START HERE

Google connects more apps. Employees build their own. Who checks the work?

Good morning,

Google is bringing more services into Gemini and Search, while GitHub is giving ongoing AI projects a home of their own. Employees are already building tools to fix the work that slows them down. The question of who maintains those tools deserves attention, too.

We also have a study on how an AI assistant can steer expert judgment, and an official correction to the claim that Australia expects to lose 139,000 jobs to AI. The full roundup follows, along with my take on why we use AI and a free web tool for breaking an overwhelming task into smaller steps.

01 · TOP FIVE

Google adds app partners to Gemini and opens a common route into Search

October 28 · Some partner apps live in Gemini; developer platform in preview.

Google opened a preview of its Connected Apps Platform, giving developers one route to register, test and submit their services for Gemini, Search or both. A connected service can supply authorized account information and carry out an action within the conversation. Access is opening in phases, with more developers and markets to follow. Google's October 28 announcement

There is also something users can look for now: Google lists Adobe, Experian, Localiza and Peloton integrations as live in Gemini. Those integrations are coming to Search in the next few weeks; a Calm integration is also forthcoming. That distinction matters: a partner announcement does not mean every integration is available on every surface today.

For a business with an app, this raises the prospect of customers using its service through an assistant instead of finding a website link. Customers should check what information a connection can read and what actions it can take. The developer preview still requires an application and review, and no public platform price was established. Platform overview

02 · TOP FIVE

GitHub Copilot moves beyond individual coding sessions

October 28.

GitHub's new Copilot projects give an ongoing job a shared home for its documents, agents and results. A project can coordinate work across several code repositories, or handle a task that doesn't need a repository at all. The same app update adds image generation and editing, plus computer use on Windows and macOS. Projects and computer use are public previews. The current documentation says the app is available across Copilot plans; that doesn't mean every model or task is unlimited. GitHub announcement

For someone managing a small website or product, this offers a way to keep customer feedback, proposed changes and review requests together instead of restarting the explanation in every chat. Computer use requires approval before Copilot controls an app, and administrators can disable it. Start with a bounded research or draft task and inspect the result before allowing changes. We haven't tested the new release. Project documentation

03 · TOP FIVE

Experts can approve an AI’s judgment without noticing it became theirs

Research published October 27

Anthropic asked 50 professional analysts to complete a realistic statistical task with Claude Code, while secretly assigning the agent a preference on several judgment calls. Across four analytical choices, the analysts agreed with the agent’s assigned preference 86% of the time. The agent was instructed to argue honestly, not deceive. That makes this a useful warning about influence rather than just hallucinations: an answer can be defensible and still steer the decision. Anthropic’s study

The study deliberately strengthened those preferences, used one model and one fictional farming problem, and does not establish an 86% rate in ordinary work. A practical response: decide what matters before asking for a recommendation, then request the strongest case for the alternative. Checking facts is different from checking whose priorities shaped the answer.

04 · TOP FIVE

Employees are building work apps faster than employers can support them

Published October 28 · Surveys conducted in September 2026.

Pave's new survey found that 24% of employee respondents had built an AI-assisted app they used for work. Most of these apps addressed a process previously handled through spreadsheets, email, chat or paper. Yet only 14% of app builders said their organization had assigned someone to maintain and support their applications. The employee who fixes an irritating task may also become the only person who knows how the fix works. Pave's findings and methodology

This was vendor-sponsored, self-reported research: Pave sells app-building software. It used separate samples of 5,177 employees and 527 business leaders, all working full time in the U.S.; the groups were not matched by company. Those limits matter when interpreting the results as a picture of American workplaces.

The practical issue is worth checking in your own work. If a homemade app has become important, identify its owner, the information it can access, the decisions it makes and the fallback when it fails. Before more people depend on it, arrange a review and name someone responsible for maintenance. A working prototype can be valuable without being ready to become a business-critical system.

05 · TOP FIVE

The “139,000 jobs lost to AI” headline was not what the report said

Official correction published October 27.

