Private AI as an executive assistant for scheduling and correspondence
A lot of executive-assistant work is language: turning "find 45 minutes with the board chair before the 14th, avoid Tuesdays" into a clean set of proposed times and a polite email, or taking a rambling voice memo and producing a correspondence draft that sounds like the executive wrote it. Cloud AI scheduling assistants do this well, but they typically need read access to a calendar and inbox to do it, which means every meeting title, attendee list, and email thread involved passes through that vendor. For an executive whose calendar includes board prep, M&A diligence calls, or compensation discussions, that's a meaningful amount of exposure for what is, at its core, a drafting task.
What's actually at stake
A calendar entry titled "Board comp committee: CEO pay review" or "Call w/ [acquirer] counsel re: diligence" tells a reader a lot even without opening the invite. A scheduling assistant that ingests a full calendar to find open slots and draft invites necessarily sees all of that, plus the attendee list, which can itself reveal who's involved in a still-confidential process before it's meant to be known. Correspondence drafting is similar: an email thread being summarized or drafted around a sensitive personnel matter or a not-yet-announced deal carries the same content risk as the document itself, because the AI tool now has a copy of it.
None of this means cloud scheduling tools are reckless. It means the tradeoff is worth making deliberately for a general assistant, and worth reconsidering specifically for the subset of an executive's calendar and correspondence that touches board matters, M&A, or anything under an NDA.
What self-hosting changes
Running the assistant model on infrastructure you control means calendar data and draft correspondence stay inside the environment the executive's office already operates in, rather than passing through a third-party scheduling product's servers. This doesn't require self-hosting your entire calendar system, most teams keep using their existing calendar and mail platform, it means the AI component that reads that data to draft invites and messages runs on hardware you control instead of a vendor's.
Where it's genuinely useful
The strongest use is turning unstructured input into structured drafts: a voice memo or a scribbled note becomes a clean meeting invite with a proposed agenda, or a set of bullet points becomes a properly formatted email in the executive's usual tone. It's also good at first-pass triage, summarizing a long email thread into three sentences so the executive can decide whether to read the whole thing, and drafting routine scheduling replies ("happy to do Thursday, does 2pm work") that a human still reviews before sending.
It's a poor fit for anything requiring real judgment about relationships or politics, like deciding how to phrase a delicate decline to a board member's request, or reading between the lines of a terse reply from an acquirer's counsel. Those calls belong with the assistant or executive who knows the relationship, not with a model pattern-matching on email tone.
Quality tradeoffs, honestly
For structured drafting, turning notes into invites, summarizing threads, drafting routine replies, a self-hosted model does the job well and the output needs only light editing. It's noticeably weaker at matching a specific executive's voice on nuanced or high-stakes correspondence without a fair amount of example text to work from, and it won't pick up on office politics or relationship history the way a longtime human EA does. Expect to keep a human in the loop on anything that goes out under the executive's name to a board member, investor, or counterparty, which is good practice regardless of which AI is drafting the first pass.
Setup effort
Most of the setup is connecting the model to read-access views of the calendar and inbox it needs to draft from, which typically takes an integration with your existing calendar and mail platform's API rather than a wholesale platform migration. Give it a handful of example emails in the executive's actual voice before asking it to draft correspondence; without that, its default tone will read as generic. Plan for a short calibration period where the assistant or chief of staff corrects draft tone and phrasing until it settles into something usable.
Where the hardware fits
This workload is lighter than document-heavy use cases, mostly short drafts and summaries rather than long-context reasoning, so a single Spark at $0.79/hour, with 128GB of unified memory to spare for running the model alongside the retrieval layer that reads calendar and email context, comfortably covers one executive's office. Running it dedicated rather than shared with other workloads keeps response times snappy for real-time scheduling back-and-forth.
A note on scope
None of this argues for running a full-blown personal assistant with unrestricted access to everything the executive touches. A narrower scope, calendar and correspondence around board and deal work specifically, gets most of the benefit with less exposure than a general assistant granted access to an entire inbox and calendar by default. It's worth deciding deliberately what the assistant reads rather than granting broad access because a setup wizard made it the easy default.