries by email.
- Limits and escalation defined
- Knowledge base from your documents
- Connected to one channel
- Trial run with read-along
- Short introduction for the team
Agent with system access
from €4,900An agent that doesn't just reply but actually closes cases.
- Several channels, one body of knowledge
- Access to CRM, inventory or calendar
- Actions with approval levels
- Review of real cases after launch
- Documentation and training
Operations & training
from €290/monthFor everyone who doesn't want to sharpen the agent themselves.
- Monitoring and troubleshooting
- Keeping the knowledge base current
- Fine-tuning based on real cases
- Model changes when something better lands
- Cancel monthly
The price depends on how many channels the agent serves, how many systems it may touch and how much knowledge sits behind it. On top come running costs for models and hosting, which you get in writing before the start.
When I advise against an agent.
An agent is the most involved form of automation and not always the right one. In these four cases I'll say so upfront.
The process is always the same
If there's no decision to make, a classic automation is cheaper, faster and more reliable. Then that's what we build.
The knowledge doesn't exist anywhere
If the answers only live in two people's heads, they have to be written down first. That's work for your team, not a technology project.
A mistake would be expensive
In legally or financially delicate cases the agent only prepares. The final decision stays with a person, otherwise we leave it.
It's only a handful of cases a month
Building and maintaining an agent only pays off above a certain volume. Below that, a good template or snippet library serves you better.
AI agents for small and mid-sized companies
By Chris Pump, developer since 2021Last updated: August 2026
The word agent is currently used for everything from a chat window to a fully automated company. A narrower definition is more useful: an agent is a program that gets a goal, gathers the information it needs, uses tools and picks its own route to the result. That is exactly where both its strengths and its risks come from.
Agent or automation?
An automation is a fixed path, an agent picks its own. That usually makes the decision easy: as long as the process can be written down as if-then, it belongs in an automation. As soon as every case looks different and the answer has to be pieced together, an agent makes sense. Many projects end up with both: an automation hands the agent clean, prepared data and takes over the fixed steps afterwards.
Why limits matter more than the model
The question of which language model is best loses its meaning every few months, because the ranking keeps changing. What remains is the question of what the agent is allowed to do: only reply, or also book appointments, change data, move money? So every project starts with a list of permitted actions, an approval step for anything touching money or contracts, and a clean route to a human. An agent without those limits isn't a feature, it's a risk.
Hallucinations and how to contain them
Language models make things up when information is missing. Talking to them doesn't help, engineering does: the agent gets a knowledge base built from your real documents and answers with a source wherever possible. If it finds nothing, the permitted answer is an honest "I don't know" plus a handover. Before launch it runs in the shadows, so your team sees the quality on real cases before a customer does.
GDPR and running it
An agent usually sees personal data, so we settle early which data that is, where it gets processed and how long conversations are kept. Depending on the requirements, European providers, your own server or locally running models come into play. Transparency belongs to that too: people should be able to tell they're writing with an assistant, and reach a human at any time.
Common questions about AI agents.
What's the difference between an agent and a chatbot?
A classic chatbot follows a decision tree someone built in advance. An agent understands the request, gathers the information itself and can take actions in your systems. Both can sit in a chat window, but what's behind them is different.
How long until an agent is live?
Two to six weeks depending on scope. Building it goes quickly; the time sits in the trial run. Only once the answers hold up on real cases does the agent face the outside.
Does the agent make up answers?
Without safeguards, language models do. So the agent gets a knowledge base from your documents, answers with a source and is allowed to say "I don't know". If a question stays open it goes to a human instead of into an invented answer.
Will my customers notice they're writing with an AI?
They should, and that's required. The agent introduces itself as an assistant and always offers a route to a human. The tone still stays yours: how it addresses people, the wording and the limits are set together.
Can the agent do things, not just answer?
Yes. Book appointments, create CRM records, trigger orders, generate documents. For anything touching money or contracts I add an approval step so a person confirms.
What does running it cost?
Model costs are billed per request and land in the low three figures a month at typical volumes. On top come server or platform costs. You get an estimate for your volume in writing before the start.
What happens when the agent gets stuck?
It hands over to a human, with a summary of the conversation so far and a suggestion. That handover gets built first, because it decides whether an agent is usable day to day.
Can an agent handle phone calls?
Yes, voice agents take calls, note the request and book appointments. That's most useful outside office hours or as a safety net when nobody can pick up.
How do you make sure no data leaks?
The agent only gets access to the systems and fields it needs for its job. For sensitive data we use European providers or locally running models, and conversations are deleted after an agreed period.
Does the agent replace a job on the team?
In practice it takes on the repetitive part of the work and frees up room for the cases that need experience. People working in support handle fewer standard questions and more real requests.
Which languages does the agent answer in?
All the common ones. The knowledge base can stay in German while the agent answers in English. It's still worth setting and proofreading the tone once per language.
Do I own the agent?
Yes. Knowledge base, prompts, rules and code live in your accounts. You can keep working on it yourself or let someone else take over at any time.
Tell me which enquiries at your company always look the same.
30 minutes, free and without obligation. Afterwards you'll know whether an agent fits, where its limits are and what it costs.
Usually answered the same working day.