Private AI for investor relations and board reporting
A self-hosted model can turn numbers finance has already finalized into board-deck prose and investor updates, without putting material non-public information (MNPI) into a general-purpose AI tool. This post covers that drafting workflow. It is not securities-disclosure advice: what you must disclose, and when, is a question for counsel.
What's at stake
A board deck in progress routinely contains MNPI: unreleased results, a pending acquisition, a fundraise in progress, or a churn number not yet shared outside leadership. An investor update draft has the same problem in miniature, such as a candid line about a missed target or a customer at risk of leaving.
Running that draft through a cloud AI tool creates a copy of MNPI on infrastructure outside your control. Depending on your insider trading policy and information barriers, that alone may be worth flagging to whoever administers them, separate from any general privacy concern.
There's also a practical leak-surface issue: a board deck usually goes through several drafts before the numbers are final, and an early draft with a wrong or soon-to-be-revised figure sitting in a third-party log is a liability even if the final version is accurate and properly disclosed.
What self-hosting changes
Running the drafting model yourself means draft financials, deal status and board commentary go to your own model, not to an AI vendor's API, so no vendor retention or training policy applies. On hardware you own, that removes the third party entirely. A rented Spark is a dedicated machine assigned only to you, but it is GPUwerk's hardware operated under a DPA, so clear it under your MNPI policy like any other provider.
Either way, your disclosure obligations and insider trading policy apply exactly as before.
Where it helps
The strongest use is formatting and consistency: turning a spreadsheet of quarterly metrics into board-deck prose with the numbers stated correctly, or checking that this quarter's update uses last quarter's metric definitions. When that slips, investors get confused. It also drafts a first-pass narrative around finalized numbers, which frees the CFO or head of IR for the parts that need judgment: how to frame a miss, what to emphasize, what tone the board needs.
It's a poor fit for disclosure judgment: deciding what's material, how to characterize risk, or how to phrase forward-looking statements. Those decisions need a human who understands the company's specific legal exposure, and ideally counsel's review, not a model's best guess at standard IR language.
Quality tradeoffs
For turning finalized numbers into readable prose or checking a draft against last quarter's terminology, a self-hosted model is reliable. It's weaker at the persuasive narrative that frames a difficult quarter for a board or investors. A frontier model does better there, and even its draft needs heavy editing by someone who knows the audience. Use it for the mechanical first draft and keep narrative judgment with your CFO or head of IR.
Setup effort
Most of the value comes from giving the model your last several quarters of board decks and investor updates as reference material, so it can match your established format and tone instead of producing something generic. That's a straightforward document-loading exercise, typically a day or two, not a technical integration. Keep the model's access scoped narrowly to whoever prepares these documents; this isn't a workload to expose broadly across finance or the company.
One Spark covers a board cycle
Board packs with several quarters of history attached are long, and a single Spark's 128 GB of unified memory keeps that history in context for consistency checks without splitting it up. At $0.79/hour in Europe, North America or East Asia, billed per minute, running it around board prep keeps the cost small next to what a leaked draft figure could cost in investor trust.
A note on scope
Limit access to people who already have legitimate access to the underlying numbers: finance, the CFO's office and the head of IR. Self-hosting takes the AI vendor out of the workflow. It doesn't solve internal access control, and MNPI still needs the same need-to-know discipline inside the company as it would with no AI involved. Treat the model as another system that touches sensitive data, subject to the same access review as any other finance tool.