mAIn Street #355: Gemini now lets you exclude chats from training; Google Maps looks for appointments and calls businesses; Notebook turns your sources into interactive websites


mAIn Street #355 | Useful AI, clearer boundaries
mAIn Street Daily Newsletter. AI news for people who actually have jobs to do.

ISSUE #355 · TUESDAY, NOVEMBER 17, 2026

Useful AI, clearer boundaries

Today’s useful changes come with questions worth asking. Does turning off training also stop storage? Can an assistant find an appointment, or actually book it? And what does it need to do good work while you sleep? We lead with five practical answers, then separate tools you might use from infrastructure your builder should understand.

Today’s top five

1. Gemini will let you save chats without using them for model training

Google is separating two choices that people often treat as one: keeping a useful chat history and allowing those conversations to help train its AI models. The updated controls begin rolling out globally in the coming weeks. They are not necessarily in your account today.

There is an important catch. Turning History off will still leave future chats stored for roughly two months for service, feedback and safety purposes. Opting out of training does not mean the conversation disappears from Google’s systems.

Before uploading client information, check both settings and your organization’s rules. A saved conversation can be useful when you return to a project next week. That convenience does not settle whether the material belongs in a personal AI account.

Read the source

2. Google Maps is starting to look for appointments and call businesses

Ask Maps is adding help finding service appointments, including beauty services, pet care, auto work and home repairs. It can look for openings and call businesses as part of that search. Finding a slot is not the same as guaranteeing a completed booking.

The Android and iOS rollout began November 16. Appointment discovery and route tips start in the United States. Itinerary creation and searching your saved lists cover markets where Ask Maps is available.

For a salon, groomer or repair shop, this makes accurate business information more useful. Keep hours, service descriptions and booking options current. If an automated caller asks about availability, staff still need a clear process for confirming what can actually be promised.

Read the source

3. Gemini Notebook can turn your source documents into an interactive site

Google says Gemini Notebook can build visual explainers, dashboards and project trackers from a collection of sources. The resulting site can open in another tab or be shared. The web rollout is global for adults 18 and older, including Workspace users; mobile support comes later.

That offers a practical alternative to sending someone a long summary. A checked set of onboarding documents could become a guide a new employee can explore, for example. Start with material you are allowed to use and a small audience.

Check the generated facts, calculations, controls and sharing settings before handing it over. A site that looks finished is not necessarily production-ready software, and using approved sources does not guarantee every conclusion is correct.

Read the source

4. AI-assisted attacks make a reliable patch routine more important

Google’s Threat Intelligence Group and Mandiant report seeing agent frameworks in live intrusions and automated analysis of software patches. Their concern is that attackers can use AI to investigate known weaknesses and move through parts of an attack more efficiently.

These are Google’s observations, not evidence that every attack is autonomous. The report also promotes Google’s defenses. It does not give every business a universal new deadline for installing every patch.

The useful question for whoever maintains your website and business systems is simple: how do we learn about a serious flaw, decide its urgency and verify the fix? Include account protections and access to AI services in that conversation. Buying another assistant is no substitute for maintaining the systems you already rely on.

Read the source

5. An assistant that works overnight needs more than a long prompt

Anthropic’s November 13 guide explains the parts of an always-on agent: dependable events and schedules, saved task state, memory, permissions and a way to deliver results. This is architecture guidance, not a new one-click app. An always-on assistant can spend most of its time waiting between runs.

For a business owner, the assignment matters as much as the software. “Watch for a change and tell me once” needs a source to check, a definition of a meaningful change and a record of what was already reported.

Start with read-only work, such as checking a public deadline or preparing a morning research brief. Say when the task should stop and which actions require your approval. A useful overnight result is one you can inspect quickly, not a pile of duplicate alerts or commitments you did not authorize.

Read the source

20 more useful reads

The first few are close to everyday work. The builder and research briefs that follow are labeled so you can skip what does not fit your job. Product claims below come from the linked providers; they are not independent performance tests.

6. Workspace skills can save instructions, but shared copies need maintenance

Google’s reusable skills bring instructions and reference material across Gmail, Docs, Slides, Drive and Chat. Sharing makes independent copies, so later edits do not automatically reach everyone. Source-file permissions remain separate. The English-language rollout is for eligible adult accounts; scheduled-release domains begin November 30. Sheets and Meet are not included yet.

Read the source

7. Judge AI spending by a completed, checked task

An AWS guide makes a useful distinction: cheap model usage is not the same as cheap work. Include retries, human review and whether the result is usable. Try measuring a small batch of recurring assignments before adding another subscription. The guide’s worked examples illustrate a method, not guaranteed savings.

Read the source

8. Google’s Tabletop experiment gives group planning a shared workspace

Tabletop can turn links, files and prompts into collaborative planning tools such as polls, itineraries and expense splitters. It could suit a volunteer event or small-team outing. This is a Google Labs experiment with a U.S. adult waitlist, not an app everyone can immediately open and use.

