Once an organisation runs hundreds of AI agents, it needs onboarding, access reviews, performance management and offboarding for them. The function that already knows how to do that is HR, not IT. This article looks at the evidence on agent sprawl, identity and accountability, and what it means for regulated firms.
Public benchmarks tell you very little about whether an AI system works for your organisation. This article looks at what the research says about benchmark quality, why agent scores are easy to overstate, how to build an evaluation set of your own, and what regulators expect you to be able to show.
Workslop is AI-generated work that looks polished but has no substance, so the person who receives it does the thinking the sender skipped. This article looks at the research behind the term, what it costs, why it happens, and what it turns into when it reaches a court or a regulator.
Agentic AI spend is variable, non-deterministic and has no natural ceiling, which makes it behave far more like a trading position than a technology purchase. This article looks at the evidence on what agents actually cost, the emerging discipline of FinOps for AI, and why the controls that fit are the ones financial services already knows how to run.
Prompt engineering was about writing a good instruction. Context engineering is about deciding what a model should be looking at when it answers. This article explains the shift, the research showing that longer context makes models less reliable rather than more, and what that means for organisations building on AI.
Grounding stops AI from making things up by connecting models to verified, up-to-date sources at query time. Here's a plain-English guide to RAG, tool use, and why grounding matters most when the stakes are high.
Dark factories - where a specification goes in and finished work comes out autonomously - started in software. This article explains the concept and maps how the same pattern could apply to audit, banking operations, legal, and compliance.
Learn how retrieval models, embedding vectors and retrieval-augmented generation are combined in modern AI systems. This article covers semantic search, the role of embeddings in supporting generative models, and the practical trade-offs of embedding-based retrieval.
An overview of how generative models produce text, images and code in modern AI systems. We look at how next-token prediction works in practice, why generative models can sound fluent without truly understanding, and how generation is typically combined with retrieval.
AI tools are now part of how many students learn. Used well, they can support understanding and independent practice. The real challenge for schools is how to assess learning fairly and meaningfully when these tools are part of the process.
© Cow-Shed 2004 - 2026
