START HERE

A quick catch-up

I’m sorry for the brief gap. Thanks for sticking with me.

ChatGPT is adding locked conversations, Gemini is bringing personal context to the car, and a new manager survey asks a question many workplaces are still working out: how do you recognize someone’s contribution when AI helped produce the result?

OpenAI is also opening ChatGPT ads to more countries. Further down, the roundup follows AI into construction sites, schools and factories, along with new polling on who trusts it and who doesn’t.

Then there’s a Thursday workflow for finding a new offer in the requests customers already make, plus an early look at a design tool for menus, stickers and packaging.

01 · THE LEAD STORIES

ChatGPT lets you hide private conversations behind a PIN

The next time you share your screen, a personal ChatGPT conversation needn’t sit in plain sight. OpenAI has begun rolling out Locked Chats, which removes selected conversations from ordinary history and search. You reopen them with a four-digit account PIN or supported phone authentication. The feature covers Free through Enterprise and Edu plans, although it may not have reached your account yet.

This is useful for keeping a family matter or surprise gift out of view. It does not change OpenAI’s retention or training settings, create end-to-end encryption, or remove information ChatGPT remembered before you locked the conversation. Existing shared links also remain accessible unless you delete them separately. Workplace compliance access stays unchanged.

The important catch is recovery: resetting a forgotten PIN permanently deletes your locked chats and content protected only by that PIN. Support cannot restore it. Forgetting the number alone does not erase anything. Read the reset warning carefully before treating it like an ordinary password change.

02 · THE LEAD STORIES

Gemini brings your personal context into Android Auto

Finding a flight confirmation while heading to the airport is a good job for a hands-free assistant. Google is bringing Personal Intelligence to Gemini in Android Auto, letting it draw on apps you choose, including Gmail, Calendar, Photos and Maps. Google’s example combines two requests: find the flight gate and check whether the flight is on time.

The Android Auto rollout began Tuesday. Cars with Google built-in are due to receive the feature in coming months. Personal Intelligence is optional, but if you already enabled it in the Gemini app, that setting carries into the car. You can choose connected apps and turn personalization off in Gemini settings.

Access is limited to eligible personal Google accounts belonging to adults, excluding Workspace accounts and users in the EEA, UK, Switzerland and Nigeria. It requires Keep Activity to be on; Google says connected-app summaries, excerpts and inferences may be used for model training or human review. Before your next drive, check which apps you have connected and whether you want their details available through the car assistant.

03 · THE LEAD STORIES

Managers are struggling to see the human contribution in AI-assisted work

When AI helps write the report, managers still have to decide what the employee contributed. In new research commissioned by monday.com and conducted by Nielsen, 53% of surveyed leaders said AI made employees’ original contributions harder to evaluate. Another 78% said they had changed their performance-assessment criteria.

The survey covered 1,701 directors and senior managers at companies with at least 500 employees in seven countries. Fieldwork took place in April; the report was released Wednesday. These are large-company leaders’ reported experiences, not findings about all workers or small businesses.

The report also found 58% expected to manage more AI agents than people within a year of the survey. That is an expectation, not a forecast. For someone using AI at work, a practical response is to make the human judgment visible: explain what you checked, what you changed and why. Managers need an agreed way to recognize that contribution alongside the quality of the finished work.

04 · THE LEAD STORIES

ChatGPT advertising expands to 89 countries

More businesses can now reach customers while they are asking ChatGPT what to buy. OpenAI has added 26 countries to ChatGPT Ads, bringing the total to 89, with self-service access through Ads Manager or advertising partners. The company says more than 100,000 advertisers have run campaigns.

Ads are limited to Free and Go plans. OpenAI says paid placements are labeled, separate from answers and do not influence those answers. It also says conversations stay private from advertisers. Buying an ad therefore does not buy a recommendation inside ChatGPT’s response.

For a U.S. business, the standard minimum daily campaign budget is $25, although bidding settings can raise it. Daily means an average: spending can reach twice that amount on one day, with a seven-day ceiling of seven times an unchanged budget. A small test needs a clear spending cap and a way to count sales or qualified inquiries. The advertiser total shows participation; it does not tell you whether this channel will pay off for your business.

