Matthias Eilenbrock
Data Scientist · IT Forensics · Developer of AUDIPY
Hosted at Symposium · Statistical Auditing · 20 May 2026
Limperg Instituut
?
 
Audit Analytics Summit · 20 May 2026
Beyond Excel.
Before Code.
When your data talks back.
M
Matthias Eilenbrock
Data Scientist · IT Forensics · North Rhine-Westphalia · Developer of AUDIPY
n= 142 points
σlocked
shapeunknown
 
01 / 21
?
A Premise
Efficiency is not
faster spreadsheets.
Efficiency is coverage per hour.
How much of the books actually gets examined — at any level — per auditor per day. That is the measure. Everything else is motion.
 
Past
Chapter 01 / 03
Where audit lives today.
The tools, the habits, the daily reality of 13,000 German tax auditors — before we talk about what comes next.
01
 
03 / 21
III

Respectedtools.
Familiarlimits.

A tabular mindset: sample, look, conclude.
Analytics
IDEA
CaseWare's query-driven audit workhorse. Scripting for macros. The professional backbone since 2001.
Spreadsheets
Excel
Still the most common analysis surface. Pivot tables, filters, sums. Every auditor knows it.
Source Data
DATEV · SAP · GoBD
Legally mandated data access. The raw reality of every larger audit engagement.
"
But with this toolkit, we only ever see what we already knew to ask.
 
04 / 21
 
Same mean.
Same deviation.
Different stories.
If you trust only the numbers — you don't yet know the data.
datasaurus · dino · n=142
#
x
y
001
59.23
38.33
002
50.00
55.26
003
56.15
14.10
004
49.74
17.56
005
30.00
29.10
006
64.36
24.10
007
22.31
61.79
008
51.28
14.87
…
…
…
mean x
54.26
mean y
47.83
sd x
16.76
sd y
26.93
corr
−0.06
 
Live · 05 / 21
 

The statistics are locked.

Watch the numbers on the right while the shapes on the left change. They won't move.
Now showing: dots
Matejka & Fitzmaurice, 2017
dots
Statistics · two decimals
Count
142
Mean X
54.3
Mean Y
47.8
Std X
16.8
Std Y
26.9
Corr X/Y
-0.06
Thirteen shapes. One set of statistics.
 
Live · 06 / 21
Same numbers · different worlds
There was a dinosaur
hiding in your data,
and nobody  saw it.
 
μx = 54.26   ·   μy = 47.83
σx = 16.76   ·   σy = 26.93
ρ = -0.065
 
Present
Chapter 02 / 03
Three new ways
of seeing.
Visualization came first. Then machine-assisted detection. Then conversation with the data itself.
02
 
08 / 21
Wave One

Visualization comes to audit.

German tax authorities adopt Power BI at scale — shared dashboards, live refresh, one common visual language.
POSTINGS · €M · FY24 → PERIOD-END ANOMALY LIVE · 08:42 120 90 60 30 JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC flagged · cut-off risk
The auditor could finally look, not only query.
but a dashboard is not a dialogue.
 
09 / 21
A Fact About People
13,000
tax auditors.
Will not become Python developers.
The data-science toolkit lives in code. The practitioners do not. This is the wall we have to get around.
 
10 / 21
A Purpose-Built Tool
The data-science stack, with an interface auditors can actually use.
Import
Source → structure
DATEV, SAP extracts, GoBD folders, e-invoices - streamed into an analysis-ready format in one step.
Analyze
One click, not one script
Isolation Forest, parallel coordinates, Benford's law, sampling. No code needed to run them.
Orchestrate
Agent playbooks
Multi-step audit workflows described in plain text, executed by an AI agent. We will come back to this.
The power of a data-science notebook, the approachability of a spreadsheet.
Free of charge for tax authorities and educational institutions — under a dedicated cooperation agreement. AI-powered features are licensed separately.
Eligible public bodies may also qualify for a Government Open Source Agreement, granting source-code access for sovereign deployment.
 
11 / 21
How the data arrives

Audit at the speed
of Netflix.

You don't download a film before you watch it. You stream it — the bytes that matter, the moment you need them. Audit data should work the same way.
Source
40 GB
Journal entries · 3 fiscal years · raw SAP export
Only what you ask for columnar reads · pushdown filters · in-memory
Auditor
< 1 s
The five columns, the thousand rows, the question asked — at the speed of thought
DuckDB inside AUDIPY — columnar, just-in-time. Forty gigabytes, one query, from a laptop.
 
Demo · 100M
Live · One hundred million rows

Pivot at 100 M.

A real journal. One hundred million lines. The auditor types a question — the pivot returns before the thought finishes.