At Debitus, an accounting firm in Porto, bank reconciliation ran on stationery.
Client statements arrived as PDFs — most of them scans of paper — and were printed again so an accountant could work down the page with two felt-tip pens: green for movements that matched the ledger, red for those that needed chasing. When we started building software to take over the job and asked for the firm’s reconciliation rules, nobody could recite a complete set. There was no set to recite. The rules lived in the pens, and in the hands that held them. Also, in the minds of those involved in the process.
The motivation was the obvious one: stop losing skilled time to a manual ritual. Machines were supposed to have arranged this decades ago — banks have exchanged machine-readable statements since the SWIFT MT940, and the ISO 20022 successor was published in 2004. What reached me instead had made a stranger journey: data born structured inside a core banking system, rendered to PDF or printed, filed, scanned, and delivered — at which point I paid a machine-learning service, per page, to undo the work of a printer.
Where fifty hours went
The build took roughly fifty hours. None of them were spent on models. Azure Document Intelligence ships pre-trained models for OCR and table extraction; the intelligence, in the marketing sense, was rented by the page and worked out of the box — on clean documents. The chart shows where the effort actually sat.
The largest share, about a third, went on the interface. This surprised me and should not have. The movements a machine matches confidently never reach a human; the ones that do are, by construction, the cases the rules fail to decide. The screen that handles them has to make an exception legible in seconds — to people whose previous interface was a highlighter, which is a formidable competitor on usability.
Close to another third went on making pre-trained extraction generalise across formats. Every bank lays out its statement differently: columns in different orders, balances interrupting the transaction list, descriptions that wrap unpredictably, and scan quality inherited from whatever the client’s photocopier was feeling that day. A fifth went on the rules themselves — extracting them from the people who applied them, one worked example at a time. “We match on amount and date” turned out to carry unstated tolerances, exceptions for particular counterparties, and habits nobody remembered adopting. We deliberately did not hard-code what emerged: the rules sit as parameters the firm can edit, because the rulebook belongs to Debitus, and adjusting a matching tolerance should not require a software release, or me. The remainder went on testing, which mostly meant discovering which unstated rules I had still missed.
The doors it opens
The immediate benefit is the one Debitus asked for: the time saved. Reconciliation was a recurring tax on the team’s attention, paid in printing, ticking and chasing, and the module removes most of it. On the argument I have been making in this series, those savings will not stay as margin for long — reconciliation is compliance work, and in a price-sensitive market automated compliance competes down to the client’s fee. The interesting question is what the automation leaves behind, and there the answers kept arriving after the build was done.
The first thing it leaves is the rulebook. Debitus now owns an explicit, parameterised statement of how it decides that a movement is settled — the first in the firm’s history. It does quiet work. New staff learn the firm’s judgment from a screen instead of a shoulder. Every rule is applied the same way twice. And when an auditor, or a client, asks why a movement was treated as settled, there is an answer that does not depend on who held the pen that month. The process can now vouch for itself, which is assurance in its plainest form.
The second is stranger, and I did not see it coming: the process can now see itself. The manual version was unmeasurable by construction. Nobody knew how long reconciliation really took, where it stalled, what share of movements needed human hands, or whether one bank generated three times the exceptions of the others, because a felt-tip pen leaves no telemetry.
The automated version time-stamps every step as a by-product of running. Auto-match rates, time from exception to resolution, throughput by week, exception rates by bank: the raw material of statistical process control, the method factories have trusted since Walter Shewhart drew the first control chart at Bell Labs in 1924 and W. Edwards Deming carried it to post-war Japan.
A century of industrial quality management passed the office by, for a simple reason: clerical work never produced the measurements the method feeds on. Marked paper in a folder is not a dataset. These time stamps are the first time this particular process has generated the telemetry a production line has had for a hundred years.
The third door follows from the second. Once a process can see itself and its rules live as parameters, action can be wired in. An exception that fits a known pattern does not need to wait for a human to notice it: the system can request the missing document itself, route the case to the right desk, escalate whatever ages past a threshold — and do it at nine in the evening, with a patience no accountant possesses at month-end. The loop the management textbooks draw, measure–decide–act, closes inside the process, with people kept at the decisions that deserve a person.
A book I have coming out soon argues that managing is deciding, and that you cannot decide about what you cannot observe; this module is that argument met in the wild. The manual process was slow, but its deeper defect was invisibility. The least advertised feature of the software is that it made a piece of the firm observable, and therefore manageable.
The fourth door points outward. The same governed history — every movement structured, every match and exception recorded — has analytical uses nobody at Debitus has touched yet: which exceptions recur, and in which months; which account has begun to show the small irregularities that tend to precede larger trouble. Each is the opening line of an advisory conversation, and that work does not compete down to the fee, because it rests on data only this firm holds, about clients only this firm knows.
None of this was in the specification, mine or theirs. Debitus commissioned a way to stop wasting hours; the hours will duly come back, and they will be the least of it. The build was harder than I expected — fifty hours of formats, rules, screens and people, wrapped around an “AI” rented by the page — and the difficulty sat exactly where the loud version of the AI conversation never looks. The pens, though, were the real finding. For years they had been recording, one stroke at a time, a process the firm had never needed to write down — and measuring nothing, remembering nothing, chasing nobody. The software made that process explicit, observable, and able to act on its own findings. What Debitus used to file away as marked paper, it now accumulates as an asset.
The green pen has earned its retirement 😃.
If you’re looking to automate your firm’s document workflows with AI and Power Platform, see how we can help.