THE ROUNDUP

Everyday tools and creative work

05 OPPO brings phone-level agents to its new flagships

OPPO unveiled the Find X10 series with ColorOS 17 and a Manus-powered Agent Mode that can work with files and photos, plan trips, and create presentations or webpages. That puts multi-step assistance closer to the material already on your phone. Availability begins in late October and varies by market. The hardware launch is new; OPPO announced the Manus partnership earlier this month.

06 Meta Muse reaches more phones

Sensor Tower estimates Meta Muse exceeded 10 million U.S. downloads in 40 days, TechCrunch reports.

07 ElevenLabs adds live translation for business calls

Eleven Realtime Translation lets people speak different languages and hear translated speech during the same conversation. ElevenLabs is targeting customer-service calls, telecom networks and live events, with support for more than 90 languages. A support team could potentially serve more callers without transferring each one to a language specialist. Access is currently an alpha for ElevenAPI enterprise customers, rather than a consumer app anyone can download.

08 DigiSelf takes conversational avatars global, with an offline option

EmbodyMe is expanding DigiSelf globally after its March launch in Japan. A photo can become an avatar for video, meetings or live conversation, giving educators and service teams another presentation format. Businesses can also run avatars locally without sending live data to the cloud. That offline option is a separate software license starting at a one-time $9,900, not the consumer subscription.

THE ROUNDUP

Marketing, shopping and customer service

09 Picsart connects ad creation, testing and Google Ads

Picsart expanded Marketing Studio to connect ad planning, creative production, Zappi-powered testing and Google Ads launches. Small teams can move from product photo to campaign without stitching several services together. Google Ads is available now; Meta and TikTok connections are coming soon. Predictive scores can guide creative choices, but actual campaigns still have to prove themselves.

10 BigCommerce makes its stores easier for AI shoppers to use

Commerce and Cloudflare added WebMCP support to BigCommerce stores, letting compatible browser assistants look up products, compare options and build carts using store information. Merchants keep checkout on their own storefronts. For shoppers, the aim is fewer mistakes from agents guessing at website buttons; for retailers, it preserves the direct customer relationship. The feature is enabled by default for eligible merchants, who can switch it off.

11 Sierra’s Stripe partnership keeps payment inside the support call

Sierra announced a partnership with Stripe that lets customers resolve a billing question and pay during one AI-handled phone call. In the payment step, customers enter card details on their keypad for Stripe to process. Medical-billing company Cedar is an early user. The partnership and marketplace connection are this week’s news; Sierra already offered payment capabilities. The practical benefit is avoiding another transfer or separate payment link.

12 Therabody’s new shopping assistant stays inside its own store

Therabody launched Jason, an AWS-built shopping assistant, on its website. It asks about needs, budget and lifestyle, then recommends products and explains the research the company uses to support them. Shoppers can find it through “Chat to Shop.” This shows how a retailer can add conversational help without sending customers elsewhere. Its recommendations concern Therabody’s own range and should be read as sales guidance, not independent medical advice.

13 ThinkingStreams assembles campaigns from a marketing brief

ThinkingStreams announced a marketing platform that uses brand information and Microsoft Dynamics 365 data to draft emails, text messages, landing pages and campaign sequences. The appeal is less rebuilding and moving copy between a chatbot and marketing software. People must approve material before it sends or publishes. Teams are being directed to demos or early-access requests, so the announcement should not be mistaken for unrestricted availability.

THE ROUNDUP

Workplaces, schools and the jobsite

14 Executives put more trust in AI than workers do

Charter’s new AI Pulse report finds executives roughly twice as likely as workers to trust AI over their own judgment. Employee involvement in adoption and hands-on training were associated with better experiences. The August sample covered 5,002 U.S. office workers using AI at employers with 100-plus staff. For managers, the practical question is whether employees have a say in implementation. Reported trust does not establish actual decision-making behavior.

15 Construction crews can ask their project files questions aloud

Procore AI Go gives construction crews spoken answers drawn from project drawings, specifications and logs, reducing the need to hunt through screens at a jobsite. The mobile app is in open beta for North American customers with eligible Digital Coworker packages. A Meta glasses version is in closed testing. Creating logs and drafting requests for information remain planned capabilities, so the initial use is retrieving answers.

