Focus: Moving from AI Pilots to AI Performance

Trend 1: The "Jagged Frontier" of AI Execution

One of the more striking patterns we keep seeing is how unevenly AI performs across different tasks. It can reason through genuinely complex problems and then get something surprisingly basic wrong. That uneven performance means the real skill for your teams isn't just knowing how to prompt AI well – it's knowing how to push back on what it gives you, spot where it's gone wrong, and make a clear-eyed call on when to trust it.

Read more:

• One Useful Thing: The Shape of AI: Jaggedness and Bottlenecks
• Harvard Business Review: A Systematic Approach to Experimenting with Gen AI
• Caversham House (LinkedIn): Plants grow in weird places – and so does AI

Trend 2: AI Agents vs. Authenticity

As AI agents take on more of the actual work – not just assisting but actively participating – the question of who is really behind something becomes harder to answer. The organisations we are paying attention to aren't trying to obscure this; they're making their human involvement visible and deliberate. Transparency about where people are in the loop has quietly become a marker of quality rather than a sign of limitation.

Read more:

• People Management: AI Agents and Authenticity
• One Useful Thing: Management as AI Superpower
• Caversham House (LinkedIn): 2026: the year of human–AI synergy

Trend 3: The End of the Productivity Tax

What separates the organisations genuinely seeing returns from AI from those still waiting isn't the tools they have access to. The CEOs making real progress have used AI as a reason to rethink how their businesses actually work – their processes, how decisions get made, how work is structured. And crucially, they're doing the learning themselves rather than handing it off. That combination of redesign and personal fluency is where the difference shows up.

Read more:

• World Economic Forum: Davos 2026: Why Scaling AI Feels Hard
• McKinsey: How the Best CEOs Are Meeting the AI Moment
• Caversham House (Thought Leadership): AI in 2026: solving the 6% problem

Tool 1: Cursor (AI-Powered Code Editor)

Philippa Cameron, our Learning Director, built ten interactive strategy tools in a single day using Cursor – and she'd be the first to tell you her coding experience was minimal going in. What changed was the nature of the work: instead of writing code, she was describing what she wanted clearly enough for the AI to build it. The real lesson from that day wasn't about Cursor specifically; it was about what becomes possible when technical execution stops being the bottleneck.

Value: Your strategic leads can now build the tools they actually need – designed around specific problems in your business – without waiting for developer capacity. That's a meaningful shift in how quickly an idea can become something you can actually use.

Tool 2: Gemini 3 Flash (Model Optimisation & Efficiency)

We've largely stopped asking "which model is most capable?" and started asking "which model is right for this task?" For a recent project with Venato.ai processing thousands of regulatory documents, Gemini 3 Flash matched the quality of the more expensive Pro models at less than half the price. The capability gap between tiers has narrowed considerably, and in many workloads it has effectively closed.

Value: Don't buy a 10-ton truck to deliver 4 tons of bricks. If you're running premium models across the board, you're almost certainly overpaying for a significant portion of your workload – and that budget could be covering considerably more ground.

Tool 3: GPT-5.2 "Thinking" Toggle (Reasoning over Reflex)

We used the Thinking toggle recently to overhaul a technical specification. Rather than the usual instant response, it held back, worked through cross-functional dependencies, and caught logical contradictions in the brief before it started writing. What came out the other end was something we could actually build from, rather than a first draft that needed pulling apart.

Value: There's a version of speed that just creates rework further down the line. For anything where getting it wrong is costly, asking the model to think it through properly first is worth the extra few minutes – the equivalent of asking a senior colleague to sit with a problem before they respond, rather than firing back immediately.