Hong Kong·Hong Kong · IRD-compliant · Sub-minute

An operator.
Not a chatbot.

Autolee HR turns a single command into a fully reproducible payroll artefact — final pay, IR56, eMPF batch. Parsing is probabilistic. Math is deterministic. Every line is written to an append-only event spine designed for long-horizon reproducibility — built so that an IRD audit five years from now becomes a replay, not a reconstruction.

── No card. Full features. 14 days.

ledger · tty/payroll
command · 2m 14s ago
process jason cheung's resignation effective 31 mar, build IR56F
parsed · computed · sealed

Last day of service

2026·03·31

Unused AL (713 ADW)

HKD 6,420

MPF clawback

HKD 1,200

Net final pay

HKD 38,940

sealed · IR56F · ready to filesha256 · 8f2c…a91d

── The product

One surface.
Three engines.

A single command bar is the only surface you operate. Behind it sits a parser that maps intent, a calculator deterministic to the cent, and a ledger that retains every event indefinitely.

── 01/Intent

A parser that knows what HR managers actually say

“Run March payroll”, “Process Jason's resignation”, “Generate IR56B for the year”. Spoken in English, Cantonese, or Mandarin.

── 02/Math

Calculations that survive an IRD audit

713 ADW averaging, MPF caps, §11AA final pay, statutory holidays — implemented from first principles, unit-tested against published examples.

── 03/Ledger

Reproducible from an immutable event spine

Filings are derived from the event spine, not from spreadsheets. Replay the same events and you get the same artefact — no more “which version is final” debates. (Cryptographic hash-chaining is on the roadmap.)

Time per cycle

−86%

Artefacts reproducible

100%

In our pilots with 12 Hong Kong SMEs, the median payroll cycle — from raw timesheet to filed IR56 — collapsed from 14 hours to 21 minutes. The remaining minutes are review, not arithmetic.

── evidence / customer log

“For the first time, payroll feels like a database query — not a ritual. We closed March in twenty minutes and our auditor stopped asking for spreadsheets.”

WL
Winnie L.
Head of People, Series-B SaaS · Central, HK

── FAQ

Honest answers, no marketing.

Is this a wrapper around ChatGPT?

No. The parser uses an LLM for intent classification only. All calculations are deterministic code that runs in Hong Kong, unit-tested against IRD published examples.

What happens to my data?

Your event spine lives in your tenant. We never train on your data. Exports are one click — CSV, JSON, or the original signed PDF.

How do you charge?

14-day trial, no card. Then either an annual enterprise contract with a named success manager, or pure usage — HKD per filing. Pick what fits your stage.

Do I need to migrate from my current payroll?

No. Run Autolee HR in parallel for one payroll cycle and reconcile to the cent. Switch over when you're confident — most teams do so after their second run.

── ready when you are

Type one command See the rest disappear.