Practice
Not the model, not the tooling. They fail because nobody wired up the retries, or the one person who understood the process left, or the API changed on a Tuesday and no alert fired. Here is the unglamorous checklist we run before calling anything finished.
14 August 2026 · 9 min read
Practice
On one engagement the pricing step took four minutes and everything around it took two and a half hours. We would have automated the wrong thing entirely. A week of timing beats a month of assuming.
31 July 2026 · 6 min read
Engineering
The happy path is the easy twenty per cent. Idempotency, backoff, dead-letter handling and an alert that reaches a human — build those at the start and the system survives its first bad week.
17 July 2026 · 7 min read
Strategy
We talk clients out of roughly one project in five. A spreadsheet that one person maintains happily is not a problem waiting to be solved — it is a solution that already fits.
2 July 2026 · 5 min read
AI
Two hundred real tickets with known-good answers, run against every change before it ships. Without that you are not iterating, you are guessing and calling it iteration.
19 June 2026 · 8 min read
AI
Support deflection is usually a documentation problem wearing an AI costume. On one project the jump from 41% to 62% came from rewriting eleven help articles, not from touching the model.
5 June 2026 · 6 min read
Data
When two of your nine spreadsheets disagree about revenue, a model will be confidently wrong — which is worse than having no model. Fix the pipeline first. That is usually the whole project.
22 May 2026 · 7 min read
Practice
Day by day: who we sit with, what we time, which questions tend to surface the expensive answers, and what the costed roadmap contains at the end of it.
8 May 2026 · 10 min read
Engineering
Roughly one quote in six goes to a human on the freight system we built. That is not a shortfall in the automation — it is the design decision that kept the error rate near zero.
24 April 2026 · 6 min read