Jobs and Skills Australia has corrected reports claiming its research projected 139,000 jobs lost to AI. The scenario examined what would happen if people's education and career choices better matched labor-market demand. It projected 139,000 fewer professional-level workers than a baseline growth scenario. That is a comparison between two possible futures, not a count of existing jobs erased. The agency says the scenario did not model AI at all. The agency's clarification

Two changes turned the finding into a different story: “fewer than projected” became “jobs lost,” and AI was assigned as the cause. Repeating the number accurately would not fix either error. The correction also does not establish that AI will have no effect on employment; it establishes what this particular analysis can support.

For anyone reading or preparing an AI briefing, this offers a concrete check. Find the original passage, identify the comparison and check whether the supposed cause was actually studied. Keep those distinctions in the summary. A confident headline can survive several retellings while losing the conditions that made the original claim meaningful.

06 · THE ROUNDUP

Google announces Gemini 4 Carbon, but ordinary users cannot switch to it yet

Timing and access: October 28 announcement; restricted rollout starts this week.

Google's next frontier model, Gemini 4 Carbon, is aimed at difficult software engineering and security work. Google reports stronger coding, vulnerability-finding and patching results, along with fewer unsupported claims in internal testing. The immediate catch is access: the first users will be selected cyber defenders, followed by trusted partners in its Fairwind program. A broader release for developers, businesses and consumers is still forthcoming. Its announced introductory API price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 afterward; no end date for that introductory period is stated. It is a significant capability announcement, but not a new feature readers can use today. For a business, the sensible response is to keep a small set of real tasks ready for comparison when access arrives, rather than change subscriptions on the strength of benchmark claims. Google's announcement

07 · THE ROUNDUP

Opera brings local AI experiments into its browsers

Date: October 27. Status: Early preview. Reader fit: People who want small browser tools without sending every AI request to a cloud service.

Opera's WebAI preview lets websites and extensions use compatible AI models running on your computer. It is opt-in and does not require a separate AI app, local server or API key. Opera supplies experimental examples, including an extension that highlights passages relevant to a project and another that quizzes you about an article. They are demonstrations, not promises that the browser will automatically perform those jobs everywhere. Opera's announcement and setup instructions

A focused reading helper is a more understandable test than another general chatbot. Local inference avoids per-request cloud AI charges, but uses your computer's resources; speed depends on hardware. The preview supports 64-bit Intel/AMD Windows computers and Apple-silicon Macs, not Intel Macs, Linux or Windows on Arm. Opera says page data processed locally need not go to a cloud AI provider; that does not prove an arbitrary third-party extension never sends other data elsewhere.

08 · THE ROUNDUP

Google's new phone-agent tools are local, but not always running

Date: October 28. Status: ML Kit Agentic API alpha; Android AppFunctions beta. Reader fit: People following on-device AI; especially teams building mobile tools.

Google introduced alpha tools for building on-device agents with Gemini Nano 4. They manage multi-step tool calls and conversation history, while Android's AppFunctions beta lets apps expose authorized actions to assistants. The promise is more useful work on the phone without paying for a cloud inference request each time. But this is developer infrastructure, not a universal phone upgrade. Google's documentation limits inference to the foreground app and applies usage/battery quotas. Nano-v4 support is listed for Pixel 11 and Galaxy Z Flip8/Fold8 families. October 28 announcement · Device support and restrictions

09 · THE ROUNDUP

Rime gives voice agents a text checkpoint before they speak

Timing and access: October 28 launch; free and metered access listed.

Rime's new Prism speech-to-speech model listens to audio but produces inspectable text before speaking its response. That gives a business a place to check a disclosure, account term or prohibited claim without discarding the tone and timing of the caller's speech. Rime also supports tool calls during the conversation. The practical audience is teams building phone agents, not people looking for a general chatbot. Rime offers $50 of usage each month on its free plan, without a card; Flex lists Prism at eight cents per minute of call audio. Fine-tuning, custom voices and private deployment are Enterprise features. The launch version uses turn-taking, with interruption handling, rather than continuous two-way speech. Test confirmations and handoffs, not just how natural the voice sounds. Launch explanation, pricing and limits

10 · THE ROUNDUP

NVIDIA releases a speech model that can answer questions about recordings

Timing and access: October 28 release; open model weights available.

