Sous: an AI system on its own computer that does your written work, learns from your corrections and keeps what you know, run for you by Serene Labs
If your week is full of messages, reports and replies you write again and again, Sous drafts them for you. It is your own AI system, on its own computer on our servers, in your own dedicated environment, on your own Claude account. A separate agent does each job you hand it. They work in a browser and in the accounts we connect, keep a library of what you know, learn from your corrections and adapt as your work changes. When you want it to work differently, it writes the change and shows it to you first. Nothing goes out until you approve it. Jacob Ralph is the one accountable lead.
Sous is new, and we have no client outcome to show for it yet. The pattern it uses runs in our own system: in the 7 days to 2026-10-22, 639 asks were recorded and only 15 of them at some point waited on the owner personally. In the same 7 days its reviewer read 316 drafts and sent back or escalated 183. Those are our own system's counts, not Sous's and not a client's.
What changes in your week
- The work you repeat arrives drafted in your Teams chat, and a reviewer reads each draft before you do.
- Your corrections become lessons. Each kind of work earns more trust as approvals come back clean.
- You stop writing the same reply twice, and you keep the final say on everything that leaves.
- Email stays in your hands: a card opens the draft in Outlook and you send it.
- You do not install, host or update anything. We do.
What Sous does in a week, step by step
- An email you owe an answer arrived. The agent for that job read it against your library and drafted the reply. The reviewer read the draft. You saw one card in Teams with the draft on it. You approved, the draft opened in Outlook, and you sent it yourself.
- You changed a line in a draft. Sous turned your edit into a lesson in its daily playbook. You saw nothing more. The next draft of that kind followed the lesson.
- You asked it to work differently. Its builder wrote the change and tested it in a sandbox. You saw the change and approved it. It applied only then, and "undo the last change" rolls it back.
- You handed it a question to research. It looked things up in a browser, drafted the findings and had the reviewer read them. You saw the report in Teams and approved it.
What we set up
A short written map of the work you repeat, then your Sous connected to your Teams chat and your own Claude account, the first agents built from your own examples, and a supervised first stretch until each drafts to your standard. We need Microsoft Teams and Outlook, examples of your real work and your written sign-off. A small team gets one Sous per person, each with their own approver, data and Claude account.
Not included: the code, which stays with us. Coming, and not offered today: Slack, and one shared library for several people.
You need a Claude account in your own name, on an API key or a Team or Enterprise plan, never a personal subscription. We confirm in writing that its terms allow automated use. Anthropic's own consumer terms bar automated access except through an API key (vendor figure, Anthropic).
Price and terms
Pricing on request. The setup has a fixed price, agreed in writing before work starts and paid when you accept it against its test. A sample of your real tasks runs through Sous for that test. Extra agents are fixed add-on pieces, and the managed run is recurring.
If anything fails the agreed test and is found within 30 days of acceptance, we fix it at no extra charge.
We work async-first and in writing, so you keep a clear record and can answer when it suits you. A call is yours on request.
Proof
Sous is new, and we show no client outcome for it yet. What we show is the control, that nothing leaves until you approve it, and the terms above.
The same pattern runs in our own system, which we run our own business on. The counts at the top are our own system's, not Sous's and not a client's. From 2026-10-14 to 2026-10-22 08:00Z, of 542 scored approvals, 481 were approved untouched.
Our lead, Jacob Ralph, built a multi-model system where supervisor models review what the worker models produce before a user sees it.
Next step
Reply with the messages or reports that take most of your week, and we send your setup scope in writing, or book a call: https://calendly.com/serenelabs/30-minute-discovery-call