The AI worked. That doesn't mean it helped.
19 September 2026 · 9 min read
An AI programme can plainly work while the business plainly does not move. Why value delivered is not value realised: take a baseline before you start, measure the whole value chain against the numbers you already trust, establish cause link by link, and find the constraint quietly eating the gain.
AI's easiest business case is its weakest
14 September 2026 · 8 min read
Sold as a headcount play, agentic operations does not survive the first budget cycle. A headcount cut is a level that laps; operational performance is a slope that compounds. Why the durable move is to automate up, not down — and two tests to tell a genuine case from a fashionable one.
Your AI agents are an estate, not a toolbox
7 September 2026 · 6 min read
The hard part of agentic AIOps is not teaching one agent to act — it is governing a population of them. Treat your agents as an estate: runbooks as guardrails, a supervised and tested improvement loop, and a register with an interaction graph, before you hand over the keys.
Your solution to a problem is incomplete until you have defined the wrong outcomes
10 August 2026 · 5 min read
A maintained evaluation harness is the most useful control on an AI workflow. Building one is less about engineering than judgment: knowing what a wrong answer looks like, and writing that boundary down so the software carries it every time.
AI workflows look cheap today. CIOs need a test for when they don't.
4 August 2026 · 7 min read
AI's subsidised pricing will not last. A commercial feasibility test for the people who sign off AI spend: stress each workflow against the day the discount ends, and map it by how hard it would be to switch provider.
AI won't grow your business. It can make growth cheaper to deliver.
27 July 2026 · 8 min read
AI's job in a growing company is not new revenue but making growth cheaper to deliver. GRADE is a five-gate test for which workflows are worth funding.
AI is an amplifier, not an escalator
23 July 2026 · 7 min read
Give two organisations the same AI tool and you get very different returns. Why your operating model, not the tool you buy, now decides your AI return.
The better the AI got, the closer I had to watch it
21 July 2026 · 8 min read
I built a production platform without writing a line of code, to test whether AI can replace the human. The productivity gain was real. So was something I did not expect.