A working journal, writtenarguedsketchedin the open
The Journal is written by the people who deliver the work — not by a content team and not by a model. Field notes from engagements we are mid-flight on, opinions formed by being wrong about things, and the occasional argument with a fashionable idea. We publish when there is something useful to say. Sometimes that is monthly. Sometimes it is not.
Pieces worth going back to.
The AI programmes authorised in 2024 are reaching their renewals. Most cannot demonstrate what they produced.
The enterprise AI programmes authorised in 2024 are now reaching their renewal cycles, and in most cases the teams responsible cannot demonstrate what was produced — not because the systems failed, but because the evidence was never built to survive the workflow changes that deployment caused. The board asks for the ROI number; the programme lead has recall metrics and usage data. Here is why the measurement gap is almost universal, what it costs at renewal, and what the four programmes that passed their renewals did differently from the start.
The context window is not a retrieval architecture.
Loading a million tokens into context is not cheaper, faster, or more reliable than retrieval — it is the reverse of all three, in ways that only become legible after launch. We have rebuilt four long-context deployments to retrieval architectures since January 2025, and in each case the result cost thirteen times less to operate. Here is what makes the long-context shortcut so persistently attractive, and what it eventually costs.
The latency budget is back, and the systems we built while it was gone are showing it.
For two years, response time was the most forgiven performance problem in enterprise software. Models were novel enough that waiting four seconds felt reasonable. In 2026, those systems sit on the critical path of synchronous workflows where four seconds compounds into thirty minutes of daily waiting. Here is what changed, where the architectural debt lives, and what we are rewriting across fourteen production deployments.
After month twelve, your AI evaluation is measuring the world your system created.
After twelve months in production, an AI system's launch-day benchmarks routinely become the thing obscuring its deterioration rather than revealing it. We have now seen this pattern in eleven deployments — the evaluations pass, the metrics hold, and the system has drifted in ways nobody is measuring. Here is what causes it, which sectors accumulate it fastest, and what we now require before any system leaves our hands.
Enterprise procurement has found its questions. Most AI vendors are still rehearsing their answers.
The enterprise AI buying cycle changed materially in the first quarter of 2026 — longer timelines, real technical due diligence, and procurement questionnaires most vendors cannot honestly complete. The shift is being driven by boards who authorised AI programmes in 2023 and 2024 and have since seen what they produced. Here is what the smart buyers are now testing for, and why the firms that have only ever sold demos will not pass.
What breaks in agentic systems at month seven, and the three structural gates we now require before any deployment.
Agentic AI systems fail in ways that demos do not reveal and evaluations rarely test. After eighteen months of production deployments, three failure modes have become reliably visible — none of them are about model capability, all of them are about architecture. Here is what we have learned to require before we approve a build.
Why most enterprise AI pilots never reach production — and the four conditions that change that.
Eighteen months of pilot work across three sectors taught us that the gap between a working demo and a deployed system is not a technical gap. It is an organisational one. Here is what we have learned to look for before we agree to build.
The case for retrieval-first architectures over fine-tuning, in seven failed projects.
Fine-tuning a frontier model looks like the obvious move when the demo data is yours. It almost never is. A note from the engineering bench on why we keep reaching for retrieval, what fine-tuning is actually good for, and the projects that taught us the difference.
Boards have stopped asking what AI is. They are starting to ask better questions.
A shift in the conversations we have been having in board rooms this year. The interesting boards are no longer asking what artificial intelligence can do. They are asking what it should be allowed to do, and on whose authority.
The Journal does not run on a content calendar. We publish when a piece of work has produced a finding the rest of the practice wants to share — and when the client has agreed to the framing. Some quarters that means three pieces. Some quarters it means one. We have never published filler and we do not intend to start.
If you would like to be told when something new goes up, the most reliable way is to ask. There is no newsletter, no opt-in, no automated drip. A senior at the firm will add you to a short list and email you a link the day a piece publishes.
Ready to build something that matters?
Tell us about your challenge. We’ll scope a solution, give you a realistic assessment, and tell you exactly what it takes to win.