Designing an executive email assistant that drafts in her voice and never hits send
An executive drowning in a mailbox and a calendar didn't want an assistant to take over, she wanted one she could still trust completely. The answer wasn't a smarter AI. It was making sure the send button stayed out of its reach, by design, not by promise.
Category
Workflow automation with a human kept firmly at the controls.
What I was trying to accomplish
An assistant that fit an executive's existing habit of working from a running checklist: sort the day's mail and meetings into one list, draft replies that sound like her, and leave every send decision with her.
The data challenges
Building a voice guide from 25 to 30 of her own sent emails, a standard-answers file that grows as recurring questions get settled, and a staged rollout, copy-and-paste drafting first, then the live mailbox and calendar, then the shared mailboxes she delegates. Figures pulled from the property management system, which isn't a live connection, had to be labeled with an as-of date rather than presented as current.
The result
A complete, rebuildable project package, instructions, saved prompts, voice guide, standard-answers templates, rollout plan, and guardrails, built so the client can recreate it in their own environment without kapchur consulting. Send permission stays off at the connector level, so review is enforced by configuration, not by a promise.
Technical details
Tools: Claude Projects, Microsoft 365 mailbox and calendar connector, saved prompts
A company executive was carrying a heavy inbox, a full calendar, and a steady stream of repeat questions from owners and residents. She already worked from a running checklist. The goal was an assistant that fits that habit.
The design has three inputs: her own mailbox and calendar through the Microsoft 365 connector, a voice guide, and a standard-answers file. It lives in a dedicated Claude Project with custom instructions and three saved prompts: draft a reply in her voice, build the morning checklist, and prepare for a single meeting. Before anything touches her account, a smoke test is specified to run in a company account: read the inbox, draft a reply without sending, and check that sending is switched off.
The main design challenge was trust. Send permission stays off at the connector level, so human review is enforced by configuration rather than by a promise. When the assistant cannot answer from what it can see, it is instructed to say what is missing instead of guessing.
Pull up a chair.
Let's talk through where things stand, and where they could go.