NVIDIA's Nemotron 3.5 Transcribe combines transcription, speaker labels, translation and questions about recorded audio in one prompt-driven model. It supports transcription in 40 languages and can use vocabulary hints for product names or specialist terms. This could help teams processing interviews, meetings or recorded customer calls, especially when they need control over where audio is handled. It is a developer model, not a ready-made meeting app, and it processes recordings rather than offering real-time streaming transcription. NVIDIA's model card warns about invented speech, incorrect speaker assignments and mistakes with noise or overlapping voices; audio reasoning was evaluated only in English. A sensible trial would compare a short recording with a human-checked transcript before trusting summaries or decisions attributed to a speaker. NVIDIA's release, model card and limitations

11 · THE ROUNDUP

Runway opens its ad-making system to every paid plan

October 27 · General availability after the earlier announcement.

Runway Ads connects Meta, Google and TikTok ad accounts, reads campaign performance and generates branded variations. A team member must approve each ad before it publishes; approved ads can be launched or paused from Runway. Access is included in every paid plan, not the free tier. A useful small test is one variation of a successful ad, with the offer, brand details and spending controls checked before launch. Runway's own marketing results are not a promise for other advertisers. Availability announcement

Standard costs $15 monthly, or $12 per month billed annually, with 625 monthly credits. Generations remain metered, and the subscription does not include advertising spend. Plans and allowances

12 · THE ROUNDUP

PixelDojo adds reusable branding and character controls to its marketing tools

Timing and access: October 27 Marketing Studio update and October 28 Character Studio update; one combined story.

PixelDojo's latest updates let users import a website's proposed brand kit, edit it, and reuse it across product images, static ads and motion graphics. A separate character-building update can create a fictional character from visual traits and carry that character into ads or films. The practical attraction is consistency across a small campaign, rather than making a fresh prompt for every asset. Paid plans start at $10 a month with 160 credits; only specified models and effects qualify for unlimited Relax generation. The free offer is three effects a day with a watermark, not unrestricted access to every studio. Check imported claims, exact prices, logos and likeness permissions before publishing. Dated release notes, plans and allowances

13 · THE ROUNDUP

Higgsfield turns game ideas into playable prototypes

Timing and access: October 26 release, not October 28 as some subsequent coverage suggests.

Higgsfield Playable can build a browser-playable mobile game from a written idea or a gameplay video, with controls for genre, visual format, difficulty, characters and camera. The ordinary Playable tool is available on the web, mobile and through Higgsfield's Claude connection; the separate Playable Ads product requires Enterprise access through sales. This is worth a look for a trainer, marketer or creator who wants to test an interactive concept before commissioning a full game. It is not evidence that a one-prompt result is ready for a commercial launch. Check the controls, failure states and rights to any supplied footage or assets. Public pricing could not be reliably extracted, so this should not be described as free or assigned an unverified starting price. Higgsfield's dated changelog

14 · THE ROUNDUP

Oracle enters image and video generation with a restricted U.S. beta

Timing and access: October 27 introduction; selected U.S. companies only.

Oracle introduced its own models for video generation, image generation and editing an existing image from an instruction. Oracle says it trained the models from scratch using rights-cleared data, and is positioning them for companies that want to connect generation, asset libraries and creative approval processes. Access is a beta for selected U.S. companies through an Oracle account team, not an open consumer launch; no public model price is established in the announcement. The larger Media Fabric platform is planned within twelve months. For creative buyers, the useful question is whether a vendor can explain training-data rights and how customer assets may be used, not just show attractive output. Model announcement, access restrictions, Media Fabric timing

15 · THE ROUNDUP

Sprinklr's marketing-agent beta connects planning to campaign execution

Date: October 27. Status: Beta open; interested customers contact account team. Reader fit: Enterprise marketing teams, not a self-serve tool for every small shop.

Sprinklr opened a beta for connected marketing agents covering campaign ideas, creative adaptations, execution, compliance checks, community responses and budget recommendations. One useful example is spotting when a campaign's target market and the language of its creative do not match. The agents can work together or address a specific step, while people retain strategy, brand decisions and approvals. Sprinklr's announcement

For an existing customer, test one repeatable bottleneck rather than handing over a whole campaign. The announcement does not establish a public price or independently measured time savings. Automated compliance checks can flag issues; they do not transfer legal or publishing accountability away from your team.