16 Oracle’s HR agents take on staffing and payroll preparation

Oracle introduced Fusion Claw-powered HR applications for workforce planning, recruiting, compensation and payroll. They can combine records, flag problems and prepare next steps across longer-running processes, such as checking payroll readiness or evaluating staffing options. Oracle describes the applications as available within Fusion HCM. Employers still retain responsibility for hiring and pay decisions, with recommendations and consequential actions routed through organizational review and approval.

17 Frontline connects school administration through AI-assisted workflows

Frontline Education announced a plan to connect school HR, student services and finance so district staff can see priorities and coordinate work across departments. An invoice-processing agent went live in Texas districts October 15, with district staff retaining control. The new announcement sets out the broader direction: shared-workspace and sign-in tools are in early access, and further capabilities will roll out through 2027.

THE ROUNDUP

Public opinion and health policy

18 Data-center expansion faces broad public opposition

A new ABC News/Washington Post/Ipsos poll finds 57% view AI as bad for society, while about two-thirds oppose more U.S. data centers. Opposition to construction in respondents’ own communities spans party lines. That makes AI expansion relevant to local planning and elections as well as technology buyers. The probability-based survey included 2,545 U.S. adults October 9–13; it measures opinions, not the actual effects of individual facilities.

19 Frequent AI users also want stronger regulation

Navigator’s new findings show 66% of respondents favor more AI regulation, including 62% of frequent users. Using the tools and wanting limits on them clearly coexist. Global Strategy Group surveyed registered voters October 1–5 using an online opt-in panel, with a 1,000-person base sample and demographic oversamples. These results describe voter attitudes, not all Americans, and the broad question does not establish support for any particular proposed rule.

20 AI’s appeal depends on what happens to the jobs that remain

In new American Compass/YouGov findings, 78% preferred a future where AI improves jobs over one where it replaces nearly all human labor and distributes the wealth. The choice puts job quality alongside job-loss fears in the public debate. Researchers compared separate samples of 1,000 respondents surveyed in May and September. The results reveal preferences; they do not prove what causes pessimism or track the same people changing their minds.

21 Southeast Asia offers another view of AI adoption

Google, Temasek and Bain’s new regional report estimates that 243 million Southeast Asians actively use generative AI, equivalent to 58% of the working population. It describes growing use in shopping and travel planning, offering a counterpoint to U.S. skepticism. Its projected $368 billion digital economy for 2026 covers commerce and other online sectors. That figure is transaction-related economic activity, not AI revenue, and the report’s estimates come from industry-sponsored research.

22 Seoul declaration puts access alongside medical AI development

South Korea and the WHO director-general unveiled the AI Healthcare Seoul Declaration at the World Bio Summit. It calls for fair access to AI’s health benefits, shared research and oversight matched to medical risk. The concern is who receives useful care as well as how quickly technology advances. These are cooperation principles, with implementation dependent on national policies, rather than a binding worldwide law or proof that any particular medical tool works.

THE ROUNDUP

Building and funding AI

23 Glean brings company-agent building into coding tools

Glean’s October 21 announcement details a generally available way to build, test and publish company agents from tools such as Claude Code and Codex. Teams can review changes and roll back mistakes while keeping Glean’s data permissions and audit trails. It currently supports auto-mode agents; workflow-mode agents still require Glean’s visual builder.

24 Blacksmith raises $66 million to handle more software testing

Blacksmith raised $66 million in a Sequoia-led round at a $1 billion valuation to expand infrastructure that builds and tests software. The bet is that a flood of AI-written code increases demand for checking whether changes actually work. The funding will support hiring and computing capacity; it does not itself introduce a new product.

25 mabl adds a separate check on AI-written software

mabl released a testing harness that checks software changes alongside coding agents, identifies affected tests and highlights gaps in coverage. Teams can define what agents may do and bring in existing test cases. That gives businesses another checkpoint before new code reaches customers. The announced verification tools are available now; desktop-application testing is still coming.

26 OpenAI models are coming to Google Cloud

OpenAI models and Codex workflows are coming to Google Cloud’s Model Garden, with purchases through its Marketplace and existing cloud-spending commitments. That gives Google Cloud customers another way to buy and manage AI within their current setup. General availability is planned for the coming weeks; the current invite-only preview is limited to approved development and internal testing.