Read the source

9. Gemini Decisions is built for bounded choices, not long answers

The developer API scores yes/no, multiple-choice and rubric-based questions, including questions about different kinds of media. A potential use is routing routine support requests while leaving uncertain cases for a person. It needs integration work, and a confidence score is not a guarantee that a decision is correct.

Read the source

10. AdMob’s planned assistant targets ad setup and troubleshooting

Google announced Ask AdMob, agent skills and a planned connection for other AI tools. App developers may eventually have less manual integration work. Testing of the assistant is expected over the coming months, with more autonomous action planned for next year. Treat this as a roadmap rather than an available replacement for your current process.

Read the source

11. Search’s holiday guide shows where booking help still has limits

Google highlights AI Mode help with hotel reservations, dining, tickets and availability monitoring. The examples show how search can move toward a specific task, but access varies. Hotel checkout covers selected properties in U.S. English, and dining features vary by market. The guide is not a claim that every featured capability launched yesterday.

Read the source

For the people building your tools

12. A2UI gives developers a way to show useful controls instead of more text

Google’s A2UI 1.0 developer protocol supports interfaces an agent can present, such as selectors and forms. For someone commissioning a customer-service tool, that is a useful design question: would a clear choice work better than another paragraph? This is developer infrastructure, not a finished customer-service product.

Read the source

13. Google demonstrates speech recognition and speech output on the device

AI Edge Gallery includes examples of local speech tools. Processing on a device can be worth exploring where connectivity and data handling matter. The announcement is a developer showcase. It does not mean every Android app now has these capabilities, or that a particular implementation needs no privacy review.

Read the source

14. Give someone responsibility for an agent’s unusual behavior

Google’s hands-on Agent Anomaly Detection guide focuses on monitoring agent activity. The practical management question is who receives an alert and what they can do about it. Logging suspicious behavior is useful only if someone can investigate. Monitoring adds a safeguard; it does not prove an agent will stay inside its assignment.

Read the source

15. A waiting AI session can still have a computing cost

A Google Kubernetes Engine guide argues that builders should plan capacity around agent sessions rather than individual requests. This is specialist engineering guidance, but the purchasing question is plain: what are you paying while an assistant waits? Ask a contractor how idle sessions and unfinished jobs affect your bill.

Read the source

16. Google’s teamwork research separates proposing an answer from checking it

Google describes science and engineering collaborations using Antigravity’s /teamwork feature, including separate critics and verification. These are reported collaborations with human experts, not unaided discoveries. The broader habit is useful outside research: ask for an independent check of an answer, then test its important claims against the original evidence.

Read the source

17. AWS Emulator lets builders test locally, with security limits

AWS’s free, open-source emulator supports local testing through familiar AWS interfaces using Docker on Windows or macOS. That can reduce trips to a live cloud environment during development. It does not reproduce important controls such as IAM and resource-policy enforcement or cross-account and cross-region behavior. Local success is not a production security sign-off.

Read the source

18. Nitro Sandboxes add isolation for agent-generated code

AWS describes hardware-enforced boundaries for supported workloads running generated code. The business point is to contain experimental software rather than let it run freely alongside important systems. There is no additional sandbox charge, but the underlying cloud resources still cost money. This is infrastructure for builders, not a safety switch in a consumer chatbot.

Read the source

19. Cognito token exchange can give an assistant a narrower access pass

AWS’s November update adds token exchange for Cognito Essentials and Plus users. In plain English, a system can exchange an access token for one suited to a delegated task rather than simply passing along the original credential. Implementation matters: ask which systems and actions the resulting permission actually covers.

Read the source

20. AgentCore changes the default runtime, not every existing deployment

AWS says its new default runtime can scale idle sessions to zero. It also reports faster starts and lower billed memory in its measurements. Existing runtimes are not automatically moved over, and regional availability applies. If you pay for a custom assistant, have its builder check the deployment rather than assume a lower bill has already arrived.

Read the source

Research and the bigger picture

21. NVIDIA’s scientific-agent release includes more than a model

NVIDIA announced Nemotron for Science with models, data and recipes for research workflows. The audience is researchers and developers rather than a small business shopping for a general assistant. The useful distinction is that specialized work needs suitable data, tools and evaluation, not just a chatbot given a scientific job title.

Read the source

22. Quantum and conventional computing remain a research partnership

NVIDIA, IBM and Fermilab announced work on a quantum-classical codesign platform, alongside error-correction research. This concerns future scientific computing. It is not evidence that a practical quantum system is ready to run an ordinary business, and laboratory improvements should not be treated as promised commercial results.

Read the source

23. NVIDIA’s Vera expansion is a reminder that AI needs more than GPUs

NVIDIA is extending its Vera CPU platform into scientific computing. General-purpose computation still has a role alongside the graphics processors associated with AI. Performance and efficiency projections come from the vendor. This is infrastructure news, not a reason for a small team to replace computers that already meet its needs.

