Blog
Field notes from work that had to run in production
This page is empty on purpose. Everything we write comes out of an engagement we actually ran, which means it appears when the work is finished and the numbers are in — not on a content calendar.
No newsletter cadence, no reposted vendor announcements.
We would rather leave this page empty than fill it with a restated vendor announcement.
Articles
Each piece states what it stands on: one of our four delivered projects, or published research and method we have not yet shipped for a client. Both are marked at the top of the article.
- OperationsSix ways an automation project quietly fails after go-liveAutomation rarely fails loudly. It degrades — in the exception queue, the silent integration, the owner who left. Six failure modes and the check that catches each.29 August 2026 · 9 min read
- StrategyFive automations we would argue against buildingA task running twelve times a year at ninety minutes is eighteen hours of labour. No build pays that back. Four more cases where the honest answer is don't.29 August 2026 · 7 min read
- OperationsYour confidence threshold is set by error cost, not savingsThe arithmetic of a document review queue, why the costly failures are the confident ones, and the checksums that decide routing better than any model can.29 August 2026 · 8 min read
- LogisticsWhat 21.6% empty running actually costs your fleetEurostat splits empty running 25.8% national against 12.6% international. Converting an empty kilometre into a loaded one earns margin but saves no toll.29 August 2026 · 7 min read
- Logistics"95% ETA accuracy" is unfalsifiable — make it checkableA percentage with no window and no horizon cannot be wrong. Specify hit rate within a stated window at a stated horizon, MAE in minutes, and coverage first.29 August 2026 · 7 min read
- RegulationThe AI Act was not delayed. Two of its dates wereThe Digital Omnibus moved Annex III to December 2027 and left three duties in force since 2025. Where automation lands, and who actually holds the duty.29 August 2026 · 8 min read
- EngineeringBuild the evaluation set before you pick the modelA test set built from your near-duplicates, 200 dpi scans and mixed-script pages, why accuracy hides two unequal errors, and how to keep an LLM judge honest.29 August 2026 · 7 min read
- ProcessThe event log is the specification, not the workshop diagramA diagram shows the process people believe they run; timestamps show the one they run. The four fields that make a log analysable and the completeness floor.29 August 2026 · 8 min read
- StrategyWhy most GenAI pilots return nothing measurableMIT NANDA found ~95% of GenAI pilots show no P&L return. What that study actually measured, what it cannot show, and the six things a kill criterion contains.29 August 2026 · 8 min read
- ManufacturingYour OEE number is probably wrong — here is how to tellManual downtime logging captures 60–70% of stopped minutes. Automated capture should exceed 95%. The honest OEE it produces is lower than the one you report.29 August 2026 · 8 min read
- EngineeringSeven ways a RAG system fails that a demo never showsA demo only asks questions the corpus can answer. Seven RAG failure modes — chunk bleed, ranking, self-graded evals, stale indexes, leakage — and their checks.29 August 2026 · 10 min read
- Customer experienceThe 800-millisecond budget behind a voice agent that sounds normalEight hundred milliseconds, split across codec, VAD, ASR, LLM and TTS. What each stage actually costs, and the extra stage a Bulgarian voice agent has to pay for.29 August 2026 · 8 min read
- IntegrationThe dangerous ERP write is the one that succeedsReleased interfaces, idempotency keys and read-back assertions — and the write that succeeds against a deprecated endpoint while every dashboard stays green.29 August 2026 · 8 min read
If one of these is the problem you are living with, say so and we will send it to you first. Tell us which one.
Rather have the conversation than wait for the article?
Most of what would go in these pieces we will simply tell you on a call, applied to your process instead of a generic one.
Thirty minutes, no slides, and a written summary afterwards.