27 OpenAI buys a coordinator for drug-development work

OpenAI acquired Orchestra Bio and is combining its team and technology with Rosalind, its life-sciences tools. Orchestra tracks studies and dependencies, helping teams see how a problem such as a failed manufacturing batch affects a trial’s schedule. The acquisition broadens OpenAI’s role in drug-development operations. A combined-product release date and purchase price were not announced.

28 NVIDIA gives agents more structure around company data

NVIDIA expanded its Agent Toolkit with an Agent Knowledge Foundation that connects business records to concepts such as customers, products and orders. New Kumo models support predictions from structured data, while OpenShell controls agent access and actions. The practical aim is agents that understand a company’s records. These are building blocks for technical teams, rather than a ready-to-use office assistant.

THE ROUNDUP

Engineering and industry

29 Mistral and NVIDIA target engineering design and simulation

Mistral and NVIDIA announced work connecting AI models with engineering simulation and 3D-design tools. They also plan public benchmarks and open datasets to test whether agents produce technically correct results in fields such as automotive and aerospace. That could give buyers better evidence than a polished demo. This is a development collaboration, with no finished engineering suite announced.

30 Rockwell plans a coordinator for factory-engineering agents

Rockwell announced FactoryTalk Conductor, software intended to coordinate AI agents across industrial design, simulation and testing. Manufacturers could connect work that currently sits in separate engineering tools, including Rockwell’s Logix and Emulate3D products. Buyers will have to wait: a public preview is planned for Automation Fair, with commercial availability expected in the first quarter of 2027.

31 Oracle previews an agent builder for industry workflows

Oracle announced AI Agent Studio for Industry Applications, a low-code/no-code environment for agents working with existing business data and permissions. Potential uses include spotting construction delays and managing retail catalogs. It could reduce the custom integration work needed for those tasks, but availability is planned within the next 12 months in the United States, with other regions later.

THE ROUNDUP

Security and investigations

32 CrowdStrike reports AI agents conducting real intrusions

CrowdStrike disclosed four previously unreported groups that used AI agents to choose actions and carry out intrusions, including attacks on healthcare and government targets. Its research covers May through August, so these are newly published findings about earlier activity. The cases suggest less-skilled attackers can do more; the attribution and conclusions remain CrowdStrike’s assessments.

33 Website owners: ask your host about the critical PHP fixes

AI-assisted security research helped uncover critical flaws in PHP, the software running many websites, and PHP’s maintainers released fixes on October 21. Ask your host or developer to confirm that your live server’s PHP runtime has those patches, including any provider-supplied backport. WordPress says these are PHP vulnerabilities; updating a plugin will not fix the underlying problem.

34 SentinelOne adds controls for agents while they work

SentinelOne introduced Singularity Agentic Detection and Response to inventory AI agents, restrict their permissions and watch what they actually do. The system is designed to stop a misbehaving agent without stopping an entire workflow. For companies giving agents access to real systems, that addresses mistakes as well as attacks. It is available in early access, not general release.

35 Corma uses browser agents to manage software access

Corma introduced browser-based agents that add or remove employee access in applications lacking provisioning APIs. Working through the interface could spare IT teams manual account changes or expensive upgrades purchased just for integration features. A customer quoted in the announcement describes the feature as beta; Corma does not provide a general-availability date.

36 Cellebrite expands AI-assisted evidence review

Cellebrite’s autumn release expands device support and adds capabilities to Genesis, its AI-assisted investigation system. Agents can incorporate investigators’ procedures while helping sort large collections of digital material. The company’s case stories suggest time savings, but they are customer accounts, not controlled evaluations. A version agencies can run on their own premises is expected in early 2027.

RESEARCH EXPLAINERS · OLDER STUDIES

New explanations of earlier research

These explainers were published October 21. The underlying studies are older, as noted below.

RESEARCH EXPLAINER

Older research, new explainer: AI agents sort biological clues

OpenAI’s October 21 explainer describes thousands of agents analyzing mouse-brain data and scientific literature to propose research leads. The underlying preprint was posted October 16, 2026. OpenAI, PerturbAI and Scripps report completing the parallel analysis in 32 minutes. Its value is helping scientists choose follow-up experiments; proposed disease mechanisms still need testing, and a preprint is not a validated treatment.

