Lessons

Real estate is a marketing and relationship business sitting inside one of the most heavily regulated areas of consumer law. This lesson maps where AI helps across listings, leads and comparables, and covers the discrimination rules that make targeting and copy riskier here than almost anywhere else.

If research and production compress, a consulting practice has to be clear about what it is selling. This lesson covers the disclosure conversation with clients, advising clients on AI when your own use is in question, what a defensible proposition looks like now, and the honest reckoning on which parts of the profession this actually threatens.

The concrete production workflows and how to run them safely. This lesson covers deck and document drafting, interview synthesis across many conversations, financial and data analysis where the difficulty was never the arithmetic, proposal writing, and the sourcing discipline that keeps a deliverable defensible.

Consulting bills for a bundle of research, analysis, synthesis and judgement, and AI compresses those unevenly. This lesson separates them, examines which parts clients were actually paying for, and confronts the pricing problem that follows when the visible artefact becomes cheap to produce.

A support deployment is only as good as the documentation behind it and the escalation path in front of it. This lesson covers curating a corpus from ticket history, designing escalation that recognises when to stop, what happens to the team as volume composition shifts, and the failure modes that end these deployments.

The higher-return deployment is usually the one that helps agents rather than removing them. This lesson covers drafting and knowledge surfacing, why suggestion acceptance is a misleading metric, the effect on new agents learning the job, and how assist tooling changes what an agent is measured on.

Support automation is usually sold and measured on deflection, which optimises for the customer going away rather than the problem being solved. This lesson examines what deflection actually measures, the difference between resolution and abandonment, why the two are indistinguishable in most dashboards, and what to measure instead.

A project manager's authority rests on being believed, and generated artefacts can erode that quickly. This lesson covers disclosure with stakeholders and teams, how the tooling interacts with different methodologies, measuring whether it helped when your own metrics are the ones in question, and the failure modes specific to this function.

The concrete uses, built so they hold up. This lesson covers plan decomposition and what it is actually good for, the documentation nobody writes, chasing and dependency tracking, retrospectives where clustering helps and candour does not survive recording, and the portfolio view a PMO can genuinely gain.

Project management is largely communication and judgement under uncertainty, which splits cleanly into work AI does well and work it cannot touch. This lesson separates the two, covers the administrative load that is the genuine target, and explains why estimation is the seductive case that mostly does not work.

Marketing's usual metrics cannot answer whether AI helped, because output volume rose and attribution is already hard. This lesson covers what to measure instead, the brand and legal risks that concentrate in this function, disclosure norms with clients and audiences, and the honest reckoning on which claimed gains survive scrutiny.

Where the work actually changes. This lesson covers a content pipeline built so claims stay grounded, personalisation and where its returns stop, what changed about search when answers replaced links, creative testing done properly, and the customer research synthesis that is quietly the highest-value use in the function.

Marketing adopted AI faster than almost any function and has the least reliable evidence about whether it helped. This lesson maps the capability honestly across the funnel, separates the uses where output volume is the point from those where quality is, and explains why marketing's measurement problem makes the question unusually hard to answer.

The obligations become concrete in three places: what you ask a vendor before buying, what you measure on your own applicants, and what you owe the person a system decided about. This lesson covers all three, plus the AI literacy duty as it applies to recruiters, and the questions worth asking before adopting anything.

Where AI actually helps an HR function, and where it creates exposure that outweighs the gain. This lesson works through the lifecycle: sourcing and job descriptions, the screening decisions that determine everything, interview support, onboarding as the best early deployment, and performance work where the label problem returns.

HR is the function where AI meets employment law most directly, and one of the eight Annex III high-risk areas. This lesson maps what the tools actually do across the employee lifecycle, which uses sit inside the high-risk tier, and the two practices that are prohibited outright rather than merely regulated.

Self-hosting is a permanent operational responsibility rather than a project. This lesson covers the security surface a local model creates, model upgrades and why they are harder than they look, capacity and cost control, what to monitor, and the honest signals that the deployment should be retired.

A model that runs is not a model that serves. This lesson covers what an inference server does that a naive loop cannot, continuous batching and why it dominates throughput, the memory arithmetic that decides which hardware you need, quantization for serving, and how to size a deployment from a traffic estimate.

Self-hosting is usually justified on privacy grounds, and the argument is often weaker than assumed. This lesson examines what actually stays private, which regulatory requirements genuinely demand it, the reasons that hold up better than privacy, and the honest total cost against an API.

The fine-tuning versus retrieval framing hides several options that are often better than either. This lesson covers long context and why it does not replace retrieval, prompt caching as a cost lever, agentic retrieval, continued pretraining, and how to revisit a decision that was right when you made it.

The decision is only half the work. This lesson covers what building each actually involves: parameter-efficient fine-tuning with LoRA and why it made the technique accessible, the training data problem that stalls most projects, the retrieval pipeline end to end, and how to combine them into one system.

Fine-tuning and retrieval are usually presented as alternatives when they change different things. This lesson separates them precisely: fine-tuning adjusts weights and teaches behaviour, retrieval supplies context and teaches nothing, and the question that decides between them is whether your problem is knowledge or form.

Dense vectors miss exact terms, so production retrieval combines them with lexical search. This lesson covers hybrid retrieval and reciprocal rank fusion, a tuning method that starts from a stated recall target, capacity planning, the operational failures that catch teams, and how to choose between running your own index and buying a service.

Three ideas carry almost all production vector search. This lesson covers HNSW as a navigable small-world graph with its M and ef parameters, IVF as coarse partitioning with nprobe, and the quantization schemes that cut memory by an order of magnitude, plus how they compose into the hybrid indexes real systems actually run.