Enterprise AI

Governance starts by naming who owns a system before a model is chosen. This episode covers mapping what you run, the rules you already answer to, and Italy's data-protection case against OpenAI, which turned on ownership rather than model accuracy. Cost is what one accepted result costs, read against Klarna's own words about its assistant and the later correction about quality. It also covers shadow AI, a thin slice versus a platform, and one page for one system. Rob is an AI voice.

One good answer is not a test. This episode covers the questions a demo hides, checks a program can decide, whether to run them once or every time, and what changes when the judge is itself a model. The public case is Mata versus Avianca, where a court filing cited cases a chatbot had invented and the person who signed had checked the writing for flow. It closes with traces and release gates. Rob is an AI voice.

Some actions cannot be rolled back, and a long-running agent has to survive waiting, a step that runs twice, and a real stop. This episode covers saved state, durable execution, and undoing forward with a compensating action. The public case is Jason Lemkin's report that a Replit agent deleted a production database during a code freeze, and Replit's confirmation that development and live data had shared one database. The question at the end is how much autonomy to allow, and for which actions. Rob is an AI voice.

An agent can be turned against the person it works for when untrusted text, private data, and a way to act line up together. Simon Willison calls that the lethal trifecta. This episode sets that frame next to the OWASP list for 2026, then three public cases: a 2024 report on Slack AI, a 2025 Microsoft disclosure about Microsoft 365 Copilot, and a 2023 Chevrolet dealer chatbot that was taken offline. The open question is whether an approval gate is enough, or whether the system must be built so it never holds all three at once. Rob is an AI voice.

A tool is an interface, and the code behind it runs. This episode covers two protocols and which direction each one faces, where the agent runs, whose authority a call uses, and what happens when a credential expires in the middle of a job. It treats agents as non-human identities, and it covers vetting and pinning MCP servers before they are switched on. The check is a permissions table. Rob is an AI voice.

A chat is not the same thing as memory. This episode separates three kinds of memory, whose memory it is, and how a stored memory is supposed to go away. A skill is a procedure kept in a folder, and some of those steps belong in code. The public case is Johann Rehberger's SpAIware report on planted memories in ChatGPT. Then who approves a memory, and whether a skill should stay in words or move into code. Rob is an AI voice.

The files a system can read set the ceiling on its answers. This episode covers owners, dates, duplicates, conflicts, who can already open a file, which document wins, and why a deletion has to reach every copy search made. The cases are an over-shared file in Microsoft Copilot and New York City's MyCity chatbot. The open question is whether to clean the library first or ship a checked slice. Rob is an AI voice.

Retrieval is how a system finds the passages to put in front of the model. This episode covers when you can skip it, how documents are cut into chunks, keyword search beside search by meaning, putting a chunk's surrounding context back, a second pass, and why permissions belong in the query. The public case is what Anthropic measured. It closes with how to choose a first build and how to measure retrieval on its own. Rob is an AI voice.

Context engineering is deciding what the model sees on each call, not how large the window could be. The opening case is a workshop write-up in which a tutor lost a planted fact and the judge still scored the answer as good. The episode covers how a step is built, why longer is not better, what caching changes, sub-agents and compaction, and the ACON method. The check is what to log before you change anything. Rob is an AI voice.

A prompt can go to a public hosted service, a private endpoint in your own cloud, or weights you run yourself. The public case is Samsung's temporary restriction on generative AI tools after company data reached ChatGPT. This episode compares what a public API promises, what a private endpoint still keeps, and what running the weights yourself asks of you. Sort the work first, then price it. Rob is an AI voice.