Security

Payroll AI Without Storing Your SSN

2026-10-01

The usual objection to "AI for payroll" is a single question: where does my employees' data go? It's the right question. Most AI answers are bad answers.

The typical pipeline ships sensitive text to a model provider, stores transcripts full of SSNs and birthdates, and hopes the retention policy holds. Ragbee was designed from the other direction: answer the question without ever keeping the sensitive parts.

Three rules that shape everything

Structured answers never touch an LLM. When you read your employee ID or a verification code aloud, that value goes through strict format parsers — not a language model. There is no prompt to inject, no model to leak.
Identity lives in flight, not at rest. SSN and date of birth are used to verify you and then discarded. What persists is a short-lived, scoped token — and every stored message and transcript passes automated PII redaction before it touches disk.
Every request is scoped to what the person actually owns. An employee's credentials unlock exactly their records, for exactly their active companies, for exactly as long as the conversation lasts. Phone sessions end when the call ends.

Prove it, don't promise it

Claims like these are cheap, so we attacked our own system: prompt injection, tenant isolation, forged tokens, brute force, webhook spoofing, user enumeration. The interesting results are the boring ones — the cross-tenant request returns 403, the injected instruction returns "I didn't catch a valid Company ID," and the brute-forced PIN locks the line after five guesses.

AI for HR is going to be judged on exactly this axis. The platforms that treat identity and data minimization as the architecture — not a compliance afterthought — are the ones that will pass procurement. We built Ragbee that way because it's the right answer, and it turns out to be the defensible business answer too.