Designed and architected by Tom Scholtes
Live demo // a working demonstration on synthetic data

Run the after-state, live.

Point AI at the real finance workflows. It finds the discrepancies, builds the booking template, and drafts the email. The operator stays in control and signs off. Everything below runs on synthetic data, right in this page.

Workflow automation for regulated finance / Luxembourg fund services

The back office, done by the operator and the machine together.

Point AI at the real finance workflows. It finds the discrepancies, builds the booking template, and drafts the email. The operator stays in control and signs off. This is a working demonstration on synthetic data, not a slide.


What this shows

One concrete process, end to end. Then the same engine across the company.

Live demo / synthetic data

Pick a department, then a process.

Every process listed below is fully interactive: download the synthetic source files, upload them back, watch the machine run, and review the report, the booking template, and the draft email. A greyed process in each area shows exactly what it takes to map a new one.

Select a process on the left.

Start with Controlling → Bank Reconciliation for the full interactive run.

Interactive simulation on synthetic data. No real client, fund, or counterparty information is used.

How it scales

From one process to the whole company.

The live demo is one workflow built by hand. The same engine discovers and maps every other workflow automatically, from the mailbox the team already uses, into one shared map.

The live run

// live discovery run mailbox AI mapping shared map

A real session: AI reads a mailbox, reconstructs the workflow, and writes it into the shared map.

The chain

Mailbox AI mapping Shared map Extract Automation
Workflow Mapboard view
WorkflowDeptFreqAuto
Bank ReconciliationCTLMonthlyHigh
Distribution ReconciliationCTLEventWiring
NAV Pack PreparationFAQuarterlyMedium
Audit PBC ResponseTAXAnnualMedium
Investor ReportingIRQuarterlyHigh

Illustrative map. The discovery method itself is proprietary and is not shown here.

The same reconciliation, two ways
rework loop Pull back file Match in Excel Chase by email Fix exceptions Sign off by hand
To switch on a greyed process

Connect the account once: Monday.com plus Outlook via the AI connector. A discovery pass then maps that workflow automatically from the existing mail and writes it into the shared map. No new software for the team to learn.

// conviction

The joy is the argument.

This is not a pivot. It is a through-line finally called by its right name. Wherever I have worked, I found the manual repetition and I quietly killed it. That was always my real method; I just have the language for it now, and the tools to do it properly.

For years I felt like someone who loved a sport that did not officially exist. I played it anyway, on the side, in the margins of every job. Now the sport is real and it has a name: AI engineering. I get to do the thing I was always going to do, and this time it counts. I love using AI to build, to automate, to take a slow, human-draining workstream apart and make it quiet and fast. I build infrastructure late at night that nobody asked me to build, because I cannot not do it.

And here is the rare part: I already speak the language of funds, reconciliations, audits, the back office where the real numbers live, and I ship working automation on top of it. Most people have one side or the other. I want to be genuinely excellent at the seam where they meet, not merely adequate.

Working this way feels like standing next to an intelligence amplifier; one I engineered myself rather than bought off a shelf.

The joy is the argument.

Shipped systems, running in a regulated finance desk

I find the manual workflow, map it, then automate it.

I take the hand-run processes that live in spreadsheets, email chains, and one person's head, and turn them into documented, partly automated flows that hold up under audit.

discovery-signature.svg manual to mapped
real-process.svg before
// undocumented: the real process, before XLSX EMAIL RECON AUDIT FILES 1 person holds it all
The real process, undocumented
One person holds the whole sequence in their head; the status lives in an inbox, not a system.
25 to 70% of a process runs outside any logged system / the real flow rarely matches the official flowchart / one person's head is the single point of failure
// before and after, on the same process

Click any step to see what it became.

This is a real recurring workflow in fund controlling: the cash reconciliation cycle. On the left it is run by hand. Switch to the after view, then click any node to read how the manual step was discovered, mapped, and where it could be automated.

7manual 0automated
bank statements GL extract NAV file email threads one controller recon report audit request
manual step automated step single point of failure
// the detail, on demand

Open the parts that matter to you.

Two short, expandable explainers. The first contrasts the manual cycle with the automated one. The second walks the chain that turns a discovered map into a running tool. The discovery method itself stays proprietary; only the shape is shown.