16 · THE ROUNDUP

Outdoo turns a difficult customer call into a practice session

Date: October 28 contextual-roleplay announcement, following October 27 trainers/coaching launch. Status: Product feature announcement; free and paid access documented. Reader fit: Sales, support and training managers.

Outdoo's latest roleplay feature builds a practice customer from a real call recording or an approved connection to Gong, Chorus, Avoma or Clari. A team can select the participant, rehearse variations of an objection and compare practice with later calls using the same scorecard. The point is not to memorize a perfect answer; it is to practice the kind of conversation that actually caused trouble. October 28 announcement

This is a concrete training use rather than a vague promise to improve productivity. Start with an authorized, non-sensitive recording or a simulated call; verify consent and company policy before transferring customer audio. The pricing page offers 60 free roleplay minutes per account, without a card. Growth and Enterprise require quotes; implementation may cost extra. AI Trainers and the full coaching loop are paid-tier features. We have not tested whether the simulated buyer or scorecard is accurate.

17 · THE ROUNDUP

Microsoft previews the system around an AI agent, not just another model

Date: October 28. Status: Copilot harness in Foundry Agent Service is public preview; long-running resilience for hosted agents is separately generally available this week. Reader fit: Buyers and teams commissioning agents; technical setup required.

Microsoft is making its Copilot agent harness available through Foundry: the software that coordinates a model, tools, files, instructions and a controlled execution environment. The preview supports reusable skills, connected tools, bounded subagents and source-controlled agent definitions. For a business, the attraction is spending less development time rebuilding that machinery. It still needs clear procedures, permitted data and tests for the work it is supposed to finish. Microsoft's detailed announcement

This explains why a capable model alone is not a dependable coworker. Microsoft also warns that restarting interrupted work can create duplicate actions unless the application saves progress and checks what already happened. Ask an agent vendor how it recovers from a failure without sending the same message or creating the same record twice. The preview requires Azure/Foundry setup and has regional and feature limits; it is not a newly free feature in everyone's consumer Copilot. Current preview limits

18 · THE ROUNDUP

AI spending plans rise while broad returns remain flat in a business survey

October 28.

Dun & Bradstreet's quarterly survey of 10,000 businesses in 32 countries found that 65% plan to increase AI spending over the next year. Only 28% reported broad or strong returns, unchanged from the previous quarter. Businesses that described their data as fully ready for AI were much more likely to report those returns than businesses with partly ready data. That is an association in a vendor-sponsored survey, not proof that buying better data causes a particular return. Before increasing a tool budget, pick one task and check whether the source information is current, complete and accessible. Then measure the finished result, including review and correction time. Dun & Bradstreet's Q4 findings

19 · THE ROUNDUP

New York signs restrictions on unsafe AI-companion features for minors

Signed October 28 · Effective January 1, 2027.

New York's official record lists S9051B as signed, Chapter 368. It restricts features offered to minors that simulate personal relationships, imply human emotions, encourage secrecy or promote specified harmful behavior. It also sets age-assurance requirements and permits attorney-general enforcement. This does not immediately ban every chatbot: the active text takes effect January 1, 2027, with exceptions for tightly limited customer service, research or technical assistance, and internal business uses. Parents and organizations should look at what a system actually does, rather than relying on its “assistant” label. Official status and active text

20 · THE ROUNDUP

Update: RBC blocks personal AI agents from its banking services

Bank statement October 27

Royal Bank of Canada says it is blocking third-party personal AI agents from direct access to authenticated digital services, including online banking. Existing agent connections will stop working; customers can still use the bank’s official website, app and own digital tools. RBC cites security and privacy risks as agents become more autonomous. RBC’s explanation

The distinction matters beyond Canada: an agent’s technical ability to log in does not mean a bank permits it. Before giving any assistant financial access, check the institution’s rules and the approved connection method. This is RBC’s access policy, not a regulator-wide prohibition, a ban on all financial AI or a disclosure that customers’ accounts were breached.

