At 16:40 on a Thursday, a competitor bank raised its flagship savings rate by forty basis points. By 17:15 the news was in every treasurer’s inbox in the country, including at Solvenz, a mid-market lender and payments group with about 400,000 accounts. The asset-liability committee was scheduled for Monday at nine. The decision on the table was simple to state and expensive to get wrong: match the rate across the book, ignore it, or do something cleverer.

The treasurer knew exactly which numbers he needed, and he could have listed them from memory. How much has flowed out since Thursday evening, hour by hour? Out of which balance bands? Are the leavers rate-chasers with short tenure, or ten-year relationships? Is the outflow concentrated in accounts that also hold a loan with us - in which case the relationship math changes entirely? Do the wallet-settlement records show the money moving to the competitor specifically, or scattering?

He also knew, from a decade of Mondays, what asking would produce. The request goes to an analyst, who translates it into a ticket for the BI team, which is forty tickets deep. The core ledger, the transaction store, the CRM and the collections system each have a different owner, so the join itself is a small negotiation. The polished answer arrives in two to three weeks - a fine piece of work, describing a decision that was taken eleven days earlier by people arguing from anecdote. Every finance leader recognizes this loop, and the research is blunt about what it does to the profession: Deloitte found finance teams spend 41 percent of their time simply gathering and processing data before any analysis begins. The tragedy is not the wasted hours. It is that under time pressure, leadership stops asking, and a bank that stops asking prices itself by gut.

Todaytypical mid-market 40%25%15%20%

Analytics Engineillustrative target 10%10%75%

Gathering & reconciling dataAssembling reportsWaiting on BI / IT queueAnalysis & judgment

Anchors: PwC: ~30% of finance time on collecting/reconciling data; Deloitte: 41% on gathering and processing.

On the old flow, Monday’s committee would have faced two defensible-sounding options: match the forty basis points across the book - call it seven figures a year in extra interest expense, paid mostly to depositors who were never going to leave - or hold the line and accept whatever outflow came. Both are guesses wearing suits.

Friday morning, in plain language

Solvenz runs the Analytics Engine across the four stores - read-only, every generated query logged against the asker’s access scope. On Friday morning the treasurer typed his questions the way he had phrased them in his head.

Net deposit outflow since Thursday 17:00, hourly, by balance band and account tenure. The chart landed in seconds and told a very specific story: the outflow was real but narrow. Under three percent of accounts were moving, almost all above €100,000 in balance, almost all under three years’ tenure - and that thin slice carried close to forty percent of the departing value. The long-tenure, mid-balance book, the one a blanket rate match would have repriced at enormous cost, had barely moved.

Of the accounts that left or shrank, how many hold an active loan with us? What is their total relationship value? A second cut, thirty seconds, and the retention target list wrote itself.

By Monday the committee wasn’t debating a rate. It was reviewing a targeted counter-offer for a few thousand rate-sensitive, high-balance accounts, priced against their actual relationship value, with the book-wide match held in reserve if the following week’s data said the bleed was spreading. The following week’s data - a standing daily digest by then - said it wasn’t. The difference between that decision and the blanket match is the kind of number that funds a product team, and it was available only because the person who owned the decision could interrogate the data himself, over a weekend, without a single hop through the queue.

The deposit weekend set a precedent that outlived it. Two months later the credit committee was weighing an expansion into a borrower segment the sales team had championed for a year, on the strength of a pitch deck and two case studies. The chief risk officer spent an afternoon with the Analytics Engine instead: roll-rates and loss curves for every historical loan that resembled the target segment, cut by origination channel and vintage, cross-referenced against deposit behaviour for the borrowers Solvenz already banked. The segment survived the examination - but the channel the sales team preferred did not, its historical vintages deteriorating twice as fast as the same borrowers acquired elsewhere. The expansion launched through the other channel. That distinction was worth more than the decision itself, and under the old flow it would never have been drawn, because nobody commissions a three-week study to challenge a deck everyone already likes.

This is the part of the Analytics Engine built for the top of the house: the unrepeatable question, asked at the moment it matters, against every system at once. No committee ever asks the BI team for “hourly outflow by tenure cross-referenced with loan holdings” - not because it isn’t the right question, but because the asking costs more than the answer seems worth. Remove the cost, and better questions simply start getting asked. The institutions that industrialised exactly this reflex are compounding it: DBS Bank, whose data-and-alerting build-out Harvard Business School documented as a teaching case, expected the measured impact to clear a billion Singapore dollars in 2025. They built the capability with thousands of engineers. The capability, not the headcount, is the point.

The furniture of an ordinary week

The dramatic Friday is what sells the story; the standing layer is what runs the bank.

Solvenz’s controller used to lose the last two days of every close reconciling settlement files against ledger postings - archaeology, performed monthly, on discrepancies that were by then three weeks cold. One sentence turned it into a control: every morning at 07:00, compare yesterday’s settlement totals by scheme and currency against the ledger, and email any mismatch above €500 with the offending batch IDs. Mismatches now arrive batch-attributed and one day old. The close shortened not because anyone worked faster but because the month-end spike stopped existing.

The head of risk went after her queue the same way. Her team was triaging several hundred monitoring alerts a day with no visibility into which rules generated the noise; a single question - false-positive rate by rule over ninety days, ranked - produced the retirement list for the five worst offenders, with the evidence attached for the compliance file. In the other direction, a tripwire now texts her when confirmed-fraud volume in any merchant category runs three times its thirty-day average inside six hours: the pattern that a queue of stale alerts buries, surfaced while it is still a pattern and not a loss.

And the CFO’s board pack, three analyst-days of assembly per month, became a pinned dashboard - originations, margin, roll-rates by cohort, settlement mismatches - refreshing hourly. Directors now open it mid-month, unprompted, which has changed the texture of board questions more than anyone expected.

BI / data-team ticket queue 5–15 business days

Analyst self-serve in spreadsheets hours – days, error-prone

Analytics Engine query under a minute

Every day of latency is a day of decisions made on stale or partial numbers.

Bar lengths schematic (log-compressed). DBS Bank (HBS case, 2024): 800+ AI models; >SGD 1B measured impact expected in 2025.

Where the people went

Nobody at Solvenz lost their job to this. What changed is what the jobs are. The eleven-person finance team had been operating, by the industry’s own measurements, as a part-time query engine - the gathering, reconciling and formatting share of their week going to work a machine does better. That share collapsed, and the recovered capacity went where it embarrasses the old arrangement: the analysts now spend their time on funding-mix scenarios and cohort-level credit questions, using the Analytics Engine’s ad-hoc layer to test in an afternoon what used to be a quarter’s project. The BI team still exists and still matters - for the genuinely hard modelling - but it stopped being the tollbooth every routine question had to pass through.

That is the honest shape of the value. Transparency, because leadership sees the institution’s actual state instead of last month’s rendering of it. Awareness, because thresholds watch the data continuously and route each signal to the one person who can act. And informed decisions, because the question that decides a Monday no longer costs three weeks to ask. The rate-match that didn’t happen never shows up in any system as a saving. It was simply a better decision, made by people who could see.