AI and automation
Our AI service connects language models such as ChatGPT and Claude to your documents, systems and workflow. We hand repetitive work like support, quotes, document reading or reporting to automation, and we write down what counts as success before we start.
- We connect the model to your own documents and systems.
- Success is defined on day one, with sample questions and expected answers.
- Model accounts and code are in your name from day one.
- Bug fixes are free for 30 days after launch.
Who is the AI service for?
It is for companies whose teams spend their time answering the same questions, reading documents and copying them somewhere else, or rebuilding the same report every week. The most common requests are an assistant that answers customers, a flow that reads invoices or contracts into your system, and a tool that drafts quotes for the sales team.
Not every job needs AI. A task with clear rules is often done cheaper and more reliably by plain automation. We tell you that on the first call.
What do we build?
First we decide what the model works from: your product catalogue, support history, contracts or internal documents. The model grounds its answers in that material and shows its source. This is called RAG; instead of making things up, the model learns to say it does not know.
Then we connect it where the work happens: your website, your panel or your messaging channel. Automations connect to your CRM, accounting software or e-mail, built the same way as our integrations: every step is logged, and you hear about it when something fails.
The job decides the model. Some jobs need a fast, cheap model, others a stronger one. Model accounts are opened in your name, and switching models later does not mean rewriting the code.
How do you know the AI works?
Before we start, we build an evaluation set of sample questions and expected answers. The definition of done is written against it: how many it must get right, and when it must hand over to a person. The set runs again after every change, so the result is a number, not a feeling.
Risky steps get a human approval step. The model drafts, your team decides. Personal data is masked before it reaches the model where needed, and which data goes where is written down.
What is included
-
An assistant on your own data
Answers from your documents and records, and shows its source.
-
Document reading
Reads invoices, contracts or forms and enters the data into your system.
-
Workflow automation
Runs repetitive steps between your CRM, e-mail and accounting on its own.
-
Evaluation set
Sample questions and expected answers show, as a number, how well the model works.
-
Human approval
On risky steps the model drafts and your team makes the final call.
-
Cost tracking
Model usage and monthly cost are visible in the panel, with no surprise bills.
How the project runs
- 01 You write, we reply within a business day Day 0 you tell us the work you want to hand over. Day 1 we reply. Day 2 we talk for 30 minutes about the job and your data.
- 02 One-page quote and definition of done Day 4 you get the quote. Day 5 the definition of done is signed, including the result the evaluation set must reach.
- 03 A small trial We first build a trial on part of your real data. If the result is not good enough, you can stop before the larger project.
- 04 Build and measure We build the connections and rerun the evaluation set after every change. You see the results too.
- 05 Go live and handover We go live on the promised date, watch the first weeks closely and hand over notes and logins.
What happens after launch?
We fix bugs for free for 30 days. Language models change quickly; when a new one comes out, we run the evaluation set on it and suggest a switch if it is better or cheaper. Updates like this are part of optional monthly maintenance and support. The code and model accounts are already yours. If you want to continue with another team, an e-mail is enough.
What you get at delivery
Each of these is written into the definition of done. If one is missing, the work does not count as delivered.
- The AI feature or automation written in the definition of done
- An evaluation set of sample questions and expected answers
- Usage and cost tracking
- A human approval step for risky or uncertain answers
- Code repository and model accounts in your name
- All access details handed over in writing
Frequently asked questions
01
Will our data be used to train the model?
On business API accounts, model providers do not use submitted data for training by default. The quote says which provider runs under which terms, and personal data is masked before it is sent where needed.
02
What if the AI gives a wrong answer?
The model grounds its answers in your documents and hands over to a person when it is unsure. On risky steps your team makes the final call. Accuracy is tracked as a number with the evaluation set.
03
How much will the model cost per month?
It depends on usage and the model chosen. After the trial we estimate the monthly cost from real usage, and you can follow it in the panel.
04
Which models do you use?
We pick per job. OpenAI and Anthropic (Claude) models are the ones we use most. We build it so that switching models does not mean rewriting the code.
05
Can it connect to our existing system?
If the system offers an API, database access or file exchange, usually yes. We check together on the first call.
06
Who owns the code and accounts?
You do. The code repository, server and model accounts are opened in your name from day one.