The AI writes the letter. A C# function does the salary.

MITS is a staffing firm carrying 600 to 700 people on payroll, with 300 to 350 of those changing in any given year. Every one of those changes generates a set of employment documents: an offer letter or a contract, and depending on the terms, a joining bonus letter, relocation or conveyance reimbursement letters, and a security document.

The documents were produced by hand. A recruiter’s offer details went to a member of the HR team, who read them, calculated the compensation breakdown, opened the right template, and typed the values in. Several hundred times a year, across two employee groups, BNP and MITS2, each with its own legally reviewed template family.

The letters are legal instruments. Notice period wording, confidentiality terms and contract language had been through legal review and had to appear identically in every letter, permanently. Anything that generated a document from scratch was out, regardless of how good the output looked.

The compensation arithmetic carried a second, sharper constraint. An error in it would not surface during testing or review. It would surface roughly three weeks after somebody started a new job, when their first payslip failed to match the letter they had signed, and the person discovering it would be a brand-new employee with no way to tell which figure was the correct one.

Two rules shaped the build. First, the compensation engine is deterministic C#, not a model. It derives every component from a banded ladder and statutory rules, then balances one component as the remainder, so the letter reconciles to the agreed total exactly. Second, language generation is template-filling, never document generation. The model reads the recruiter’s document and refines the candidate-specific wording. The Word templates hold everything that is about the company. The model never sees a blank page.

An ASP.NET Core 9 MVC application using Microsoft Semantic Kernel over GPT-4o, with OpenXML and ClosedXML for document generation. The pipeline runs six steps: server-side text extraction from the uploaded document; a structured extraction call returning a typed JSON contract; the deterministic compensation engine; two separate language calls, one refining responsibilities and one compressing them to fit the template’s physical space; template filling across both employee groups’ template families; and packaging into a single ZIP.

Reading and writing are deliberately separate model calls, so a defect in the output can be attributed to one or the other. Every run writes its extraction JSON to disk alongside the letters it produced, so any later question about why a letter says what it says has a mechanical answer rather than a reconstruction. HR can override any extracted value before generation, and the engine still does the maths on whatever it is handed. A person reviews the output before it is sent. That was a design assumption, not a limitation discovered later.

A full letter set that took a person around forty minutes now takes about five, most of which is review rather than production.

The change that matters more to MITS is a category of error that has stopped occurring. Compensation figures are now computed the same way every time, from one place. During the build we found a boundary defect in the engine, fixed it in a single character, and proved the correction across every input the system had ever seen. That is not a property manual entry ever had.