AI automation that removes repetitive work, not control
A language model on its own does not solve a process. It has to be connected to the right data, given clear rules, and stopped at a defined point so a person can step in. That is the work we do, on the tasks your team repeats every week.
AI automation is a repetitive business process run by software that uses artificial intelligence models for the parts requiring interpretation of text or documents. Unlike traditional automation it does not follow fixed rules alone, so its output has to be reviewed by whoever remains accountable for the process.
The problem we solve
The conversation about AI in business has split between people expecting whole departments to disappear and people treating it as a passing fashion. Day to day the reality is duller: there are repetitive tasks made of reading documents, rewriting text, cross-checking figures and moving data from one system to another, and they eat hours every week.
On tasks like these a language model works well, under two conditions. First, it needs access to the right data: without the company's own information it produces generic or invented answers. Second, the output needs a human check whenever an error has real consequences, such as a quote sent to a client or a figure landing in the accounts.
The usual mistake is to start with the most complex and most visible process, the one that sounds impressive in a meeting. The opposite works better: pick a narrow task, repeated often, with an output that is easy to judge right or wrong. If it works there, the rest of the organisation adopts it without being persuaded.
How we work
Useful automation starts from a process someone already performs by hand today. We begin there, measure it, and automate only the part that survives a review.
Choosing the process
We look at the repetitive tasks together and pick one with high volume and clear criteria for a correct result. If nobody can say when an output is good, that process is not ready to be automated yet, and it is better to say so upfront.
Access to company data
We connect the automation to the sources it needs: business systems, document archives, spreadsheets, mailboxes. We define which data leaves the company and which does not, because that is the question that carries the most weight once third-party models are involved.
Prototype and evaluation
We build a first version and run it against real cases that are already closed, where the outcome is known. Comparing the output with what was actually decided is the only honest way to know whether the automation holds before anyone relies on it.
Human oversight and rollout
We decide where the flow pauses for approval and where it can continue on its own. Orchestration runs on n8n or purpose-built code, integrated into existing systems through APIs and webhooks, without asking people to switch tools.
Monitoring and cost
We track errors, run times and model API spend. When a cheaper model is enough, we switch; when the process changes, instructions and rules are updated. An automation left unattended degrades within a few months.
Technologies
The model is only one piece. What decides the result is how data is gathered, how steps are orchestrated and how decisions are recorded.
Models
- Claude
- GPT
- Gemini
- Open source models on dedicated servers
Orchestration
- n8n
- Webhooks
- Scheduled jobs
- Serverless functions
Data and context
- PostgreSQL
- Vector search
- PDF extraction
- Business system APIs
Integrations
- Google Workspace
- Stripe
- Postiz
- Email and chat notifications
Automations in service
Three processes that now run without a person handling every single step.
Allestio · AI assistant inside the system
An assistant built into the management system for audio-video contractors: it reads the request, proposes materials and cost lines during quoting, and leaves the operator the final call on quantities and prices before anything reaches the client.
CookedThumb · Visual content generation
A web app that produces professional video thumbnails in batches from a short brief, using Gemini for image generation. Built for people publishing regularly who do not want to redo the same design work for every upload.
MAXYMIZE social publishing · Internal infrastructure
A system that prepares and distributes content across social channels through a self-hosted Postiz instance. Posts start from project release notes and stay as drafts until they are approved.
Frequently asked questions
How much does it cost to automate a process with AI?
There are two lines: building the automation, and the recurring spend on model API calls. A first narrow workflow starts in the low thousands of euros, while monthly usage depends on volume and is usually the lighter of the two figures.
How long until we see the first result?
An automation covering a single process can be prototyped in a few weeks and judged straight away against closed cases. From that review to production usually takes four to eight weeks, the time needed to integrate it and set the checkpoints.
Will our company data be used to train the models?
On the business plans of the major providers, data sent through the API is not used for training. Where information is sensitive we look at anonymising critical fields, or at models running on a server the client controls, so nothing leaves the perimeter.
What happens when the AI gets it wrong?
It will, and the project has to be built on that assumption. We define upfront where a person must approve, what gets logged so any decision can be traced, and which thresholds halt the flow. Where the outcome is contractual, a human keeps the final word.
Can it work with the system we already use?
In most cases yes, through APIs or webhooks. When a system exposes no interface, we work with its scheduled exports or the files it produces. We check feasibility before quoting, because this is where projects of this kind get complicated.
Let us find what is worth automating
Thirty minutes on your repetitive processes to work out which one to automate first, what it takes to make it reliable, and what is better left to people.