---
title: "How we use AI in audit and why a human signs"
description: "The firm uses language models for drafts, recalculation and finding provisions. Three rules do not change: models are not trained on client data, personal data enters the contour de-identified, and a human signs the document. Here is how it works."
date: 2026-10-30
lang: en
rubric: "AI in audit and finance"
tags: ["ai", "methodology", "quality control"]
issue: 1
author: "tilepin"
publisher: "TRI-S-AUDIT"
canonical: https://tri-s-audit.uz/en/blog/ai-contour-three-rules/
license: "CC BY-ND 4.0 — cite with a link to the canonical URL"
---

# How we use AI in audit and why a human signs

The firm uses language models for drafts, recalculation and finding provisions. Three rules do not change: models are not trained on client data, personal data enters the contour de-identified, and a human signs the document. Here is how it works.

## Key points

- AI at the firm is a technological resource of the system of quality management under ISQM 1: it is entered in the register of approved resources, has an owner and a quarterly check of its settings.
- Models prepare draft documents, recalculate computations independently and find provisions on lex.uz; a human makes the decisions and signs.
- Client data is not used to train models; personal data enters the contour only in de-identified form.
- Every use of a model leaves a trace in the working papers: what was requested, what was received, who reviewed it.

## Why use it at all

Audit means working with large datasets and long documents. A model reads an accounting system extract faster than a person, notices unusual entries and prepares a draft working paper from a template. The human spends time on judgement: whether something is material, what it means for the opinion, what to tell the board of directors.

## Three rules

**Rule 1. Models are not trained on client data.** We use only tools and plans in which training on user data is switched off, and we check this setting quarterly and record it in the register. A tool that is not in the register is not applied to client data.

**Rule 2. Personal data only in de-identified form.** Before an extract enters the contour, the names of employees, individual counterparties and identifiers are replaced with codes. This is a requirement of the Law "On Personal Data" and of our privacy policy, and it is not negotiable.

**Rule 3. A human signs.** A model does not issue documents. A draft, a calculation or a draft report is checked by the preparer, reviewed by the supervisor and, for reports, reviewed by the partner responsible for quality. The partner signs and is responsible for the document as a whole.

## What it looks like on an engagement

The firm's methodology is written in two layers: the Manual for people and a machine layer for the model, which is extracted from the Manual and cannot contradict it. The model receives instructions from the machine layer: which journal entry tests to run, how to calculate materiality, in what format to prepare a working paper. The result goes to the engagement's drafts folder and waits for review. Independent checking of calculations is performed by a separate agent that does not edit the document but only recalculates the figures and reports differences.

## What this gives the client

Procedures over the full population of data instead of a sample, consistent quality of documentation on every engagement, and transparency: we can show which step was performed by a model and which by a human. What it does not give: reports "from AI". The auditor's opinion remains a human's opinion, and the firm is responsible for it.

## Sources

1. International Standard on Quality Management (ISQM) 1, paras 32–35: resources and technological resources — https://www.iaasb.org/
2. Law of the Republic of Uzbekistan "On Personal Data" (ZRU-547), Articles 15 and 27-1 — https://lex.uz/
