// ai_agents · automation
Custom AI agents that actually do the work
An AI agent is a program that can read context, decide what to do next, call your tools, and write the result back to your systems. Not a chatbot that answers and forgets. A real worker for the boring, repetitive parts of a business: triaging support tickets, searching internal docs, qualifying leads, moving a workflow forward step by step.
We build these from scratch, scoped to one job at a time and wired into the tools you already run. We build on Claude and on open-weights models, picking whichever fits the task and your data rules. We do not use OpenAI or GPT. When a job genuinely needs a frontier model we say so; when a smaller open model on your own hardware is the right call, we say that too.
We are a small engineering shop in Jönköping working with companies across Sweden and the Nordics. Direct, honest, no enterprise theatre. If an agent is the wrong tool for your problem, we will tell you before you spend money.
What a custom AI agent actually is
There is a lot of noise around the word "agent". Here is the plain version. An agent takes a goal, breaks it into steps, and uses tools to get there. It can read a ticket, look something up in your database, draft a reply, and update the record, all in one run. A plain chatbot can only talk.
Ours are built for one job and shaped to your setup, not bought off a shelf. They run against your industry, your size, and the way you actually work. They handle several steps and tools in a single workflow instead of stopping at the first answer. They get sharper as you correct them and feed them real examples. They plug into the tools you already have rather than forcing you onto a new platform. And they run when nobody is around, which is the whole point of automating the night shift of admin.
None of that is magic. It is software with a language model in the loop, built carefully and tested against your real cases.
What you get out of it
The point of an agent is to take work off people, not to add a gadget. Where it earns its keep:
Less repetitive work
The routine, rule-based tasks get handled automatically, so your team spends its hours on the work that actually needs a person.
Support that answers right away
An agent can field common questions, walk a customer through a process, and hand the hard cases to a human with the context already gathered.
Answers from your own data
Point an agent at your documents and records and it can find the answer, summarise it, and cite where it came from, instead of someone digging through folders.
Scales without a new headcount
When volume doubles, the agent handles more of the same without a proportional jump in cost. You pay for compute, not for another desk.
Where we put them to work
Concrete jobs we have built agents for, not a wishlist:
Support. An agent that reads incoming tickets, answers the common ones, troubleshoots known issues, and escalates the rest to a person with the relevant history already pulled together. It takes the repetitive volume so your support people handle the cases that need judgement.
Internal knowledge. An agent that indexes your handbooks, wikis, and past tickets and answers staff questions from them directly, with a link to the source. Instead of someone reading through documentation for twenty minutes, they ask and get the answer.
Leads and sales. An agent that watches incoming interest, qualifies it against your criteria, enriches the record, and flags the prospects worth a human call. It does the first-pass filtering so your salespeople spend time on real opportunities.
Workflows. An agent that moves a multi-step process forward across systems: read a form, check a rule, update a record, trigger the next step, notify the right person. The kind of approval chain that used to take days can run in minutes when the routine decisions are handled automatically.
The tools we build on
We are honest about the stack because it matters for your data and your bill.
We build on Claude from Anthropic as our main model, and on open-weights models you can host yourself when the data has to stay on your own hardware. We pick per task: the strongest available model where the reasoning is hard, a smaller efficient one where it is plenty. We do not use OpenAI or GPT.
Around the model we use proper retrieval over your documents so an agent answers from your real content, and orchestration tooling like LangChain and n8n to wire the steps together. Everything connects to your existing tools, APIs, databases, and channels rather than asking you to migrate. The agent fits into your stack; you do not rebuild your stack around the agent.
How we build it, and your data
An agent project is normal engineering, not a research experiment. We start by scoping one job: what it should do, what it reads, what it writes, where a human stays in the loop. We build that, test it against your real cases, and put it in front of people before we widen the scope. Smaller and working beats big and flaky.
On data: your inputs and documents are yours. We build so they are used to do the job and nothing else, never to train a shared public model. Where the data is sensitive or regulated we can run the whole thing on your own hardware or on EU infrastructure we operate, so nothing leaves your boundary. We work to common standards like GDPR by default, and to others such as SOC 2 or HIPAA when your situation requires them, scoped honestly to what each project actually needs.
We keep it low-maintenance and support it after launch. As your business changes the agent gets refined and extended. And we stay straight with you the whole way about where it helps and where a person is still the better tool.
// faq
Frequently asked questions
How long does it take to build a custom AI agent?
It depends on scope. A focused agent doing one well-defined job can be live in a few weeks. Something that spans several systems and needs careful integration takes longer, a couple of months is realistic. We would rather ship a narrow agent that works and expand it than promise a sprawling one on a date we cannot hold.
Will the agent work with the tools we already use?
That is the default. We build agents to connect to your existing tools, APIs, databases, and channels rather than moving you onto a new platform. If something exposes an API or a sensible integration point, we can usually wire an agent into it. We check the specifics against your stack during scoping.
Is our data safe when we use an AI agent?
Your data stays yours. We build so your inputs and documents are used to do the job and never to train a shared public model. When the data is sensitive or regulated we can run the agent entirely on your own hardware or on EU infrastructure we operate, so nothing leaves your boundary. We work to GDPR by default and to standards like SOC 2 or HIPAA when your situation calls for them.
Do you use OpenAI or GPT?
No. We build on Claude from Anthropic and on open-weights models you can host yourself, picking whichever fits the task and your data rules. We are happy to explain why, and to compare honestly against whatever you are using today.
How much maintenance does an agent need?
Not much day to day, but it is not fire-and-forget either. We provide support and updates after launch, and the agent gets refined as your processes change and as you feed it more real examples. Plan for it as a system you keep improving, not a one-off install.
How do we know it is worth the money?
We will not throw fabricated ROI percentages at you. The honest answer: the value comes from hours of repetitive work removed, faster response on the routine cases, and capacity that scales without new hires. We scope a first job small enough to prove out quickly, so you can judge it on your own numbers before committing further.
Got a repetitive job an agent should be doing?
Tell us which process is eating your hours and we will scope an agent that does that job, wired into the tools you already run. If an agent is the wrong tool, we will say so.