Conversational AI Design: what changed

Full content check: corrections

2026-11-03T00:00:00Z

[Updated] Module 23: memory policy and code comment: The memory policy table now says that its consent and deletion rows become test cases, a note explains that memory about a person is personal data under GDPR and LGPD, and the comment in the code example now matches what the code does.
[Updated] Module 24: what to grade: The section title said three things but its table lists five. It now reads "What to grade, and at what level". The wording on simulated users is softened to "often" and "tend to", because that behavior varies.
[Updated] Guide: Messaging and WhatsApp, wording: The caution about the Brazil status is now neutral text instead of first person, and the message-template rule states the usual 24-hour window.
[Updated] Guia (português): WhatsApp, redação: A mesma cautela sobre o Brasil agora está em texto neutro, e a sugestão de módulos virou links ("Comece pelos módulos 3, 19 e 18").
[Updated] FAQ: simulated users: The answer now says simulated users tend to be more cooperative than real ones.
[Updated] Glossary: simulated user: Now says "often kinder than real users".
[Note] What was checked: The two new modules and their quizzes, the 48 FAQ answers, the regulation tracker, the messaging and Portuguese guides, the explainer, career and human-handoff guides, five cheat sheets, both new case studies, the new sections in modules 8, 17, 18 and 19, and the 25 new glossary terms. The note that MCP became stateless (specification of 28 July 2026) was confirmed against the MCP project's own release post, and the GitHub case study against Simon Willison's original article.
[Note] What was not re-read: Modules 1 to 22 as they stood before the demand research, and the eight earlier case studies, were not read again in this pass.

Liability, human handoff and a Portuguese starting point

2026-11-03T00:00:00Z

[New] Case study: A clinic's chatbot invented medical titles: A German appeals court (Higher Regional Court Hamm, 12 May 2026) held a company responsible for false answers its website chatbot gave about its doctors' titles. What the court said, what it did not, and how to design so a bot cannot make claims about your organization. Not final.
[New] Guide: What should stay with a person: A list of ten situations an AI agent should not handle alone, what the assistant does instead, four questions to ask before automating a task, how to hand off well and what to test.
[New] Guia (português): Design de IA conversacional, por onde começar: A map of the course in Portuguese, answers to three common questions (chatbot, assistente ou agente; aviso de IA; quem responde por um erro) and a pointer to the WhatsApp guide.
[Updated] Module 18: liability watch: Adds a dated box with the Hamm ruling and California AB 316, which stops a defendant from arguing that the AI acted autonomously, and a pointer to the guide on what stays with a person.
[Updated] FAQ: two answers updated, two new (48 in total): The answers on which calls an agent should never handle alone and on liability now point to the new guide and case study. New questions: whether an "AI can make mistakes" disclaimer protects a company, and what California AB 316 changes.
[Note] Both legal items were checked against primary sources: The Hamm ruling against the court's own press release and the published full text; AB 316 against the enacted bill (Chapter 672, Statutes of 2025). Both pages say these are summaries, not legal advice, and that the Hamm ruling is not final.

Follow-up: fact check

2026-11-03T00:00:00Z

[Note] Fact check: regulation tracker: The Article 50 dates, the Digital Omnibus postponement and the 2 December 2026 marking grace were re-checked against several law-firm summaries and found consistent. The California and FCC references rest on earlier, widely documented rulings and were not re-researched. No change to the text was needed.
[Note] Fact check: messaging guide: The guide already says that reports disagree on the Brazil status and that it was not checked against Meta's own terms. That caution was kept.
[Note] What was not re-checked: Pages not named in this release were not independently re-verified in this pass, and no vendor statistics were added.

