https://hapax.fi/ https://hapax.fi/advisory.html https://hapax.fi/training.html https://hapax.fi/counsel.html https://hapax.fi/notes.html https://hapax.fi/note-anthropic-books.html https://hapax.fi/note-staff-ai-rules.html https://hapax.fi/note-eu-data.html https://hapax.fi/trust.html https://hapax.fi/work.html https://hapax.fi/shapes.html https://hapax.fi/uncertainty.html https://hapax.fi/case-erp-cleanup.html https://hapax.fi/case-invoices.html https://hapax.fi/case-contracts.html https://hapax.fi/case-board-reporting.html https://hapax.fi/case-anomaly-detection.html https://hapax.fi/case-churn.html https://hapax.fi/case-tenders.html https://hapax.fi/case-inbox.html https://hapax.fi/case-training-outcomes.html https://hapax.fi/case-talent-screening.html https://hapax.fi/case-skills-library.html https://hapax.fi/film-erp-cleanup.html https://hapax.fi/films/case-erp-cleanup.poster.jpg The automation that retires itself On 240 fabricated month-end reports from a fictional SAP system, with every correct value recorded in advance, a language model parsed each one while a second system learned the repeating formats and wrote ordinary code for them. 184 of 240 ended up on that code, at an estimated two fifths of the cost. https://hapax.fi/films/case-erp-cleanup.mp4 96 2026-10-07 https://hapax.fi/film-board-reporting.html https://hapax.fi/films/case-board-reporting.poster.jpg The board pack that answers back Two years of fabricated monthly financials for a fictional Nordic manufacturer, with four scenarios hidden in the data and the correct findings written first. Three independent AI analyses found all four, with no false positives and no wrong figures across thirty findings. The strongest evidence in one scenario went unexamined. https://hapax.fi/films/case-board-reporting.mp4 120 2026-10-07 https://hapax.fi/film-anomaly-detection.html https://hapax.fi/films/case-anomaly-detection.poster.jpg The anomaly nobody caught We built a 50,000-row purchase ledger, hid 87 irregularities in it, and screened it three ways. Rule-based checks caught 48. A statistical anomaly detector caught none. The slow vendor price drift at the centre of the design was missed by every layer. A follow-up report catches three freshly planted drifts on new data but cannot see the vendor that prompted it. https://hapax.fi/films/case-anomaly-detection.mp4 121 2026-10-07 https://hapax.fi/film-tenders.html https://hapax.fi/films/case-tenders.poster.jpg Sixteen tenders, answered two ways Sixteen invented tenders and 510 requirements, answered two ways. Fourteen simulated people with a chatbot produced per-tender medians from 64.3% to 93.3%. A workflow run once per tender answered every requirement on 11 of 14 and refused both tenders the firm cannot win. The only three requirements nobody found were mandatory forms mentioned once inside an annex. https://hapax.fi/films/case-tenders.mp4 114 2026-10-08 https://hapax.fi/film-churn.html https://hapax.fi/films/case-churn.poster.jpg The churn a usage model can’t see On 500 fabricated accounts with 3,777 support tickets and a known answer key, a usage model catches the accounts whose logins drop before they cancel. A language model reads the tickets, finds nine more genuine churners in the top hundred, and adds €1.14M in annual revenue the usage data misses. https://hapax.fi/films/case-churn.mp4 117 2026-10-08 https://hapax.fi/film-inbox.html https://hapax.fi/films/case-inbox.poster.jpg One hundred and fifty emails, three gates An invented equipment-rental firm in Oulu, a 30-clause policy handbook and 150 synthetic customer emails. A pipeline classifies each one, drafts a reply with clause citations, checks the draft against three rules and routes the result. 122 were sent automatically without a single rule violation. All seven safety and data-request emails were escalated. Six emails asking for more information were misclassified as routine. Three of six expectations were met. https://hapax.fi/films/case-inbox.mp4 84 2026-10-08 https://hapax.fi/film-training-outcomes.html https://hapax.fi/films/case-training-outcomes.poster.jpg Forty learners, three instruments Forty simulated learners sat a two-day programme with their true improvement set beforehand. The assessment recovered the overall learning effect to within 0.01 of the planted 0.13 and caught six of the seven people whose confidence rose while their competence did not. The industry-standard feedback form correlated with real learning at r = 0.10. A 94%-reliable scorer flips two of six verdicts. The cohort is simulated, so what is tested is the instrument, not the training. https://hapax.fi/films/case-training-outcomes.mp4 97 2026-10-08 https://hapax.fi/film-talent-screening.html https://hapax.fi/films/case-talent-screening.poster.jpg Two hundred and forty applications, screened four ways 240 synthetic applications for a signal-processing role at a fictional instruments maker in Oulu, screened four ways and checked against an answer key no screening arm could see. A workflow read every application against twelve requirements and returned evidence dossiers with verbatim quotes but no ranking or verdict. It caught 14 of 15 credential inflations and identified 19 of 22 keyword stuffers. Four of ten predictions were met. https://hapax.fi/films/case-talent-screening.mp4 116 2026-10-08 https://hapax.fi/film-skills-library.html https://hapax.fi/films/case-skills-library.poster.jpg Sixty questions, and what happens when you write the rules down Sixty requests to a fictional Finnish retailer, each with the correct answer written first and held out. Without written rules, an AI assistant scored 49.7%. With a single document of definitions and worked examples, 99.3%. A cheaper model with the rules scored 29 points above the main model without. Six of twelve predictions were met. https://hapax.fi/films/case-skills-library.mp4 104 2026-10-16