AI agents and automations

What problems it solves

Every business has work that needs no judgement but takes hours: copying lines off an invoice, reading a tracking number from a label, sorting the inbox, writing the same description in five languages, moving tasks from the chat into a list. People do it because nobody else will.

Language models are good at exactly this kind of work: reading, classifying, drafting. Their weak spot is that they are sometimes confidently wrong. So we do not let them act on their own. The model prepares the work, a person says yes, and only then is anything saved, sent or published.

The agent proposes, a person confirms

There is a check between the model's decision and the action. Anything that only reads data, such as orders, stock or reports, goes through straight away. Anything that sends, spends, deletes or changes a setting stops and waits for explicit confirmation, separately for each action. One confirmation never applies to the next action. If nobody answers in time, nothing happens.

The same rule applies everywhere: in chat, in voice conversations and in background checks. An ad goes live with a button, a description is saved with a button, a parsed invoice is booked in with a button. There is always one moment on screen where a person made the decision, and it stays in the history.

How it works

Every automation goes through the same steps.

1. The input arrives on its own or with a single action: an uploaded PDF, a new email, a photo, a chat question, a scheduled job.
2. The model reads and prepares: it extracts lines, fills in numbers, writes the draft and proposes what to do.
3. Anything that can be calculated exactly is calculated in code: totals, stock, SKU matching. The model receives the finished figures and only judges the meaning, without inventing them.
4. Before anything is saved, the screen shows what was read and what it was matched to.
5. The person corrects it if needed and presses the button. Anything typed by hand always takes priority over what was read.
6. The system remembers a confirmed match. Next time the same line from the same supplier is matched automatically.

Which models we use

We use Anthropic Claude for writing, analysis and agents with tools, and Google Gemini for reading PDFs and photos, real-time voice and fast sorting. For each task we pick the model that does it best at a reasonable cost. We do not give light tasks to the most expensive model.

When data must not leave your infrastructure, a local model running on your own server can be used. Documents and conversations are then not sent to any outside service.

What you get

• Invoices, PDF orders, labels and photos that the system reads and matches to the catalogue, without anyone retyping them.
• Product descriptions with the same structure for every product, adapted to each store and language.
• An inbox where every email arrives with a priority, category, language and short summary.
• An assistant that answers questions about orders and tasks in chat or by voice.
• An advertising agent that proposes campaigns and copy per market, while launching stays a button press.
• A daily review of what is growing, what is falling and where money is spent without results.
• A cap on AI spending, after which the agents stop.

What it connects to

The agents work on the same data as the rest of the system, with no separate copy.

• Shopify and WooCommerce: products, orders, descriptions.
• Meta and Google Ads: campaign results, ad copy.
• Mailboxes: incoming emails and attachments.
• The team chat: tasks and instructions in plain language.
• Warehouse and suppliers: invoices, goods-in, catalogue.