Read the source

24. Wearable-based coaching research is not a medical diagnosis

Google research explores a framework for cardiovascular, kidney and metabolic risk profiling and personalized coaching using wearable information. It is investigational work, not an approved diagnostic tool or individualized medical advice. A promising model of health risk still needs appropriate validation before people rely on it for care.

Read the source

25. Language Explorer makes linguistic diversity visible

Google’s interactive research and education resource offers a way to explore the world’s languages. It is not a newly released translation model. For organizations serving more than one language community, it is a useful reminder that translation quality involves people, context and local knowledge as well as software.

Read the source

AI Workflows | Tuesday: Context and harness management

Your AI can keep a project handoff current while you decide what needs attention.

Anthropic’s November 13 guide to always-on agents describes the setup behind assistants that carry work across sessions: events, timers, memory, permissions and delivery. It is a guide for builders. The useful lesson for the rest of us is that recurring work needs a reliable way to pick up where the last run stopped.

Suppose you coordinate a store opening. The launch date lives in a plan, the sign installer’s update is in email, and yesterday’s meeting changed who owns the inspection. You need a morning handoff that tells you what changed and which decisions need you.

Context is the material the assistant needs: the plan, current updates, responsibilities and your definition of a problem. The harness is the setup around it: approved access, saved instructions, a schedule, checks and limits on what it may do.

Read approved sources → compare with the last handoff → show changes and unanswered questions → wait for your decisions.

Give the job a home and a memory.

Use a company-approved agent with access to the relevant folder and messages. ChatGPT Work, for example, can use connected apps and run recurring work. Access depends on your account and workspace permissions. Have it verify the sources it can read before preparing the first handoff. An approved, sanitized export is a fallback when a connection is unavailable.

Keep the standing brief separate from changing facts. The brief says which project to watch, what counts as urgent and what requires approval. A dated working record holds the latest confirmed status, unresolved questions and items already reported. Have the agent preserve the previous version so you can inspect what changed.

There is a scheduling catch: OpenAI says a task created in a project cannot access files uploaded or stored in that project. Put the necessary instructions in the scheduled task and use supported, authorized sources it can actually retrieve. Test that access during a scheduled run before relying on it.

Give the agent this assignment.

Prepare my store-opening handoff using only the approved project folder and project email thread. Verify access first. Compare current sources with the last dated handoff. Report changed deadlines, blocked work, conflicting instructions and decisions I need to make. Link each change to its source and date. Mark missing information as unknown. Save a new dated handoff in the approved working folder without replacing the previous version. Do not change the project plan, contact anyone or make commitments. After I check the first result, help me configure a weekday 8 a.m. run in my time zone. Include these source boundaries and instructions in the scheduled task itself. Report an access failure clearly; never treat it as proof that nothing changed.

Test the setup with a known deadline change and an unresolved question. Both should appear with the right source. Run it again without new information: it should identify no new changes. A fresh-looking report that repeats yesterday’s alerts fails that test.

Managers can use the same handoff format across projects. Assign a person to approve changes to the standing brief and keep the actual project decisions with the responsible owner.

Pick one active project today. Ask your agent to list the sources it can access and draft the standing brief before you give it a recurring schedule.

One thing to try today: choose a recurring task, define what a useful finished result looks like, and name the point where you want to review it. That is a better starting place than asking an AI to do everything.

mAIn Street

Check out the resources I offer below and sign up for my new newsletter!

Read more from mAIn Street
mAIn Street issue 354 cover: Approval before action

mAIn Street #354 | Approval before action ISSUE #354 · MONDAY, NOVEMBER 16, 2026 Approval before action AI is moving closer to the parts of work that carry consequences: spending money, answering customers and moving information between systems. Today’s top five cover ChatGPT Wallet, business discovery, freelance project economics, healthcare agents and Adobe’s reusable creative workflows. Then we put the same idea to work in a supplier-bid comparison you can check before anyone buys...

A worker learns visual hierarchy by arranging a clear customer-pickup graphic, with a sketch showing the order of information.

ISSUE #353 · FRIDAY, NOVEMBER 13, 2026 Google’s Gemini 4 Hydrogen accepts Office files directly, ElevenLabs opens early access to a workplace meeting app, and a Capgemini survey finds AI costs running over budget at many large companies. Today’s Friday workflow uses a customer-pickup graphic to practice visual hierarchy, with YouArt in the tool spotlight. Today’s top five 1. Google releases Gemini 4 Hydrogen with native Office-file support Google lists Gemini 4 Hydrogen as generally available...

A business owner reviews sales, inventory and calendar information through an AI assistant.

ISSUE #352 · THURSDAY, NOVEMBER 12, 2026 Databricks is connecting company data to the AI assistants people already use. Railway reports an estimated $740,000 in charges from stolen AI keys, while DeepL and Otter expand the ways people can work across languages. Plus: a Thursday workflow for testing a small service your existing clients might pay for. Today’s top five 1. Databricks brings shared business definitions into four AI assistants Databricks introduced Genie plugins for ChatGPT,...