From noise to system.

A consumer retail business with its own delivery decides every day what to restock, which supplier to buy from and what to charge for delivery. E-commerce platforms solve the shop window; everything else is left out.

Stock, supplier costs, driver settlements and real profitability end up in a spreadsheet someone updates when they can, and in mental arithmetic that is almost always optimistic.

A system where selling and running the business are the same thing. The order comes in through the store, decrements stock atomically, is charged through MercadoPago, issues an electronic invoice automatically on payment, is assigned to a driver and lands on an accounting dashboard.

Around it, the support stack: WhatsApp wired into a unified inbox and an AI agent answering from the live catalogue and current prices.

Verification was a self-run audit in seven phases across 92 acceptance criteria: documentation inventory, contrast between what was documented and what the code actually does, currency of every dependency, test plan, automated end-to-end execution with Playwright, security review and verified remediation.

It surfaced faults that never show up in a demo: a collaborator could mint themselves an administrator account; the brand-colour field allowed code injection across every public page; the catalogue could be manipulated so the AI agent quoted false prices; and under concurrent orders the AI spend cap was exceeded by 150%. Every fix shipped with its regression test.

Nothing is marked “working” without executed evidence. What was not tested is delivered as unverified, never as approved.

Handling direct debits took the administration team around five hours, twice a month. Several spreadsheets had to be cross-checked by hand and the identifiers for every point of sale corrected one by one, because the bank matches each debit on an exact combination of tax ID, bank account and point of sale: a single customer changing accounts or ownership was enough to break the match.

With a hundred customers that was an afternoon. With more than two thousand it was some 120 hours a year — three full working weeks, every year, spent reconciling spreadsheets.

A web application where the administrator uploads the spreadsheets and, two minutes later, receives by email the files ready to load into the bank’s system and process the collection.

What made the replacement possible was not the coding: it was measuring first what the task actually cost. The problem was never the administrator or the spreadsheet — it was a procedure designed for a company size that had ceased to exist.

The task ceased to exist. In time, some 120 hours a year freed up; in money, a large share of one person’s salary devoted to a single task that added nothing to the business.