Demand-driven update: rules, agent security, evaluation, memory and messaging

2026-11-03T00:00:00Z

[New] Module 23: Context engineering and memory: What goes into a model's context, how long conversations fail, and how to design memory for support (what to remember, what never to, consent, forgetting, memory poisoning).
[New] Module 24: Evaluating agents: Trajectory, tool-use and state checks, multi-turn quality, calibrating LLM judges, simulated users, a 50-example starter set and evals in CI.
[New] Guide: Chatbot, assistant, agent or copilot?: A plain-English comparison and an autonomy ladder, for the most common beginner question.
[New] Guide: AI disclosure and chatbot regulation tracker: EU AI Act Article 50 (applicable since 2 August 2026), what the Digital Omnibus delayed and did not, and a design checklist. Dated, with sources, not legal advice.
[New] Guide: Frequently asked questions (46 answers): Short answers grouped by topic, each linking to the page with the detail.
[New] Guide: Becoming a conversation designer: What the role asks for in 2026 job ads, how the modules map to the skills, and how to build a portfolio from the capstone.
[New] Guide: Messaging and WhatsApp for support: WhatsApp's 2026 policy on general-purpose bots, voice notes, consent, handoff continuity and what to verify.
[New] Guia (português): Agentes de IA no WhatsApp: Versão curta do guia de mensageria, escrita em português para o público brasileiro.
[New] Case study: One tool, three legs: How a single MCP tool combined private data, untrusted content and an outbound channel, and what it teaches about support agents.
[New] Cheat sheet: Lethal trifecta test (10 minutes): Three questions to ask about every agent, and how to break one leg.
[New] Cheat sheet: Agent evaluation checklist: Datasets, graders, trajectory and state checks, and release gates.
[New] Cheat sheet: Resolution, deflection, ROI and pricing: Definitions that cannot be gamed, a cost-per-resolution formula and red flags in vendor benchmarks.
[New] Cheat sheet: Agent protocol map: MCP, A2A and the commerce protocols as layers, with a note on what is still changing.
[New] Cheat sheet: Voice go-live checklist: Latency, barge-in, entity accuracy, consent, escalation and cost checks before a voice agent takes real calls.
[New] Glossary: 25 new terms (70 in total): Context engineering, trajectory, lethal trifecta, tool poisoning, deflection rate, resolution rate, memory types, end-of-turn detection and more.
[Updated] Module 18: lethal trifecta, indirect injection, tool poisoning, policy promises: New sections on the three-capability test, attacks that arrive through tickets and forms, tool-description poisoning, and why a bot must not invent commitments. Adds one quiz question and three sources.
[Updated] Module 8: protocol map and tool poisoning: New section placing MCP, A2A and the commerce protocols in layers, plus a tool-poisoning note in the security section. Adds one quiz question.
[Updated] Module 17: resolution, deflection, cost per resolution, ROI, pricing: New sections on metrics that cannot be gamed and on outcome-based versus per-conversation pricing. Adds one quiz question.
[Updated] Module 19: entity accuracy and unit economics: New sections on getting names, numbers and codes right and on what a minute of voice really costs. Adds one quiz question.
[Updated] Module 1: link to the chatbot, assistant, agent and copilot guide: One pointer added.
[Updated] Module 9: pointer to the context engineering module: One pointer added after the memory and state section.
[Updated] Module 14: pointer to the agent evaluation module: One pointer added.
[Updated] Module 13: pointer to the messaging guide: One pointer added.
[Updated] Module 22: next steps now include modules 23 and 24 and the guides: The capstone's closing section lists where to go next.
[Updated] Cheat sheet: pre-launch safety checklist now covers launch and incident readiness: Added a kill switch, rollback, human override, incident log and an owner for every policy.
[Note] Facts that move fast carry a "last verified" date: The regulation tracker, the messaging guide and the protocol sections say when they were checked and against what.
[Note] What was left out: Vendor statistics (resolution rates, ROI, salaries) and one list of incidents that could not be confirmed elsewhere were not used. Rulings and policies that rest on a single secondary source were kept out or marked to verify.

Rebuild since the first edition

2026-10-27T00:00:00Z

[Note] LLMs lead, intents assist: Most new assistants are model-first, with deterministic flows, retrieval and tools around the model. Intent platforms are now one option among several.
[Note] MCP went stateless: The July 2026 MCP specification removed the handshake and sessions. Module 8 covers what that means for tool design.
[Note] Agents need architecture: State machines, single agents and multi-agent systems each fit different problems. Module 9 gives you a decision framework.
[Note] Evaluation is the job: Testing, tracing and red-teaming are where assistants succeed or fail in production. Modules 14 to 17 turn that into a routine for testers and analysts.
[Note] Prompts are per-model: The same prompt behaves differently across models. Module 5 shows what differs and how to keep prompts portable.