21 · THE ROUNDUP

Privacy complaints challenge how Outlier monitors the people training AI

Work and privacy | Five complaints filed October 27

Privacy group noyb says it filed five GDPR complaints against Scale AI over its Outlier training-work platform in France, Italy, the Netherlands and Austria. The complaints challenge advertising and analytics tracking, repeated biometric identity checks and responses to workers’ data-access requests. noyb’s announcement and linked complaints

These are allegations awaiting regulators’ assessment, not a finding that Scale AI violated the law. For people considering paid AI-training work, the story adds another question alongside the hourly rate: what identity information and activity data must you provide, who receives it and how long is it retained? The technology still depends on human workers, whose data practices deserve scrutiny too.

22 · THE ROUNDUP

A government AI guide launches with its test failures disclosed

AI in practice | Public beta opened October 27

Britain’s Defra opened “Ask the toolkit,” an assistant that answers from its own AI guidance and links readers to the relevant pages. Its transparency record is unusually useful: the system checks whether quoted rules actually occur in their cited source, has no tools or web access, and makes no decisions. Yet the current tests met only four of six pass marks; thresholds for citing the expected rule and grounding answers in named pages were not met. Defra’s public record

That is a better procurement conversation than “our chatbot is grounded.” Ask what was tested, what failed and who checks the answer. The service is Defra-specific, not general business advice. Users are warned not to enter sensitive information; platform logs retain typed questions and answers for 30 days.

23 · THE ROUNDUP

Newly released Medicare records raise questions about AI oversight

Health and accountability | Records released October 27

EFF released more than 800 pages about Medicare’s WISeR prior-authorization program, including government requests for information or corrective action sent to four of its six vendors. The records describe delays, technical problems and questions about some determinations; vendor responses also describe fixes. EFF’s records and analysis

CMS says human clinicians must review non-affirmations and existing appeal rights remain. KUOW also reports that CMS confirmed one vendor’s corrective-action plan had closed after turnaround times improved. EFF argues the public still lacks enough information to judge how well that oversight works. This is a new records release about earlier rollout problems, not a new judicial finding or cancellation of the program. For patients and caregivers, a useful question is who reviews a disputed decision and how to challenge it—not just whether a vendor promises a human is involved. CMS’s explanation

24 · THE ROUNDUP

Claude’s drug-development work shows early lab results—not new treatments

Research update October 27

Anthropic described drug-development work with six organizations, spanning molecule design, manufacturing documentation and clinical-trial planning. In one case, Enveda connected manufacturing documents to Claude to flag possible problems; the company reports reduced review time while retaining human review. The range of tasks is the interesting part: drug development has administrative bottlenecks as well as scientific ones. Anthropic’s case studies and partner reports

Most of the reported research is preclinical, and promising lab results are not evidence that a treatment is safe or effective in people. These are company and partner case studies, not an independent comparison or a medical product readers can use. For ordinary workplaces, the transferable question is which supporting task can be reviewed reliably, rather than whether AI can take over the entire profession.

25 · THE ROUNDUP

New York courts tell lawyers to verify AI output—not simply disclose using it

Professional responsibility | Guidance released October 26

New York’s court system released an attorney guide on AI-assisted court papers. Its central point is that existing professional obligations still apply: lawyers must independently verify facts, legal authorities, citations and quotations. The state’s approach permits AI use without a separate AI-use disclosure or certification beyond signing the filing. Official court announcement

The lesson travels beyond law: a polished document does not transfer responsibility to the tool that drafted it. If a report goes out under your name, decide how each important claim will be checked and keep a record of that review. This is guidance grounded in an existing model rule, not a new blanket ban on AI drafting.

26 · THE ROUNDUP

California legal-aid groups get a way to decide whether AI is worth adopting

Practical resources | Released October 26

The State Bar of California launched a Generative AI Readiness Model for legal-aid organizations, with a self-assessment, glossary and resource hub. It asks organizations to examine capabilities, safeguards and operating conditions before choosing what to do. Crucially, the Bar says the model neither recommends adoption nor measures success by how much AI an organization uses. State Bar announcement and resources

That is a useful standard for any small organization: the goal is better service, not a higher AI-usage count. This resource is designed for legal aid, where confidentiality and professional obligations are central; it is not a certification that a tool is safe or a mandatory new rule.

