01The Problem
One person carries most of the cooperative's admin. She handles funding and grant applications, with around 18 in flight at any time. She also follows up with the accountants, onboards new members and drafts documents.
On top of that sits a heavy email load. Email is the biggest pain. Between the inbox and the funding deadlines, the week is gone before the real work starts.
02What the Assistant Does on a Normal Day
She talks to the back-office assistant in a private chat channel, like she would with a colleague. It works in the Google Sheet she already uses for her tasks. She likes spreadsheets, so there was no new app to learn.
- Finds funding calls. It tracks funding and grant calls that are opening and stores them for her.
- Keeps the task list up to date. It adds, updates and closes rows in her own task sheet when she asks.
- Answers questions. It looks things up in the cooperative's knowledge base, so she doesn't have to dig.
- Triages email. Every 30 minutes during working hours it goes through new email, sorts what matters and drafts replies for her to approve.
- Sends a morning briefing. At 09:00 every day she gets a short summary: deadlines, open tasks and what came in.
- Runs a funding radar. Every Monday it sums up the funding calls that are open or about to open.
- Drafts, on request. It prepares replies and documents for her to check, edit and send.
Other members use the same assistant through a web chat inside the cooperative's internal tool. So it isn't tied to one chat app.
03Watch It Run
This is a replay of a typical morning. Press play. Names and content are made up, but the behaviour matches the real assistant.
04Safety by Design
This is the part that matters most. The assistant does the busywork. A person decides anything that leaves the building.
The hard rule: it drafts, but it never sends an email, makes a payment or submits anything to an official registry without her explicit confirmation in chat.
- A human approves every outside action. Drafts wait in the chat until she says yes. No yes, no send.
- The web chat is read-and-prepare only. For other members it reads, searches, summarises and prepares. It only acts when someone writes to it, and it treats anything with outside effect as not authorised.
- Everything is logged. What it did, when and why can be checked later.
05Under the Hood
- Built on my own agent framework: a Claude agent (Claude Sonnet) written in Python, running on a Linux server.
- One isolated service per assistant: each assistant runs on its own, so a problem in one can't take down the others.
- Conversation memory: it keeps context across sessions and remembers earlier decisions.
- Cost-aware triggers: the model only wakes when there is work to do, so it doesn't spend money idling.
- Works with existing tools: Google Sheets for tasks, a private chat channel for the operations lead and a web chat for everyone else.
06What Came Out of It
- Funding calls get found and stored without her having to go looking.
- Her task sheet stays current, and she never had to switch to a new tool.
- Questions about the cooperative get answered from the knowledge base, without a search through old files.
- Nothing goes out without her say-so, so she can hand off the busywork without worrying.
Want an assistant like this for your team?
I build assistants that do the admin while a person approves anything that matters.