The cases (55)
Grouped by impact. ★ = the strongest.
Time back
From transcript to ticket
Five to ten minutes after the meeting, minutes and tickets are on the board, no duplicates, correct parent ticket.
Inbox: 9,748 → 5,277
Inbox down from 9,748 to 5,277 e-mails; the target is permanently below 20.
Knowledge becomes wiki pages and decks
A wiki page including diagram in about five minutes; across decks, wiki and document store, around 80–90% less time per artefact.
Transcribed locally, audio never leaves the device
A cleaned transcript with timestamps shortly after the meeting ends; an eight-hour session day is transcribed in under two hours of compute — the audio stays on the device.
One meeting, several versions, one to-do list
Distribution-ready documents and a hand-over-ready to-do list across all meetings — from a two-day off-site a good 20 to-dos with owner, date and timestamp, plus minutes and a detailed analysis for almost 30 participants.
Paperless mail with a burden of proof
1,423 documents in the archive, paper backlog from 551 to 0, the archive's inbox from 466 to 0.
Morning brief: a situation report every morning
One glance at the phone over the first coffee replaces opening five programs.
Ticket backlog: an hour, not a day
Backlog built in about an hour instead of a working day — a good four fifths less time; around 60 linked tickets in the first run, one epic round across around 20 products in about an hour.
Birthday greetings on time, weekends included
Punctual, personal greetings even on weekends and holidays; plus a morning overview of what each meeting needs.
Email drafts that learn my voice
Drafts that need less rework with every mail sent; the learning comes from comparing with the real result.
Trips that grow on the map
24 stops without gaps from 11 Aug to 12 Sep 2026 in the data core — the map draws itself.
Waste-collection calendar from a reverse-engineered app interface
56 collection dates up to year-end in the family calendar, backed by 19 tests.
Reading the meter from a photo
38 readings from seven meters as one series from 2001 to 2026.
Money back
Three years of solar credits, overnight
Reminder out at 22:11, written commitment overnight at 08:29; roughly €1,100 paid out later.
Tax return, the AI as devil's advocate
356 items rated one by one, three genuine errors caught before filing, submitted five days ahead of the deadline.
Tenants' utility statement on request
2025 statement finalised on 7 Aug 2026, delivered, credit of €177.13 transferred to the tenants.
A ledger that matches the bank
A mis-posting of roughly €56,500 prevented; 16,138 ledger rows, 15,976 of them identical to the banking records.
Rental-car negotiation with an audited checkout page
Instead of roughly €11,350, a local provider at roughly €3,886 — and a discrepancy of roughly €720 on the checkout page caught before the click.
Warranty claim enforced: commitment within two days
Two days after the letter, the retailer's managing director committed to the replacement without reservation, return label included.
Purchase compass: three phases, live price
Documented single cases: a document scanner for €345.98, a notebook in the newer model for €997 instead of the older one.
Annual fixed-cost review with live market check
Two parallel subscriptions with the same provider discovered, hidden in the bank statements for years.
Photo in, finished classified ad out
Four pallets sold for €40; €95 total proceeds in the register by mid-September.
Risk removed
Owners' meeting prepared like a lawyer would
Discovered: 83 % of the 2027 service-charge increase is a heating advance, not the lift or the grounds.
Meeting versus project knowledge, with trust light
An analysis note with evidence and a reasoned trust level — the reliability of an AI result becomes measurable instead of felt.
Out of the cloud, with proof
Roughly 1.76 TB retrieved with checksum proof across 311,462 files; full restore of one backup track 1,623 of 1,623 files byte-identical.
Vaccination records digitised, doctor's visit prepared
Five booklets in the register, twelve entries checked, four of the doctor's statements verified against the official recommendations.
Knowledge unlocked
Spec to running app in eight person-days
Eight functional and domain building blocks with around 160 input fields in about eight person-days; requirements become visible as running features immediately and can be reviewed by experts right away.
Legacy system as knowledge base: 250 requirements
Agents derive requirements from the legacy base and reconcile them with the team's current state; all content is draft until the expert review.
41,000 e-mails as a biographical source
41,431 e-mails 100 % dispositioned; the world-trip circular letters are complete as raw material for the book.
Domain and system knowledge in one place
Around 20 domain knowledge sources mapped, around two thirds of them directly usable by AI; first knowledge maps exist for several products, and the first productive domain query returned two relevant supreme-court rulings, cross-checked and filed with sources.
A council of AI models argues
In the first test case the council delivered 18 clustered objections against 12 from the single model.
Lost GPS data: forensics and reconstruction
Roughly 5,786 kilometres across 26 states reconstructed — labelled as reconstruction, not as measurement.
World map of every trip, tracks replayable
112 countries and 47 routes on one map that loads in 3.7 seconds on a slow mobile network; 954 of 954 browser probes green.
Family & joy
Art on the wall that curates itself
31 works on the device; a new picture appears within ten minutes of the request.
Colouring booklet and animal app for safari
Three booklets in the luggage, 124 species in the app — the next booklet is one command instead of six misprints.
Grandmother's fairy tales on audio figures
99 recordings transcribed, seven packs, the first Tonie plays in the kids' room.
School letters with traffic light and actions
154 parent letters captured read-only; in operation roughly 1.6 new letters a week arrive with a traffic light.
Paintings that move
Three finished 4K loops hang on the device, seam verified bit-exactly.
System builds system
Moving the server while it runs
35 services on one server with eight virtual guests; the document archive moved with roughly two minutes of downtime.
One twelfth of the tokens via CLI
The same board query with 40 cards costs around 1,400 instead of around 17,000 input tokens — about one twelfth for the same information.
Two months to a team project
Two months from private experiment to kick-off; after the first few months there is a Business Brain, first use cases and change support.
A system that corrects itself
16 logged self-corrections in a single day; the consistency check went from 28 findings to 0.
Knowledge base checks itself, no language model
Traceable knowledge that is checked daily — around 1,650 existing notes across eight brains were back-filled with source and check date for it.
Many writers, no overwriting
One shared, versioned knowledge base — over 3,000 commits from around two dozen identities in the team brain since the start.
Trace instead of gut feeling
Measurable effort instead of gut feeling — the baseline showed zero percent navigation: the agents knew the paths from context and did not follow the knowledge graph.
Build once, use everywhere
Consistent rules and tools across nine workspaces — the second reconciliation run finished without a single file deviation.
Steering
Critical risks current within a week
Around 100 active risks checked, around 90 tickets commented with proposals; within a week the owners had updated or closed a good half of them, and a good three quarters of the critical red risks.
1,700 items in one management view
Around 1,700 items consolidated into one delivery status, around 130 epics and features measurable within it — running after a good week, cross-source for the first time.
Weekly status from around 15 contributions
A weekly mail draft with embedded dashboard excerpts for three product lines — transparency without an extra meeting.
Proving AI value continuously
An evidenced portfolio instead of individual success stories — almost 30 use cases, a good 40 percent of them mature.
Enabling others
14 interviews, analysed anonymously, change personas derived
14 interviews across roles condensed into change personas with concrete change needs in a short time — with little resource effort.
Ask Mathias: a digital twin
A sparring partner that pushes back and a twin that pre-sorts questions — wired as a pilot in the Business Brain; the test series is still open.
Guided setup: four fifths was waiting
A working workspace; the target is setup in under an hour — the pilot took almost seven, around four fifths of it waiting for individual approvals.
Passing the setup on as starter kit
Kit with 114 files, all three gates at zero; first update run: 45 changed sources, 30 adopted automatically.
Enablement: open review, office hours, FAQ agent
A recurring rhythm for transparency and enablement — and a lower bar for asking.