An agent for the operations inbox
A services company loses hours to a shared inbox. We put an agent in its environment that reads the mail, looks up the account and drafts the reply. A person still approves anything that commits the company.
The problem
Requests arrive by email in a shared inbox. A senior person sorts each one, looks up the account and writes a reply the company has written many times before. When that person is away, the queue waits.
The work is repetitive, and the cost of a wrong reply is real: a price, a date or a promise to a customer.
What we build
An agent that runs in the client’s environment.
- It reads new mail from the shared inbox.
- It looks up the account in the tools the company already uses.
- It drafts the reply, or opens the internal task, using the company’s own procedure.
- A person approves anything that changes money, a contract or a commitment to a customer. Routine drafts can be sent under a rule the client sets.
- Every run is logged. Before go-live, the agent is scored on a set of past requests the client has labelled.
How it runs
- The agent is connected to the mailbox and the account system through tools, with the client’s credentials, in the client’s cloud or network.
- A person stays in control of the decisions that matter. The agent does the lookup and the draft.
- Failures are reviewed on a schedule the client chooses. A bad draft becomes a new evaluation case.
What the client gets
- A working agent in their environment, connected to their systems.
- An evaluation set, and a short written record of how the agent behaved on it.
- Monitoring, so they can see what the agent did.
- Guardrails: which actions need a person, which data the agent may read and when it must stop and ask.
The result
People review a draft that already has the account context, and the queue moves on the day the mail arrives. The client owns the agent, the prompts, the tool connections and the evaluation set.