AI ProductFinancial Crime

Audit Assure

It runs automated audit tests on the bank's own data every night and explains every finding. Auditors decide.

About Audit Assure

The idea

A bank's internal audit team has to test its controls across every loan, deposit, ledger posting and change of staff access. Checking a sample by hand misses things, and banks need to be ready for the Central Bank's 1 January 2027 audit rules. Audit Assure gives the Chief Internal Auditor one command dashboard over the bank's own data. Each night it reads the core systems over read-only connections, masks customer identities, and runs 10 automated audit tests over every record in scope: duplicate identities, loans with no security recorded, postings to closed periods, access raised just before a posting, and more. Hits are grouped into ranked cases, and a branch risk heatmap ranks every branch by open findings, entity risk and pattern hits. Every finding shows the rule that matched, the records behind it and the innocent explanations to check first. An auditor reviews it, can ask questions in a chat that answers from the masked data, and records a decision: fraud case, control weakness, false positive or no issue. Findings are tracked through to the audit committee, and the output is a readiness pack for the Central Bank's rules. It runs on the bank's own Red Hat Linux server, and every scan, approval and decision is kept in a tamper-evident log.

Product tour

How Audit Assure works

  1. Open the command dashboard

    The dashboard shows confirmed frauds this financial year, money prevented and recovered, open findings and the last nightly run. Pattern coverage says how many tests ran and how many could not, so nothing looks clean just because it was skipped. Below it, the branch heatmap ranks every branch by open findings, entity risk and pattern hits.

    Open the command dashboard
  2. See how each test is doing

    Alerts by pattern show, for each test, how many alerts were confirmed, are still under review, or turned out to be false positives. A test that raises mostly false positives gets tuned or retired. Underneath, detections are counted day by day over the last 30 days.

    See how each test is doing
  3. Work the findings queue

    Open findings are ranked by severity, then by how unlikely their combination is to be chance. Each one says in plain words why it was flagged, and that it is a lead, not a finding. The auditor records the outcome: open a fraud case, log a control weakness, close it as a false positive with a reason, or mark it no issue.

    Work the findings queue
  4. Check the evidence

    The evidence page shows the rule that matched, the records it matched, and the known innocent explanations to check before escalating. A side panel says how strong the case is: here, about 4 in 9 subjects with this combination would turn up by chance, so it is a lead worth a look, not a strong anomaly.

    Check the evidence
  5. Ask about a finding

    The deep dive answers questions from the masked data: what the finding is, why it was raised, and what would make it a false positive. The answer names the pattern, its threshold and the dates behind it, and repeats that the finding is a lead for a person to check.

    Ask about a finding
  6. Describe a new test

    An auditor describes a new test in plain words, and the assistant drafts it: title, audit assertion, severity, threshold and the evidence columns to show, with every assumption listed so the auditor can remove it. A second person must approve the pattern before it runs.

    Describe a new test
The film · 0:41

See Audit Assure in motion

Sample data · synthetic demo
The problem

What Audit Assure catches

Case 01

A loan with nothing behind it

A facility is disbursed and sits in the books like any other. Every field on the loan record is filled in.

Red flag: No item of security is recorded against the facility.

Case 02

Access raised just in time

A staff member's access rights are raised, and a few days later they post a ledger entry. Each step on its own looks routine.

Red flag: The privilege escalation came within days of that same person's own posting.

Case 03

A change to a closed period

The period is closed and its results reported. Then a new entry appears in the ledger.

Red flag: The entry was posted after its account's period had closed.

Case 04

One identity, two customers

Two customer records with different reference numbers, opened at different branches.

Red flag: Both records carry the same national identity number.

Stage by stage

How work moves through Audit Assure

  1. Extract

    Each night it reads the bank's core systems over read-only accounts against standby copies, plus scheduled HR and directory file exports.

  2. Mask

    Customer names, identity numbers, addresses and contact details become consistent codes, so the same person keeps the same code. A named person signs off the column classification once; it then runs each night while the data layout stays the same.

  3. Test

    Every active pattern runs over every record in scope, with no sampling. The same patterns on the same data give the same results, even a year later.

  4. Rank

    Hits are grouped into cases and ranked, allowing for patterns that coincide by chance, and branches are scored for the heatmap.

  5. Review

    Auditors work the findings queue, check the evidence, ask the assistant, and record a decision: fraud case, control weakness, false positive or no issue.

  6. Report

    Findings are tracked through to the audit committee, and the output is a readiness pack for the Central Bank's 1 January 2027 audit rules.

Ground rules

Promises it always keeps

  • Not run is never a pass

    A test missing its input data is shown as untestable or unable to run, with what is missing, never as a clean result.

