Is it a problem to paste an employee's review notes into a public AI chatbot?

It depends on the vendor's data-handling terms for the plan you're on, but the underlying risk is structural: the text leaves your infrastructure and is processed on shared servers you don't control, alongside performance opinions about a named, identifiable person. Most HR and legal teams treat that as personal data requiring the same handling as payroll records, not as casual text.

Does running the model locally remove manager bias from the review?

No. A local model drafts from whatever notes a manager gives it, and it will reproduce the bias in those notes as readily as a cloud model would. Self-hosting changes where the text goes, not the judgment behind what was written. Bias review is still a human and process question.

Blog/Private AI for performance review drafting

For AI assistants

Private AI for performance review drafting

By Samuel Seidel · Updated September 9, 2026

A performance review draft is a manager's raw, unfiltered opinion of a named employee's work, sometimes including salary context, disciplinary history, or a note about a promotion that hasn't been announced. Using AI to turn scattered notes into a coherent draft is genuinely useful. Where those notes go while that happens is a separate question worth answering deliberately.

What's actually in a review draft

Managers writing performance reviews often paste in more than the final wording needs: peer feedback quotes, PIP status, comp discussion notes, sometimes a coworker's name in a story about a conflict. None of that is meant to leave the manager-HR relationship, and most of it identifies a specific employee. A cloud AI tool used to smooth the language into a final draft sends all of that context along with the request, processed on infrastructure outside the company, under whatever retention policy applies to that specific plan.

This is a narrower problem than most AI-and-HR-data articles suggest. Review text isn't usually regulated the way health or financial data is, but it's exactly the kind of information an employee would reasonably expect to stay inside HR and their direct manager, and a lot of it becomes materially embarrassing if it surfaces the wrong way, mid-cycle, before a review is delivered, or during a later dispute.

What changes when the model runs on your own hardware

A dedicated DGX Spark running the drafting model keeps the notes and the draft on infrastructure your company controls. There's no vendor server the text passes through, no plan-tier question about whether inputs get used for training, because the request never leaves your network. The workflow looks the same to the manager: paste rough notes, get back a structured first draft to edit. What differs is everything downstream of "send."

A practical setup is Open WebUI configured with a review-drafting system prompt, restricted to the managers and HR partners who need it, on a Spark that isn't shared with unrelated workloads. For teams already running retrieval over internal documents, the same instance can pull in a company's rubric or values language so drafts stay consistent without a manager having to paste the whole document each time.

What a model should and shouldn't decide

A drafting model is useful for structure and tone: turning a manager's bullet points into full sentences, checking that a draft covers the required review categories, or flagging when language reads more harshly than the underlying notes support. It should not be the source of the substantive judgment, and it shouldn't see comp bands or headcount plans it doesn't need for the task in front of it. Scoping what a manager pastes in, even on a private instance, is still worth doing. Self-hosting removes the third-party exposure; it doesn't substitute for a manager keeping the input tight.

Rating consistency and bias

Some HR teams ask a model to flag inconsistent language across a manager's reviews, for instance the same behavior described as "assertive" for one employee and "difficult" for another. That's a reasonable use of a private model, since the comparison requires reading multiple employees' review text together, which is exactly the kind of aggregation that shouldn't happen on a third party's servers. It's a check on drafting language, not a substitute for a calibration meeting, and it won't catch bias that isn't visible in word choice.

What this doesn't solve

Self-hosting the drafting model doesn't change your legal obligations around personnel records, doesn't create an audit trail HR didn't already have, and doesn't decide who should see a given draft before it's delivered. Those are policy questions for HR and legal, not infrastructure questions. What a dedicated instance removes is one specific exposure: a third-party AI vendor holding a copy of what a manager wrote about an employee before that employee has read it themselves. See our guide on pricing for what running a dedicated Spark for an HR team actually costs.