AI workflows, checked on real cases. Berlin.
The one workflow that slows you most, automated with AI, proven on your real cases.
I help operations teams take hours of manual work out of repeating chains: documents, orders, reconciliations, email. No new hires, no heavy software, no blind experiments.
14 financial institutions since 20249 markets, from Germany to South-East Asia3 working languages: English, German, Russian
Test run on 10 cases
example- 01Order, 3-page PDF
- 02Supplier invoice
- 03Email with the order attached
- 04Low-quality scan
- 05Order, PDF
- 06Complaint by email
- 07Order with handwritten changes
- 08Invoice with two VAT rates
- 09Order in Excel
- 10Email in German
matches the manual result passed to a person
The employee's answers on cases 04 and 07 are written down as rules. The next cases like them, the system handles itself.
An example protocol. Yours is built on your own cases.
≈ 1 quarter
No payback, no build
Before I build, I calculate what the change saves. If it does not pay back in about a quarter, that part waits. No automation for its own sake.
your rules
You set the rules
Unclear cases are never guessed: they go to your people. Each answer becomes a written rule you can read and change.
10 of your cases
Your cases, not slides
I test the solution on 10 of your files or requests before any big build.
Who this is for
Sound familiar?
One workflow means one job that repeats dozens of times a week: a request comes in, someone reads it, checks it and types it into a system. If one of these sounds like you, start there: one workflow, built by me or with your team.
More work, same team.Your people spend hours retyping and checking. New work keeps coming, and nobody suitable applies.
Where would AI actually help?Everyone talks about AI, but nobody tells you which task in your business it takes over.
ChatGPT yes, results no.Your team uses ChatGPT, but the answers vary and the work has not changed.
Your team should learn it too.Others in your trade already use AI. You want your people to learn it on your real work, not in a course.
Suited to teams from about 10 people, or to a workflow with a few hundred cases a week. If yours is smaller or not on the list, I will say so on the first call.
What it looks like
Today, and after.
The same picture at a property manager, a law firm, a practice, a logistics company or a trading company. Only the words change: a tenant's email, a client's file, a patient's request, a PDF order.
Can we take on more?
New work comes in, but the team is full. Saying yes means hiring, and nobody suitable applies.
→ youThe system reads and enters each request; a team member checks it. The same team takes on more.
✓ system, team member checksDoes my team get better?
Your good people spend the day retyping. Nobody has time to learn AI.
→ youThey check and decide instead of typing. They work with AI every day, on your real work.
→ your team, with AIDo I know what goes out?
Only you know how the odd cases are handled. When you are away, things stall or go wrong.
→ youNothing unclear goes out on a guess: a person decides. Your rules are written down, and your team can read them.
✓ a person decidesSupport is not repair. It keeps your rules alive: the system knows how you work today, not how you worked on handover day.
How I work
Three steps. See each result, then decide.
No pig in a poke: you see each step work before you pay for the next. From the first call to a running system takes about seven weeks.
Free process check
30 minutes
Together we pick the one repeating job that eats the most hours. I describe it on one page: where the time is lost. You get a straight answer on whether a test run is worth it.
Decision a test run, or stop
Test run on 10 of your cases
1–2 weeks
I run 10 of your real cases through a working version, each result next to the manual one. You get my written recommendation: how often it was right, where it fits your process, when it pays back.
Payback rule
If the numbers do not show payback within about a quarter, my recommendation is not now and you pay nothing beyond the €500. You decide.
Decision build · fix first · not now, and you decide
Into your tools, then you teach it new cases
weeks 3–6
I build it into the programs you already use. At first a team member confirms every result; later clear cases go through alone, unclear ones stay with your team.
€1,250 when it works on 30+ of your cases (week 4), €1,250 after two weeks live in your tools (week 6)
Support, €250 a month: your business changes, and each new case you decide becomes a new rule. Cancel any time.
Once a month: the share of cases the system handled without a person, last month and now.
Decision you accept each milestone, you renew support month by month
The €500 comes off the build if it starts within 30 days.
Your team should learn it too? Instead of step 3: a two-day workshop for 2 to 4 people, €1,000. I lead, your team builds it, a week of support is included.
Boundaries
What is included and what is not
Always included
- Checking on real examples from your work
- Acceptance criteria agreed before we build; you check every checkpoint yourself
- A data-processing agreement (AVV) before the first file; where the data run and which model provider is involved is put in writing
- Integration into the tools you already use
- Rules updated from the cases we review, for as long as the ongoing support runs
Not included
- Pilots and R&D without test cases and metrics
- Resale of software and licences
- AI-Act compliance paperwork
- A new process or new features after go-live: that is a separate build
Professional secrecy
Law firms, tax advisers, medical practices: before the first file we sign the written undertaking your profession requires. In medicine we work only on administration and scheduling, never treatment.
Evidence
Three of fifteen cases, anonymised.
The method comes from banks. It works the same at ten people or a hundred. Client names stay out until there is a written yes.
US, land surveying
150,000
zoning rules the system searches for the answer
The largest land-surveying firm in the US. Inbound letters and drawings.
Letters from builders arrive in every format. The system sorts them, drafts the drawings and looks up the answers in 150,000 zoning rules. Confident cases pass on their own, doubtful ones go to a person.
Engineers no longer search the rules by hand. In early production.
early productionCentral Asia, bank
< 1 min
per document, target accuracy 85–95 %
A bank, Central Asia. Contract documents read into the system.
Clerks used to retype data from currency-control contracts by hand, slowly and with mistakes. The system now reads them: target accuracy 85 to 95 percent, under a minute per document.
The bank's operations team no longer retypes contracts. In production.
in productionVietnam, bank
232 h
of manual work a month, counted before the build
A bank, Vietnam. Regulatory reporting.
A discovery workshop put a number on the manual work: 232 hours a month. Nine gaps between the requirements and the planned architecture were found before anything was built.
The bank decided with the numbers in hand, before spending on a build.
diagnosticThe person
One person, not an agency.
Since 2024: AI agents and their quality systems for 14 financial institutions in 9 markets.
Before that: partner at an AI consultancy, six years at Microsoft with banks.
Claude Certified Associate (Anthropic, 2026). Written specifications and tests from real examples, never demos.
The rule I work by
If the numbers do not show the investment coming back within about a quarter, I say it straight: do not build.
Next step
Bring one workflow. Leave with a number and a verdict.
Thirty minutes, free. You name one workflow; I tell you whether a closer look is worth it.