  • A person decides every finding

    Every hit is a lead, not a finding. Nothing is closed or escalated without a person deciding.

  • Detection never calls a model

    Patterns are fixed, versioned rules, so a scan re-runs to the same result. The AI assistant explains and drafts; it never picks which records get flagged.

  • Two people for every pattern

    The person who approves a pattern is never the person who drafted it, and the software enforces it.

  • Read-only against the bank

    It has no write access of any kind to the bank's systems.

  • Every action is on the record

    Scans, approvals, decisions and AI turns go into a hash-chained log, so an edited, backdated or reordered entry shows up when the chain is checked.

Under the hood

What's inside Audit Assure

Dashboard

The command dashboard, findings queue, Discover, Pattern Library and Data Sources screens, with role checks on each.

Pattern library

Audit tests kept as declarative records in seven detection kinds, such as duplicate groups, missing related records, time proximity and a first-digit (Benford) screen. No SQL inside a pattern.

Masking

Replaces identities with consistent keyed codes before analysis. A separate locked store turns a code back into a name for a permitted officer, and every use is logged.

Scheduler

Runs the nightly cycle and queues the scans users ask for. It holds no credential that reaches bank data.

AI assistant

Answers questions about a finding, branch, pattern or feed from masked data, and drafts new patterns, within set limits on rows, time and spend.

Evidence log

A tamper-evident, hash-chained record of every scan, approval, decision and assistant turn.

By design

What it deliberately won't do

  • It won't: Write to the bank's systems

    It reads over read-only accounts against standby copies and never changes a record.

  • It won't: Say that fraud happened

    Every hit is a lead for an auditor. Only a person's decision turns it into a confirmed case.

  • It won't: Let AI choose what gets flagged

    No model sits in the detection path. Patterns are fixed rules that a reviewer approved.

  • It won't: Send customer identities to a model

    Names, identity numbers, addresses and contact details are masked before analysis, and an egress check sits on every path out.

  • It won't: Close findings by itself

    There is no automatic closure or escalation. A person records every decision.

  • It won't: Test only a sample

    Every record in scope is tested on every run.

  • It won't: Run as a shared cloud service

    It runs on the bank's own Red Hat Linux server, with PostgreSQL and the application in containers.

Technical details

Audit Assure runs on-premises on the bank's own Red Hat Linux server, with PostgreSQL and the application in containers. It is four apps written in TypeScript on Node: a Next.js dashboard, a scheduler that runs the nightly cycle, a worker that extracts, masks and scans, and a host for the AI assistant. They talk through the database, and each part runs under its own database role (eight in all), so the part that reads raw data is not the part that scans it. Core systems are read over read-only accounts against standby copies; HR and directory data arrive as scheduled file exports, with nothing installed in those systems. Masking uses consistent keyed tokens (HMAC), so the same person keeps the same code across records. Patterns are declarative records over a common data model, in seven detection kinds, with no SQL inside them. Sign-in is through LDAP (Active Directory) or local accounts with Argon2-hashed passwords, with role checks, and TLS through nginx. The AI assistant reaches a configurable model endpoint behind an egress check and limits on rows, time and spend; it is never part of detection.

Target audience

Chief Internal Auditors and audit committees at banks. Also the internal auditors who work the findings day to day, and the reviewers who approve new audit tests before they run.

Questions

Does it change anything in our core banking systems?

No. It reads over read-only accounts against standby copies and has no write access of any kind to the bank's systems.

Does customer data leave the bank?

The application and its database run on the bank's own server. Before analysis, customer names, identity numbers, addresses and contact details are masked. The AI assistant calls a model endpoint the bank configures, and an egress check scrubs what is sent.

Does the AI decide what gets flagged?

No. Detection is fixed rules that re-run to the same result. The AI explains findings and drafts new patterns; a person approves every pattern and decides every finding.

What happens if a test can't run?

It is shown as untestable or unable to run, with what is missing, never as a clean pass.

Can we add our own audit tests?

Yes. An auditor describes the test in plain words, the assistant drafts it with its assumptions listed, and a second person approves it before it runs.

What happens to tests that raise too many false alarms?

Each test's alerts are tracked as confirmed, under review or false positive. Patterns that raise mostly false positives get tuned or retired.

Where does it run?

On the bank's own Red Hat Linux server, with PostgreSQL and the application in containers.

How do people sign in?

Through LDAP (Active Directory) or local accounts with Argon2-hashed passwords. Each person's role decides what they can see and do.

Book a demo

See any of the nine running on your own use case.

A 30-minute walkthrough with the team that built it.

Email us
hello@bistecglobal.com

Offices in Sydney and Colombo.

bistecglobal.com
© 2026 BISTEC Global · Intelligent design. Engineered for adaptability.Demos on this page use sample data.