Notes from the build
13 articles on getting AI past the demo: how to scope it, how to prove it works, what breaks in production, and what the work looks like inside a given industry. Most came out of real builds rather than a content calendar.
What is an AI-native product studio?
The term gets used loosely. Here is a precise definition, the four things that make a studio genuinely AI-native rather than an agency with a model behind a chat box, and how to tell the difference before you sign anything.
Telling the shapes of supplier apart, before you sign
What a studio, an agency and a consultancy are each structurally good at — and the questions that separate a team which has run AI in production from one that has run it in a demo.
- AI product studio vs software agency vs IT consultancyA studio, an agency, and a consultancy will all quote you for the same AI project — and produce three different things. Here is what each is structurally good at, and the failure mode each one carries.4 min
- How to choose an AI development partner: an enterprise checklistEvery vendor deck looks the same. These are the questions that actually discriminate between a team that has run AI in production and a team that has run AI in a demo.4 min
The decisions that decide whether it survives production
Retrieval or fine-tuning, how to know a change made things better, what breaks when an agent runs unattended, and where a model can sit when the constraint is regulatory.
- AI agents in production: what actually breaksAn agent that works in a demo and an agent that runs unattended against real systems are different engineering problems. These are the failures that show up in week three.5 min
- LLM evaluation in practice: how to know your AI feature actually worksWithout evaluation, every prompt change is a guess and every regression is invisible until a customer finds it. Here is how to build a harness that is small enough to actually maintain.4 min
- Private and on-premises LLM deployment for regulated industriesNot every compliance requirement means self-hosting a model. Here are the four real options, ordered by cost and control, and how to work out which one your constraint actually demands.4 min
- RAG vs fine-tuning: which one does your use case actually need?The question is usually posed as a choice between two techniques. It is better understood as a diagnosis: is the model missing knowledge, or missing behaviour?4 min
Methods, in the order the work actually happens
Step-by-step: writing a scope that can be tested, and closing the distance between a prototype everyone liked and a system somebody owns.
- From AI proof of concept to production: what the gap actually containsThe prototype impressed everyone in the room and then nothing happened for eight months. That gap has a predictable content, and almost none of it is model quality.4 min
- How to scope an AI project so it actually shipsAI projects rarely die from a hard engineering problem. They die from a scope that was never testable in the first place. Here is a method for writing one that is.4 min
What reaches production in each sector — and what quietly does not
Healthcare operations, customer support, logistics and industrial safety, written around the constraint each one puts on the system rather than the technology.
- AI for customer support: deflection without wrecking satisfactionThe fastest way to ruin a support operation is to optimise a bot for deflection. The interventions that actually work start behind the scenes, with the agent rather than the customer.4 min
- AI for healthcare operations: what is actually deployable todayThe AI that reaches production in healthcare is rarely the AI that makes the news. It sits in the administrative layer, where the error cost is recoverable and the workload is crushing.3 min
- AI in logistics: the use cases that survive contact with the yardLogistics is unusually well suited to AI and unusually unforgiving of it. The difference between the systems that stick and the dashboards that get ignored is fairly predictable.4 min
- Computer vision for industrial safety: what it takes to run on a real siteDetecting a hard hat in a photograph is a solved problem. Running that detection across twenty cameras on a working site, at night, without generating alerts nobody reads, is not.4 min
Have a problem worth writing about?
Most of these came out of real builds. Tell us what you're working on and we'll tell you honestly whether we're the right team.
Start the conversation