Writing

Frontier models are now reliable enough to carry institutional processes. The advantage goes to organisations that codify their expertise into governed, reusable skills.

Even flawless agents have an economics problem. Pay for research, not for agents clicking around to retrieve facts a data pipe could supply.

2026 will be defined by the Action Layer: what AI can touch, not just what it knows. Three predictions for market research.

Ask an AI the same sizing question five times and you can get five different markets. In research, consistency is trust.

AI loves tidy lists of three. That habit deletes outliers and inflates minor points, and it costs you real insight.

Why only a specialist research firm can build an AI replica of what it does: the moat is the tribal knowledge, not the model.

AI is trained to agree with you. A researcher’s job is to challenge you. On sycophancy, and how to work against it.

Tool fluency is now table stakes. The real shift is strategic fluency, and the Insight Engineers who build with AI rather than just use it.

Human-in-the-loop is everyone’s favourite guardrail, and on its own it fails. Delegate the task, never the responsibility.

Hallucinations are mostly solved. The bigger risk now is the Plausibility Trap: answers that look well researched and aren’t.