# Notes from the field.

Engineering, strategy, and operational lessons from the work we ship. Opinionated and specific. No "AI is transforming everything" intros.

## Articles

### [Why your AI agent works in the demo and breaks in production.](/content/Insights/AgentsInProduction/index.html)  
Most agent demos run for 30 seconds in a sandbox. Production agents have to work for years. Four patterns that hold up.

### [RAG is not a product. It's an architecture.](/content/Insights/RagArchitecture/index.html)  
Most teams treat RAG as plumbing. It isn't. The failure modes live in places that never show up in a demo.

### [The eval is the product.](/content/Insights/EvalIsTheProduct/index.html)  
Without evals you can't deploy with confidence, can't improve with discipline, can't catch the regression that's about to embarrass you.

### [Build, buy, or wait: a framework for enterprise AI investment.](/content/Insights/BuildBuyOrWait/index.html)  
The interesting AI initiatives are never trivially "buy" or "build." They're the ones in between.

### [The four ways enterprise AI engagements fail.](/content/Insights/FourFailureModes/index.html)  
Wrong problem, wrong data, wrong adoption, wrong governance. Spotting them early is the difference between deployed and shelved.

### [Frontier models vs. open-source: when each one wins.](/content/Insights/FrontierVsOpenSource/index.html)  
Most production systems we ship use both. The question isn't "which one." It's "where does each one earn its place."

### [What to ask when interviewing an AI consultancy.](/content/Insights/InterviewingAConsultancy/index.html)  
Most consultancies sell strategy slides. Three questions separate builders from deck-makers.

### [The hidden cost of generic AI SaaS.](/content/Insights/HiddenCostOfAiSaas/index.html)  
You don't pay in license fees. You pay in workflow drift — and that bill compounds.

### [How to run an AI proof-of-concept that doesn't waste a quarter.](/content/Insights/AiPocPlaybook/index.html)  
A POC that can't ship is an expensive demo. Three rules to make yours actually inform the decision.

### [Designing AI for people who hate it.](/content/Insights/DesigningForPeopleWhoHateIt/index.html)  
Frontline workers, clinicians, analysts — the people who actually use enterprise AI often start out skeptical.

### [Notes from a 200,000-employee frontline AI deployment.](/content/Insights/FrontlineAtScale/index.html)  
Frontline AI at retail scale isn't a feature. It's a platform decision. Three lessons.