Migrating systems, running parallel, comparing periods or auditing a payroll run? PayrollCompare AI compares your documents field by field and tells you exactly what is the same, what changed, and what is missing. No system access, no configuration. You run your documents through it and read the result.
Powered by the Khaonix platform.
Built to add markets as configuration, not as rebuilds.
Need your company-specific wage types recognised?
Contact me to scope a client-specific model: kees@khaonix.ai
When two payroll systems should agree, for example in a migration, a parallel run or an audit, confirming they actually do is slow, manual work. You export from both sides, line them up, normalise formats that do not match, compare value by value, chase down every difference, and repeat it for the next pay run. PayrollCompare AI replaces that with three steps:
- Upload file
- Compare
- Review changes
Built differently
- Bounded by design. The model reads exactly the payroll fields it was specified to read, and nothing beyond. Its limits are known before it ships, not discovered in production.
- Trained without your data. The model learns entirely from synthetic data, manufactured for the purpose. Your documents are read to produce your result, then discarded.
- Specified, not assumed. You tell the system which payroll items your company uses. It works backwards from that specification instead of guessing from your documents.
This is the Khaonix platform. PayrollCompare AI is the first product built on it. Read how the method works.
Scope
The standard product reads a defined scope of 20 payslip fields, documented and versioned. Companies with their own wage types or additional fields can have a client-specific scope built on request: kees@khaonix.ai
Two ways to run it
- Hosted API. The service runs in the EU (Finland), accessed over HTTPS with per-client API keys. Documents are processed in memory and nothing is written to storage. Run it synchronously and the result returns on the same request, with nothing remaining on the server afterwards. Run it asynchronously for scheduled workloads and results are held in memory only, for a limited time. Full data handling is documented under Privacy / Security.
- Local deployment. For organisations whose documents cannot leave their own infrastructure, the same service deploys on-premise, available on request. No GPU is required. The measured workload below runs entirely on CPU, on modest commodity hardware.
Measured performance
All figures measured on the production service over public HTTPS, on an 8-vCPU, 16 GB cloud instance with no GPU: Extraction sustains approximately 7 payslips per second. A batch of 5000 payslip pages completes in around 12 minutes in a single synchronous request. A comparison costs roughly two extractions: thousands of document pairs, one call. No ceiling was encountered at any tested size. A payroll comparison is rarely one document; it is a monthly run, a parallel period, an audit sample. PayrollCompare AI processes these as a single request.
The extraction API
Underneath PayrollCompare AI sits the same capability as a building block: structured payroll fields extracted from any payslip, returned as data. If you are building your own workflow, an integration, or an AI pipeline that needs reliable payroll data from documents, the extraction API is the piece that reads them. It returns exactly the fields it was specified to read, as structured data, and nothing beyond. Contact me: kees@khaonix.ai
PayrollCompare AI is the first product built on the Khaonix platform: a system that manufactures bounded document models from a specification, without ever training on real documents. Payroll is the first proof the machinery runs end to end, not the limit of what it is for.