The problem
Chat windows and isolated automations usually know only their current task. Company work is made of cases that wait, absorb new evidence, require approval and need to be checked again later.
The approach
I built a shared work layer for recurring operational cases. It connects email, documents, tasks and selected ERP processes. Agents handle bounded steps while state, evidence and the next check remain durable.
An input is first matched to an existing or new case. The workflow reads the relevant sources, applies fixed rules and uses a model only for the parts that require interpretation.
- One identified case instead of a loose sequence of chats
- Explicit states for open, WAITING, human decision and complete
- Model routing by task instead of one model for everything
- Approval before irreversible or external actions
- Escalation and a next-check date for every open case
What this demonstrates
The prompt is rarely the hardest part. Case identity, current state, permissions, evidence and a clear recovery path matter more when the system touches real work.
Where it applies
The model fits SMEs whose work is split across inboxes, files, ERP and personal knowledge. The value does not come from maximum autonomy. It comes from less searching, clearer decisions and cases that are followed through reliably.
Not every process should be automated. Rare cases, unclear ownership or poor master data are often an organisational problem first.