Insights
Writing about operational problems, technology decisions, adoption and evidence. No trend-chasing and no disguised sales letters: if a piece is not useful on its own, it does not belong here.
Each piece takes a question operational leaders actually face and works it through honestly, including the parts that do not flatter the technology.
Growth · 21 August 2026 · 11 min read
Britain has the talent. So why aren’t more businesses scaling?
Britain still produces talent, ideas and ambitious businesses. The harder problem is helping enough of them scale—and whether AI can strengthen established firms as well as create new ones.
Read the article →
Model behaviour · August 2026 · 5 min read
A model can think something it never says
Anthropic published research in July 2026 describing a way to read what a Claude model is privately holding, not just what it writes down. The finding is real and the caution around it is deserved. What it changes is what 'evaluate the model' can honestly mean.
Read the article →
Operating environment · August 2026 · 5 min read
The compute boom needs a social licence
Opposition to AI data centres has grown sixfold in a year, and the constraint on capacity is becoming social rather than technical. A benefit that requires becoming a customer is not compensation, and the test generalises well past data centres.
Read the article →
Resilience · August 2026 · 7 min read
The model you depend on is rented
In June a frontier model went dark for eighteen days by government order rather than by outage. Capability you rent runs on terms you did not write, and a fallback nobody has run is not a fallback.
Read the article →
Economics · August 2026 · 4 min read
Your AI bill is becoming a payroll
Agentic AI does not consume like software. It consumes like labour: variable, workload-driven, and capable of tripling while you are in a budget meeting. Managing it takes payroll instincts, not licensing instincts.
Read the article →
People · August 2026 · 4 min read
The gains are real. Who gets them?
AI productivity gains are no longer hypothetical. What remains undecided, in almost every organisation, is where they go. That decision shapes whether your people fight the technology or champion it.
Read the article →
Direction · August 2026 · 4 min read
Finding the first worthwhile use case
Most organisations do not have an AI problem. They have a queue of operational problems and a new tool of uncertain relevance. The first use case decides whether the tool earns trust or burns it.
Read the article →
Delivery · August 2026 · 4 min read
From prototype to production
The distance between a working demonstration and a dependable system is where most AI initiatives quietly die. The gap is not mysterious. It is a list of unglamorous engineering debts.
Read the article →
Workforce · August 2026 · 5 min read
Tasks are not jobs
A capability benchmark just quadrupled in eight months. A sector shed jobs while the wider economy added them. Neither number means what the headline built from it wants you to think.
Read the article →
Economics · August 2026 · 5 min read
What adopters actually do
A study of 21,000 US businesses found companies that adopted AI grew headcount while others stood still, on spend that started at about thirty dollars a person a month. Neither the jobs story nor the budget story most people assume survives the data.
Read the article →
Adoption · August 2026 · 4 min read
The gap is a management problem, not a model problem
A stronger model does not close the gap between what AI can do and what it actually delivers. It can widen it, by increasing what looks safe to delegate faster than any organisation can build the supervision to match. Closing the gap is a management job, and the evidence says who does it matters more than which model you bought.
Read the article →
Technology choices · August 2026 · 5 min read
Private AI in regulated operations
For operations governed by regulation and confidentiality, where AI runs matters as much as how well it performs. Private does not mean isolated, and control does not mean building everything yourself.
Read the article →
Vendor risk · August 2026 · 4 min read
The vendor you vet is not the vendor you get
A model provider not training on your data answers one question about trust and leaves a different one open: whether the same provider ends up competing in your market. The two risks need separate answers.
Read the article →
Evidence · August 2026 · 4 min read
Reading the entrepreneurship boom like a scientist
Solo founders and one-person, million-dollar businesses are the story of the year. Some of the data behind it has been wrong before, in the same direction, for reasons that had nothing to do with AI. That is not a reason to dismiss the story. It is a reason to grade it.
Read the article →
Organisational design · August 2026 · 4 min read
What lean AI-native companies are actually testing
AI-native startups are reportedly a quarter smaller than their predecessors at the same valuation, doing the work with fewer people rather than more tools bolted onto the old headcount. That is a live experiment in organisational design, and larger operations are the ones with the most to learn from how it fails.
Read the article →
Agentic AI · August 2026 · 5 min read
The harness decides more than the model
OpenAI and Cursor both shipped agent news in the same week, and the more interesting story was not which model scored highest. It was how much of the outcome now turns on what the model is allowed to touch, not on the model at all.
Read the article →
Evidence · August 2026 · 4 min read
A benchmark just audited itself, and lost
OpenAI withdrew its own recommendation of a leading coding benchmark after finding roughly 30% of its tasks were broken. The finding is specific and checkable. What it implies about every other benchmark chart in a vendor's launch post is the part worth sitting with.
Read the article →
Disagree with something we wrote?
Even better. The most useful conversations start with a real problem or a sharp objection. Bring either.
Your message goes to the people who would do the work.