01 Before and after on the same process

The same reconciliation, run two ways. In the manual cycle, one person pulls each source file, matches in a spreadsheet, chases open items over email, fixes the exceptions, and signs off by hand. The status of the work lives in an inbox, not in any system, and the loop reworks itself whenever a number moves.

In the automated cycle, the files are ingested on a schedule, the routine matching runs as a deterministic rule, only the breaks are surfaced, and the booking template and the email are generated for the operator to review. The person moves from lookup to judgement, and the single point of failure is removed because the process is written down.

before / manual
  • - Source files pulled and opened by hand
  • - Matching done line by line in a spreadsheet
  • - Open items chased over long email threads
  • - One person holds the full sequence in their head
after / automated
  • + Files ingested and normalised on a schedule
  • + Routine matching runs as a deterministic rule
  • + Only the exceptions reach the operator
  • + The map is documented; bus-factor of one removed
02 How the after-map gets built

An automated flow is only as good as the map underneath it, and most finance workflows have never been written down. The after-map is recovered from artifacts the team already produces, then turned into a diagram the business can read and a machine can run. The discovery method is proprietary and is not shown; the chain below is the shape, not the recipe.

Mailbox AI mapping Shared map Extract Automation

Start from the mailbox the team already uses. A mapping pass reconstructs the real steps and handoffs and writes them into one shared map. From that map the concrete steps are extracted, and only then is the automation built against it. Each process is connected once, with the owner's consent, and no new software is forced on the team.

// judgement, not a tool catalogue

What I reach for, and why it fits.

There are dozens of process-mining tools. Most assume a full event log, admin rights, and a six-figure budget. A real finance desk has none of those. These three approaches earn their place because they work inside the constraints.

interview-led

Compressed current-state assessment

Talk to each person on a consistent five-column lens, then map what they actually do against what the system thinks they do.

Why it fits: needs no prior documentation and no software install. Works when the only event log is people's memory.

metadata-light

Hand-over-of-work mapping

Read recurring email threads and shared files by counterparty and subject to recover who really hands off to whom, no surveillance agent required.

Why it fits: the one quantitative artifact you can build from data that already exists, in each person's own scope.

already on the desk

Spreadsheet step extraction

The query steps inside recurring workbooks are machine-readable documentation of the real data flow. Read them and you get the lineage for free.

Why it fits: the highest-yield evidence in a finance environment, hidden in the files people already maintain.

ruled out on purpose
  • Enterprise mining suites. They need an admin connector and a budget a single desk does not have, and they only see the 30 to 50% of work inside the core system.
  • Desktop task-mining agents. Practitioners report them as fragile, intrusive, and hard to draw insight from. Wrong fit for a small, trust-sensitive team.
  • Pure log-based algorithms. Elegant, but they assume a clean event log that most finance processes simply do not produce yet.
effort to deploy → value to the business → high value / low effort interview mapping spreadsheet steps handover map mining suite task agent
// the call I make

I optimise for value the desk feels, at effort the desk can carry.

The top-left quadrant is where a finance team gets real workflow improvement without a procurement cycle or an IT mandate. That is where I work.

Teal: low-effort, high-value methods I deploy in days.
Amber: high-effort methods that rarely pay back at this scale.
// shipped, running in a regulated environment

Systems already doing the work.

Patterns built on top of a fund-controlling desk and reused across entities. Each one runs in production today; here is what it does and how it works.

Automated regulatory notes

order-of-magnitude

Reads a trial balance, identifies the entity type, and drafts disclosure notes in the firm's chosen style: tables, movement schedules, narrative.

skillsWord generationLuxGAAP

Annual accounts quality review

pre-audit

Takes the accounts as PDF and the source reconciliation as Excel, checks every cross-reference between statements and notes, and flags inconsistencies before the auditor finds them.

PDF parsingExcelQA automation

Share register anonymisation

hours to minutes

Aggregates raw share-register data into net positions per shareholder per class, applies rule-based redaction, and produces print-ready Excel and PDF for KYC and AML reviews.

Data aggregationComplianceMulti-format

Mailbox intelligence

daily

Triages email and meeting-transcript backlogs into a prioritised morning briefing: unanswered messages, follow-ups, and half-finished commitments by workstream.

M365OutlookTriage

If any of this maps to a process on your desk, say hi.

Happy to talk about workflow discovery, process mapping, and where automation actually holds up in a regulated finance environment.