Case study · Audit automation engineering

The audit file that writes itself — and proves every number.Validated

AuditCompass turns raw client records into ISA-compliant working papers — planned, generated, validated and cross-referenced end-to-end by an orchestrated team of AI agents under deterministic control. The same engineering is available for your firm.

0
Worksheet templates
planning → completion
0
Methodology skills
ISA 200–720 codified
0
Specialist agents
orchestrated automatically
0
Untraceable figures
tolerated in output
A.1 — SCOPE

Records in. Graded working papers out.

An auditor drops the client's records — trial balance, bank confirmations, contracts, payroll, in whatever format the client sent them — and describes the engagement in plain language.

AuditCompass classifies the evidence by content, plans the engagement on a dependency graph, and produces each working paper the way a senior professional would: risk first, materiality confirmed by the human, then substantive testing with every figure traced to its source.

  • A-secPlanning chain — acceptance, materiality, entity understanding, controls, analytics, risk register, strategy
  • B-secSubstantive testing — confirmations, reconciliations, counts, cut-off, ISA 530 sampling
  • C-secCompletion — misstatement evaluation, representation letter, auditor's report drafts under Czech law
  • QA27-check gate — the file cannot ship until materiality, risk and finding chains reconcile
01 · Input

Client records, as-is

Any structure, any naming. Files are classified by content — a filename is a hint, never evidence.

02 · Engine

Agents × deterministic validation

Seven orchestrated agents draft the professional judgment; a Python engine enforces structure, traceability and exact terminology. Neither is trusted alone.

03 · Output

Graded, print-ready files

Excel working papers graded A–F per sheet with a complete audit trail. The auditor reviews and signs — responsibility stays human.

B.1 — EVIDENCE

Engineered so fabrication is structurally impossible to ship.

Generative AI in audit has one disqualifying failure mode: plausible invented numbers. AuditCompass treats that as an engineering problem, not a prompt-writing problem.

Every monetary amount in every narrative is parsed back out of the finished worksheet and traced against the imported trial balance and a hash-verified evidence ledger. An amount that traces to nothing fails validation — the worksheet is rejected, not delivered with a footnote.

The same discipline covers document numbers, sampling provenance, dropdown vocabulary, materiality propagation and cross-worksheet consistency — the things quality inspectors actually check.

$ generate.py validate B.7.1 --semantic --trace-hard autonomy=full-auto → hard traceability enforced   Sheet Grade Fill Purpose ───────────────────────────────────────── Souhrn A 100% Rekonciliace A 100% reconciliation Test 4 A 100% confirmation Test 3 A 97% physical count … 7 more sheets … ───────────────────────────────────────── OVERALL A/B 98% 0 fabricated refs   ✓ every CZK amount traced to evidence
B.1.1

Anti-fabrication tracing

Amounts, invoice numbers and counterparties must resolve to the trial balance or a hash-verified source file. UNTRACEABLE = rejected.

B.1.2

Codified methodology

29 skills encode ISA 200–720 and Czech statute as decision logic — fraud presumptions, going-concern trees, sampling math — routed per worksheet.

B.1.3

Orchestrated agent team

Planner, evidence classifier, generators, adversarial reviewer, findings evaluator, report drafter, chain auditor — dispatched automatically by engagement circumstance.

B.1.4

Complete audit trail

Every generation logged with SHA-256 of inputs; manual Excel edits diffed and recorded per ISA 230. ISQM 1 evidence, by construction.

B.1.5

Chain validation

Materiality, risk and misstatement chains reconcile across the whole file — a hand edit that breaks consistency warns immediately.

B.1.6

Human authority preserved

Opinions, materiality and going concern remain the auditor's decisions. Per-worksheet sign-off is a configurable gate, not an afterthought.

C.1 — SERVICES

Built once for my own methodology. Buildable for yours.

01

Working-paper automation

Your templates, your methodology, your review hierarchy — generated and validated instead of typed. Any framework: ISA, national GAAS, internal manuals.

02

Evidence pipelines

Client-records intake that survives reality: arbitrary formats classified by content, balances reconciled, every figure anchored to its source with a verifiable trail.

03

Quality & compliance gates

Deterministic validators that grade output like an inspector would — terminology exactness, consistency chains, completeness — wired into your delivery workflow.

04

AI operators for your team

Claude-based assistants that follow your firm's rules: codified methodology skills, guarded autonomy levels, audit-trail logging from day one.

Michail Muratov

I build production-grade automation for audit and finance practices. AuditCompass is my own platform — methodology, engine and agents engineered end-to-end — and the working proof of what I can deliver for yours.

C.2 — CONTACT

Let's automate your audit practice.

A 30-minute call is enough to identify which of your workflows automates first — and what the saving looks like per engagement.

muratov.io  →
or find Michail Muratov on LinkedIn