Finance engineer
Your finance stack is not slow because of your team. It is slow because of systems debt nobody has inventoried.
I run AI and data systems inside a multi-entity group's finance function. I write about the problems a CFO inherits from years of accumulated systems, and what it takes, in practice, to fix them with a finance engineer in the room rather than a consultant outside it.
11
entity books, each on its own close calendar
consolidated by agents that read the ledger by API
9 of 11
books tied on the first automated pass
the other two exposed the next layer of debt
1 line
per figure on the executive page
source and sign-off beside every number, workings folded
The recurring diagnostic
The five layers of finance systems debt
Debt lower in the stack makes every layer above it harder. Most groups start at reporting, because the board can see it. The hours are usually being lost two layers down.
- 1
Records
Where is a transaction first written down?
Some entities keep a side book. Bank lines are typed in. Invoices are keyed from email.
- 2
Reconciliation
Can you prove two records of the same thing agree?
Bank recs finish ten days after month end. Intercompany differences are chased by one person who knows where to look.
- 3
Consolidation
How many books close before the group number exists?
Entities close on their own calendars. FX rates are typed into a workbook. One person does the eliminations.
- 4
Reporting
How many days until the CFO sees a P&L they trust?
The pack is pasted from exports. A challenged figure takes a day to trace. Nobody is sure which version is final.
- 5
Decision
What changed because someone read the page?
The executives ask for a figure the pack does not carry. The CFO quietly distrusts one number and cannot say why.
Every case I write up names the layer it fixed and the layer it exposed next. That second part is the one a vendor will not tell you.
Why it stays
A consultancy is paid to diagnose. Systems debt is cleared by hundreds of small builds.
An outside firm has no daily read access to the ledgers, no seat in the close, and no incentive to make itself unnecessary. It produces a roadmap. The roadmap sits next to the workbook it was meant to replace.
What clears the debt is someone in the room with read access, who ships one reversible change at a time, is present the morning it breaks, and treats every automation as a change to a control.
That is a finance engineer. Someone who can read a trial balance and a stack trace in the same hour.
How I work
Five rules for putting a machine inside a finance function
Read-only first
Every build starts by reading the ledger through its API, never by writing to it. The agent earns write access one control at a time.
Rehearse on a whole-book copy
Nothing runs against a live book until it has run clean on a full copy of last month.
Tie out before trusting
An agent that reads balances is checked against the signed figures before anyone relies on it. Nine of eleven books tying on a first pass is a finding, not a result.
Every automation is a control change
When a machine writes to a book, the control that used to sit with a person is redesigned, documented and owned, not deleted.
Nothing probabilistic in a signed figure
A language model is good at reading documents, drafting pages and spotting exceptions. It never produces a number that goes on a balance sheet.
About
I started as a commodity trader in Singapore, learned trade finance from the inside, and earned the CAIA charter to understand how capital is priced. I raised a Series A to build trade finance software and moved to California to bring it to the Americas.
Today I hold the Group Head of AI seat inside a multi-entity trading group, where the finance function is the first place the AI programme had to earn its keep. Everything on this site comes from that work, with names, vendors and systems removed.
I write here once a fortnight. The short version goes to LinkedIn.
Finance Systems Debt Scorecard
Thirty minutes with your controller. A score for each of the five layers.
Twenty questions, scored live, no deck. You get the score, the layer costing you the most hours, and what it would take to fix it. No charge. I run a few of these a month.
Group CFOs and controllers of multi-entity companies. Reply within two working days.