Which task do you want off your desk?
I build AI agents that take on one specific job. You set the rules, the agent does the reading and checking, and a person approves what matters.
- Tasks
- Customer questions, quotes, invoices, applications, tenders
- Control
- A person approves what matters
- Data
- Processed on servers in the EU
- Regulation
- Built for the AI Act
01 · Three steps
One step at a time.
You see the value of one task before building on it.
- Step 1
Mapping
I look at the workflow and the systems you use.
- A written proposal with the tasks that pay off
- What it costs to run, and what is not worth it
When you know AI can help, but not where.
- Step 2
First agent
One task. The agent first works beside you and suggests, before it may answer on its own.
- An agent in production, measured against how the job is done today
- You see it work before you build on it
When one task happens many times.
- Step 3
Running and training
I keep the agent running and teach you to change the rules yourselves.
- A solution you understand and own
- No lock-in
When the first agent is running and the next task is up.
What could an agent do for you?
Almost anything that is reading, checking and writing by rules someone has already decided. Pick an area.
- 01
Answers “where is my parcel?” around the clock.
Looks up the order and the tracking and replies with the right status, even at three in the morning.
Uses
- Orders
- Shipping
- Chat
- 02
Sorts the inbox before you arrive.
Reads everything that came in, labels what it is about and drafts a reply. You approve and send.
Uses
- Help pages
- 03
Catches the unhappy ones.
Recognises a disappointed customer and sends the case straight to a person, before it becomes a bad review.
Uses
- Customer history
Do not see yours? The best ideas often come from people who know their own working day. Tell me about yours
02 · In depth · How an agent is built
Tools, rules and a person who approves.
A language model on its own answers with confidence, even when it is wrong. What sits around the model is what makes an agent trustworthy.
Tools instead of memory
The agent does not answer from whatever the model happens to remember. It looks things up in your systems, the registry or the regulations, and shows where the answer comes from.
Rules you decide
What it may do alone, what it must ask about, and what it must never touch are written as rules you can read and change.
Measured before it is let loose
The agent first runs on real cases beside you. Its suggestions are compared with what you would have done, and it does not answer on its own until the hit rate holds.
Unsure means ask
When the basis is thin, the case goes to a person instead of the agent guessing.
A cap on spend
Limits per user and per day, so a bug or misuse does not turn into a bill.
A log of everything
What it read, which rule it followed and what it did. You look it up instead of guessing.
03 · In depth · Data and regulation
What you want to know about your data.
Servers in the EU
Your data is processed on servers in the EU.
You know who
Providers are chosen for each task, and you are told which ones.
No training
I choose providers that do not train models on your data.
Only what it needs
The agent gets access to one inbox or one registry, not the whole system.
A person approves
Money, rejections and anything it is unsure about go to a person.
Says it is a machine
The AI Act requires chatbots to make that clear. The agents I build do.
04 · Limits
What I say no to.
Not everything should be done by an agent. When that is the case, I say so before you spend money on it.
Tasks that rarely happen
If it happens once a month, it is rarely worth building an agent for it.
Decisions about people
Rejecting applications, hiring and credit are decided by a person. The agent prepares.
What a better routine solves
Sometimes the answer is a form or a simple rule, not AI.
Answers that are not written down
If the knowledge is not written anywhere, the agent cannot find it. Then that is where we start.
05 · Who you talk to
You talk to me all the way.
I map the work, build it and train your team. No salesperson in between and no handover to another team. Link Utvikling, Bergen, Norway.
What people ask before they start.
What is the difference between a chatbot and an AI agent?
A chatbot answers questions. An AI agent carries out a task: it uses tools such as an inbox, a registry or a set of rules to read, check and draft, step by step, and asks a person when it is unsure.
Can't we just use ChatGPT?
For a lot of things, yes. ChatGPT is good when one person works on one text. An agent is built for one task that happens many times: it pulls from your systems, follows your rules and writes down what it did. If ChatGPT is enough, I will tell you.
Our data is messy. Does it still work?
Often, yes. Agents are good at reading unstructured text such as email and PDFs. The mapping shows what, if anything, needs tidying first.
How much of our time does it take?
Most of it at the start: a conversation about how the job is done today, and a few real examples. While the agent works beside you, a person looks over its suggestions. After that you change the rules yourselves when something changes.
Is it safe to let AI near our data?
It can be done safely. Data is processed on servers in the EU, the agent only gets access to what it needs, everything it does is written down, and decisions that matter are made by a person. I choose providers that do not train models on your data.
Do we have to follow the AI Act?
The EU rules for chatbots apply from 2 August 2026. For most companies the key rule is that people must know they are talking to a machine. High-risk uses, such as recruitment and credit scoring, get stricter requirements.
Which task should we start with?
Tell me how the task is done today. You get a reply within a day, with an honest answer on whether AI fits. Also when the answer is no.