I find the expensive work hidden around enterprise systems and turn it into software. Treasury, payments, finance ops. Free tools: https://xitter.cf/t.co/oN5yAJJJXe
Amsterdam, Netherlands
Joined March 2021
- Tweets3.2K
- Following70
- Followers3.8K
- Likes14.2K
Pinned Tweet
1/ For most of my career, understanding a business problem and building software to solve it were two different jobs. AI is collapsing that gap. The interesting consequence isn’t faster code.
1
1
219
Rule: every article ships with something you can use, or it waits.
An article I can’t turn into a tool or a checklist usually means I have an opinion, not a mechanism.
The tool is where opinions get tested.
30
A quiet AI win in treasury: the investigation request to the bank.
Reference, value date, amount, beneficiary, what they report, what you ask. Half an hour of copying from four screens.
Draft it from the record in seconds. A person checks and sends.
57
New article: Exception ageing, the number reconciliation teams don’t report.
Why count hides what age reveals, and three one-week changes.
Link below.
47
Product signal #11: the write-off that happens every month.
Same cents, same counterparties, same difference account. Nobody investigates; it’s how the month closes.
A tolerance rule that exists in practice and not in the system.
1
58
How a week of posts comes from one article:
Article written in full by Tuesday. Thread Wednesday = the spine. Product signals = its examples. AI posts = “what would a model do here?”
Nothing is written fresh at 08:30.
1
67
The key design decision in AI triage: what happens when the model isn’t sure.
Lazy: guess, attach a confidence score nobody reads.
Useful: an “unsure” outcome, routed to a named person, with one line on what it couldn’t determine.
The second one is deployable.
1
65
1/ Reconciliation teams report how many items are open.
The number that matters is how old they are. ↓
1
1
75
5/ Three changes, one week each:
Report by age bucket, not by count.
Everything past 14 days gets a named owner and a date.
Count the reasons items age. The top reason is your next rule.
1
9
Product signal #10: the file someone resends by hand.
Bank rejects at 15:40. Someone fixes the field, renames the file, uploads via the portal. Cut-off in 20 minutes.
The system has a path for sending and no path for rejection.
1
58
Column audit stores nothing. Day one, I feel the cost: I can’t see which rule fires wrong.
I’d decide the same again. The people it’s for check one thing first: do my column names leave the browser?
Rules improve the slow way: you tell me.
1
69
The expert’s edge on an exception isn’t knowledge. It’s memory: “this looks like March.”
Give the system that memory: for each open item, the 5 most similar resolved ones and how they were closed.
Retrieval, not generation. The person still decides.
1
70
Every exception path has four steps: detect, enrich, decide, feed back.
Most systems do the first. People do the second and third by hand. The fourth almost never exists.
That’s why the same exception returns next month.
53
Product signal #9: the person everyone asks.
Payment bounced? Balances disagree? Same person, every time. Not their job. They just know.
Exception routing lives in one head. Twenty rows of a table would hold most of it.
54
Column audit launch day. Two counters: audits run, columns audited. Nothing else, and never the content.
Kill criterion, set before coding: 50 audits in 7 days.
Number next Monday, whatever it is.
108
The first AI job in an exception queue isn’t clearing it. It’s explaining it.
For each cleared item: one sentence, why was this manual? Then cluster.
The queue becomes a backlog with frequencies. The top cluster is your next matching rule.
1
83