RESEARCH EXPLAINER

Older research, new explainer: AI helps date pregnancies from simple scans

Google’s October 21 explainer revisits a JAMA Network Open study published July 9, 2026. In Chicago and Nairobi, novice operators used handheld ultrasound probes while AI estimated gestational age. Among 385 participants in the primary evaluation, accuracy met the study’s comparison standard against expert measurements. The approach could broaden access where specialists are scarce. This evaluated pregnancy dating, not comprehensive diagnosis, and did not establish better pregnancy outcomes.

AI Workflows | Thursday: New-opportunity identification

Find the small service customers keep asking you to offer.

Wednesday’s Picsart Marketing Studio announcement brings ad planning, creative testing and Google Ads launch into one place. That makes it easier to promote an idea. The question to settle first: does the idea answer a need customers have already shown you?

Thursday’s skill is identifying a new opportunity. The useful result is one small offer worth investigating, with evidence you can inspect and a clear reason it might fail.

Imagine you run a cleaning business. Several inquiries ask for a kitchen-and-bathroom-only visit, while your service menu offers whole-home cleaning. That could be a smaller service people want. It could also be work that loses money once travel time is included. Your agent can find the pattern; you decide whether the economics work.

Let the agent investigate the requests.

Use a company-approved assistant with access to the records you choose. For example, Claude’s Google Workspace connectors can retrieve Gmail inquiries and Drive documents without making you copy every message into chat. These connectors are available to all Claude users; Team and Enterprise workspace owners must enable them. Restrict this job to reading. If approved access is unavailable, use a small, anonymized export instead.

1. Set a narrow question. Name the inquiry label or folder, a period such as the past 60 days, and your current service list. Ask which requests you repeatedly decline or handle as one-off favors. Exclude unrelated messages and sensitive customer details.

2. Make the evidence count. Have the agent group similar requests, count distinct customers rather than messages, and link each finding to its source. Ask it to separate a passing question from an explicit request. Require it to say when the sample is too thin to support an opportunity.

3. Compare the offer with your constraints. Ask for up to three possibilities, including who needs each one, what you would deliver, and what you still need to know. Supply your actual costs and capacity before asking for a margin estimate. Choose one that uses skills and equipment you already have.

4. Prepare a small demand check. Have the agent draft a plain-language offer and three questions for a customer conversation: When did you last need this? What did you do instead? Would the proposed scope and price solve that problem? Review the offer yourself before anyone sees it. Set a success condition before the conversations. For example, you might proceed to a small trial if two of five people who previously asked about the service request a follow-up at a price you can deliver. If responses amount to polite compliments, revise or stop. This is a pilot signal, not a demand forecast.

Keep creative scores in their place.

If you later use Picsart to compare ads for the offer, its Vera documentation describes directional predictions for a US audience, including synthetic-panel research. Those scores can help improve a message. They cannot establish that your customers will buy the service. Check plan and credit costs first, keep customer conversations out of the creative tool, and require your approval before any campaign or spending.

Give your agent this assignment

For a manager, give one person responsibility for checking the sources and deciding whether a trial is affordable. AI can organize the evidence; someone still owns the promise made to the customer.

Testing note: This workflow was checked against current documentation. It has not been tested end to end for this edition.

Start by naming one inquiry folder and asking your agent to find a request your current offer does not cover.

Tool Spotlight | Early look

UseIt Studio lets you keep editing a packaging idea.

UseIt Studio is an AI design tool for physical items such as menus, stickers, cards and packaging. This is an early look based on its published features, not a hands-on review.

The company shows editable text and graphic layers, matching visual sets, 3D previews, and files intended for printing or cutting. It works in a browser and has a Windows app.

For: A shop owner or small-brand operator who wants to explore a matching menu, bag sticker or gift box before talking with a printer.

Cost: UseIt’s terms describe credits and paid subscriptions, but current dollar prices, credit allowances and export limits were not publicly verifiable. Check them in-app before starting. Subscriptions renew automatically.

Watch the rights: Gallery remixes default to personal, noncommercial use unless the design page says otherwise. Use your own artwork or confirm commercial permission. The privacy policy describes cloud processing, so keep confidential client material out of a first experiment.

Skip it if: You need a dependable, urgent production file today. A printer should check a sample export, dimensions and cut lines before you order a run.

Testing status: The app and its exported files have not been tested for this edition.

A QUESTION FOR YOU

What do people already ask you for?

What’s one request you keep hearing that your current service or product does not cover?

Aric