Confidence, not magic

Currency

Categorising the bank feed is usually the first job a finance team hands to AI. The model reads each line (a cloud invoice from Luxembourg, a ride in Mexico City, a VAT payment in Madrid), proposes a ledger account and says how sure it is. Mostly it is right, and mostly it knows it.

The edges are where the work is. Above the bar you set, lines post straight to the ledger; below it, a person looks. Where that bar sits is a control decision, and it belongs to the CFO, not the vendor.

One day’s bank feed · 52 lines

06:02NIMBUS CLOUD EMEA SARL LU−€3,412

Hostingconfidence 0.98

Auto-postedStraight to the ledger

For review 19Auto-posted 33

Right, posted
Right, in review
Wrong, posted
Wrong, caught in review

Drag the bar, or tap a dot to read the line.

Stance

Confidence bar0.85

At or above it, a line posts. Below it, a person looks.

Learn from corrections

Replay the feed after one review cycle.

Verdict

At 0.85 the model posts 63% of the feed unattended and lets two errors through: a personal hotel bill booked to Travel, and a supplier credit note booked as revenue. Raise the bar to 0.94 to stop them, and review fourteen more lines a day.

Auto-posted

63%

33 of 52 lines, untouched

Misposted

€1,726 on the wrong account

For review

≈ 29 min a day, 10 h a month

Precision

94%

of auto-posted lines on the right account

The trade-off

Every bar from 1.00 down to 0.50. Each step up is a wrong line joining the ledger; the dot is where you are.

Errors let through

Share of the feed auto-posted →

The queue

Lines queued for human review, with the model’s suggested account and confidence.
Line	Amount	Suggested	Conf.
Marketplace EU −€1,390MARKETPLACE EU*7Q2 LUX	−€1,390	Office & telecoms Reviewer: IT equipment	0.58
Marketplace EU −€64.00MARKETPLACE EU*9D1 LUX	−€64.00	Office & telecoms	0.61
O Tasco Lisboa −€128O TASCO LISBOA	−€128	Meals & entertainment Reviewer: Employee recharges	0.62
FX conversion −€48,000FX CONVERSION EUR>GBP	−€48,000	FX differences Reviewer: Internal transfers	0.64
Courier Express −€96.00COURIER EXPRESS INTL	−€96.00	Office & telecoms	0.66
Payroll withholding −€12,480TAX AGENCY WITHHOLDING Q3 ES	−€12,480	VAT control Reviewer: Withholding tax payable	0.67
Intercompany transfer −€250,000INTERCO TRANSFER ES SL>UK LTD	−€250,000	Intercompany	0.68
Contractor platform −€8,640CONTRACTOR PLATFORM PAYOUT	−€8,640	Payroll & payroll taxes Reviewer: Contractors	0.69
Cafe Korn −€38.50CAFE KORN BERLIN	−€38.50	Meals & entertainment	0.71
Workloft Events −€4,200WORKLOFT EVENTS BERLIN	−€4,200	Rent & facilities Reviewer: Marketing	0.72
Mueller Anlagenbau +€42,800MUELLER ANLAGENBAU GMBH RE 2291	+€42,800	Revenue Reviewer: Trade receivables	0.74
StreamCo −€13.99STREAMCO PREMIUM	−€13.99	Software Reviewer: Employee recharges	0.76
Card FX fee −€21.40CARD FOREIGN TXN FEE	−€21.40	Bank & card fees	0.78
Relay CRM −€14,400RELAY CRM ANNUAL SEATS	−€14,400	Software Reviewer: Prepayments	0.79
LearnHub −€480LEARNHUB TEAM PLAN MONTHLY	−€480	Software	0.80
Contractor platform −€259CONTRACTOR PLATFORM FEE	−€259	Contractors	0.81
Taberna del Puerto −€212TABERNA DEL PUERTO MADRID	−€212	Meals & entertainment	0.82
Lex Arbor (Kestrel) −€18,500LEX ARBOR LLP PROJECT KESTREL	−€18,500	Legal & professional Reviewer: Acquisition costs	0.83
Schoon Cleaning −€690SCHOON BV CLEANING AMSTERDAM	−€690	Rent & facilities	0.84

Notes on the model

What it shows

Automation and leakage move together. Every notch you lower the bar buys unattended postings and lets more mistakes into the ledger.
Confidence is not correctness. The model is surest about errors that look like routine spend: a hotel reads as Travel whether or not the trip was business.
Feedback only fixes what a person sees. Corrections lift the merchants that were reviewed; lines that slipped through teach the model nothing.

How it works

Fifty-two synthetic lines from one day of a group’s bank feeds across Europe, the UK and the Americas, each with the model’s account, the correct account and a confidence score.
Lines at or above the bar auto-post; the rest queue for review at 1.5 minutes a line. Precision is the share of auto-posted lines on the right account.
Learning replays the feed after one review cycle: every merchant a reviewer touched comes back on the right account, closing 40% of the gap to certainty.

What it leaves out

Materiality. A coffee and an acquisition invoice count as one error each; in practice anything above a value threshold is reviewed whatever the confidence.
Scale. Real models are calibrated on thousands of lines; one day’s sample makes each error visible and each step on the curve large.
Reviewer fatigue. Review is assumed perfect. Under volume, people approve suggestions too, which is why the queue must stay small enough to read.

Next experiment · 03

Pipeline to plan

A driver-based rolling forecast built from pipeline and order book, in three cases at once.

Run the scenarios

Using this on a real decision?

Talk it through

The defaults are illustrative. The inputs that decide the answer are the ones you only know from inside the business.

Get in touch