The overdue compensation review — figures on your hardware, drafting from a sanitized brief
Finance & HR
The business problem
Compensation reviews slip. The policy says every salary is reviewed every 12 months; in practice, people pass 14, 20, 26 months without one because nobody has time to reconcile the HR export against the policy. It is unglamorous, sensitive work — and exactly the kind of work that falls between Finance and HR.
The obvious shortcut — paste the spreadsheet into a chatbot — is a bad one: names and salary figures are the most sensitive data most companies hold. The alternative is doing it by hand, which is how reviews get forgotten in the first place.
What happens
You approve the input
HR exports the review data — names, roles, last-review dates, salary figures — into a file the Work Twin is explicitly allowed to read. Nothing else is in scope.
Stage 1 runs on your own hardware
A model on a machine in your office reads the export and finds who is overdue and by how long. This is the only stage that ever sees names or figures.
Stage 2 drafts from a sanitized brief
A hosted model receives only the conclusion — how many people are overdue, by how many months, under which policy — and drafts the note to the department head. No names, no figures.
A person decides and sends
You review the findings and the draft. The run is logged: which model handled which stage, and what each stage received.
A real run, 2026-07-29
This is not a mockup. In a logged run on 2026-07-29, stage 1 ran on a GB10 on the local network with Step-3.7-Flash (16,043 ms, 530 tokens) over an input containing real salary figures and three names; it correctly flagged two people overdue at 14 and 26 months against a 12-month cycle. Stage 2 (Kimi K3, hosted) received only the sanitized conclusion — two employees overdue, at 14 and 26 months, 12-month policy — and produced a usable note to the department head. No names and no figures reached the hosted model.
Two honest footnotes from the same record. First, the routing was configured deliberately: the product supports choosing a model per stage — it does not automatically detect what is sensitive. Second, the run shows where each stage ran; it is not a compliance attestation of any kind.
What it costs you to try
A pilot is deliberately small:
- One approved HR export and one written review policy — no system integration is required for the first run.
- One assessment call to decide which stages run on your hardware and which, if any, may use a hosted model.
- Everything else — hardware, models, logging — is scoped and priced in the assessment, with the assumptions printed beside the numbers.
What this does not do
- It does not decide what is sensitive on its own. You choose the model per stage — we help you set that up in the assessment; the product does not auto-classify your data.
- It is not a compliance attestation. It records where each stage ran and what it received; your legal obligations and their assessment stay with you and your advisors.
- It does not decide compensation. It surfaces who is overdue under your policy and drafts a note; the judgment and the send button stay with people.
- It does not require the hosted stage. A sensitive-only configuration keeps every stage on your own hardware.
Frequently asked questions
- Does the hosted model ever see salaries or names?
- Not in this workflow. In the logged run of 2026-07-29, the hosted stage received only the sanitized conclusion — two employees overdue, at 14 and 26 months, under a 12-month policy. The stage that saw names and figures ran on a local GB10, and the run record states what each stage received.
- Where do the numbers 16,043 ms and 530 tokens come from?
- From the logged run on 2026-07-29: stage 1 (Step-3.7-Flash on a GB10 on the local network) took 16,043 ms and produced 530 tokens. They are measurements of one real run, not performance promises — your latency depends on your hardware and input size.
- Can we run it with no hosted model at all?
- Yes. The model is chosen per stage; an all-local configuration keeps both stages on your own hardware. The mixed configuration exists for cases where a hosted model is acceptable for non-sensitive drafting.