National Tax Service (NTS)
AI-assisted company analysis for tax investigators
About the program
AI cuts the time analysis takes. But if checking whether the result is right takes longer, nothing has been saved. In work like a tax audit, where the evidence is the conclusion, that problem is acute. So the course paired "how to produce it fast" with "how to catch what is wrong." Participants cleaned financial statements and public filings into a form AI can read, analyzed the areas that looked questionable, and moved the result into a draft report — while learning, at each step, to check the output against the source. We prepared reference material so the models handle investigation terminology correctly, and designed the practice files as processed data so no sensitive records were involved.
What we ran
Speed and verification treated as one package, taking the output to a state where it can be used at work as it stands
How the course was designed
Verification built into the curriculum, on the premise that checking AI output can cost more than the analysis itself
Investigation terminology compiled in advance so the models do not confuse specialist terms
Practice designed on processed data rather than live records, for security
[Day 1] Understanding it and getting hands on
How generative AI works and how to choose a tool
Setting criteria for which tool suits which kind of task
Collecting and cleaning source material
Step-by-step practice turning public filings into a form AI can read
Financial statement analysis basics
Walking through the structure of a financial statement and the flow of an analysis
[Day 2] Applying it and producing something that lasts
Advanced financial statement analysis
Working through the analysis area by area and surfacing the parts that warrant a closer look
From analysis to a draft report
Moving the analysis into a draft report in the standard document format
Scenario-based practice
Short repeated exercises on the scenario types investigators actually meet
Handling hallucination
Tools that cite their sources automatically adopted as the default for the analysis stage
Authoritative sources registered in advance to anchor the answers
A self-check procedure against the original documents built into the practice
How it ran
One PC per participant in the computer lab, with a licensed seat secured in advance for everyone
An assistant instructor on site for individual help during practice
Recordings of every session, so anyone who missed one could catch up or review
Ground rules we covered
Never entering taxpayer information or non-public records
Every figure and citation from AI checked against the original before use
Gallery
