AI assistants propose,people keep the final say
Assistants for company e-mail, for WhatsApp and for workflows with written rules. They run on our own server, work with your real data and never send, delete or cancel anything important without a person's approval.
Three assistants, already at work
Each one runs today in one of our projects. We don't promise more than we have seen working.
E-mail assistant
For companies with several accounts and a lot of noise in the inbox.
- hourly summary, with a phone notification
- deadlines added to the calendar automatically
- draft replies, sent only after approval
WhatsApp assistant
For stores that get questions about prices and stock.
- searches products, prices and availability in the real catalogue
- records quote requests, with an ANAF (Romanian tax authority) check and a draft quote
- hands the conversation to a person and pauses while the person replies
Procedure-based operator
For repetitive workflows described in a document.
- handles a system's events (orders, e-mails, deadlines)
- fills in records and logs every step in the history
- sensitive decisions, such as cancellations, stay with a person
Message to action, human in the loop
The same pattern in every project: the rules are written down, and a person takes the final step.
What the assistant won't do
- it doesn't send e-mails without a person's confirmation
- it doesn't delete permanently; what it discards goes to Trash, where it can be recovered
- it doesn't touch e-mails from customers, banks or institutions, or invoices
- it doesn't make up prices or deadlines in drafts
- it doesn't enter passwords and keys; a person does that
Where the data lives
E-mails, the catalogue, the rules and the history stay on our own server. For each analysis, the necessary text is sent to the Claude model through the API, and the response comes back to the server.
Every run leaves a trace: in our e-mail assistant, 104 summaries written by Claude between 30 Oct and 7 Nov 2026 are saved in the notification archive.
You can read more about the way we work on the how we work with Claude page.
Where it already runs
AI e-mail assistant
8 accounts in one place, 11,567 unwanted messages removed in 2 days, an hourly AI summary.
Automated B2B store
WhatsApp Business assistant since 30 Oct 2026 and an operator following a written procedure for orders.
E-commerce automation
The store's complete flow, of which the assistant is only one part.
What clients ask us
Does the assistant reply to customers on its own?
By e-mail, no: it writes the draft, and a person approves it. On WhatsApp it answers product questions from the real catalogue and records quote requests; when the conversation needs a person, it hands it over and pauses while the person replies.
Which AI model do you use?
Claude, developed by Anthropic, called through the API. We use the name descriptively; we are not an Anthropic partner.
Does company data leave the server?
The data stays on our own server. The text needed for an analysis (an e-mail, a question, a catalogue excerpt) is sent to the model through the API with each request. Passwords and keys never reach the assistant; a person enters them.
What if the assistant makes a mistake?
That is why the final step stays with a person. Discarded e-mails go to Trash and can be recovered, and each recovery teaches the filter. Drafts can be edited as many times as needed before sending.
How much does an assistant consume?
It depends on how often it runs. In our e-mail assistant we reduced WhatsApp reading to 4 times a day, for the calendar only, to limit consumption. We agree on the frequency together, before any work starts.
Have a workflow that repeats every day?
Tell us where the messages come from and which decisions need to be made. We'll show you what an assistant can prepare and what stays with you.