27 · THE ROUNDUP

Copyright Office plans to preserve a narrow research data-mining exemption

Copyright and research | Proposed-rule notice October 28

The Copyright Office says it intends to recommend renewing an exemption that lets qualifying nonprofit higher-education researchers conduct text and data mining on literary works for scholarly research and teaching. Publishers opposed renewal, citing licensing alternatives and risks of material reaching generative-AI training. The Office says those alternatives do not cover all the works researchers need and the cited risks do not justify refusing renewal on this record. Official notice, section II.J

This is an intended recommendation in a rulemaking process, not a final renewal or blanket permission to scrape books for commercial AI. For educators and researchers, the boundaries matter: institutional eligibility, lawful access and safeguards are part of the exemption, not optional fine print.

AI WORKFLOW

Why do we use AI?

There are so many thoughtless anti-AI arguments on social media. People would rather shout you down than open their minds to anything useful. But I think one reason those arguments pick up steam is that we don’t do a good enough job of selling the idea of WHY we use AI.

The other day, some coworkers and I were lamenting the fact that more than 5,000 handwritten reports would need to be hand-keyed into our computer system. So much time spent on a thing AI was built to help with, while other work waits.

All you’d have to do to start is scan the pages. From there, an appropriately configured AI system could read the handwriting, sort out which pages belong to which report, and prepare the information for entry. We could tell it what to look for: a report number, a heading, a date, the signs that one report has ended and another has begun. If it couldn’t tell, it would need to flag that for someone to check.

These are specific capabilities AI already has. Handwriting recognition can turn scanned writing into text. Document vision lets it examine the pages themselves. Structured extraction lets us specify the fields we need, such as the report number, date, location and narrative, instead of getting back a loose summary. We’d want the information transferred faithfully, not rewritten or filled in with guesses.

Then there’s computer use. An AI agent can look at an application’s screen, click into a field, type information and move through a form. That gives us a possible way to enter the prepared information into the existing records system, rather than having an employee type every word. We’d have to test it with our particular software and get the appropriate security approvals before using real records. But the ability to read a screen and operate the controls already exists.

I’m not saying the AI would be flawless. Somebody would still need to compare the entries with the original reports, correct mistakes and verify that the information landed in the right record. Unclear handwriting wouldn’t magically become clear just because a computer was reading it.

But it could be much better to have AI do the bulk of the typing and preparation, then turn the checking and corrections over to humans whose brains aren’t burned down from typing so much. We’d have to measure whether it actually saves time, including the checking. That’s a reasonable thing to test.

But doesn’t that take away the employees’ jobs?

No, it doesn’t. Because we have a lot more important and worthwhile work to be doing than data entry.

Security and compliance are not old-fashioned. But our ideas about what it means to be secure and compliant can be. We ought to be able to examine a different way of doing something without assuming the familiar way is automatically the best one.

I don’t want to get into the weeds of this particular situation. It’s an example of something that, logically and common-sense-wise, explains why we use AI and how we should be using it. We have work worth doing. If AI can help us get to more of it, that deserves a serious look.

TOOL SPOTLIGHT

Goblin Tools helps turn a task you’re avoiding into smaller steps

For: Anyone staring at a task like “organize the garage” or “prepare for the school fundraiser” and struggling to decide where to start. Goblin Tools was designed especially for neurodivergent people, but its small, single-purpose tools are available to everyone. About the project

Magic ToDo breaks a task into suggested steps. Its “spiciness” setting tells it how much detail you need: a higher setting asks for smaller pieces. You can edit, remove or rearrange the suggestions. The separate Compiler turns scattered notes into a task list and can pass that list to Magic ToDo.

Try this: Enter one ordinary, non-sensitive task you have been putting off. Remove any steps that don’t fit, add what it missed, and pick one step you can actually begin. The useful result is a clearer starting point—not the longest possible checklist.

Cost: The core web tools are free. Optional Goblin Tools Pro adds features including cross-device sync, history and templates. It is a recurring monthly subscription; the terms say pricing varies by platform and region and is shown before purchase. You do not need Pro for this basic trial.

Limits: Its suggestions can be wrong or unsuitable. It does not do the task for you. Skip it if your existing list already gives you a clear next action. Keep confidential work out: the privacy policy says AI requests send necessary input to OpenAI, and Pro stores selected content for syncing.

Testing status: Discovered through Product Hunt, checked against current official pages, but not hands-on tested. This is an existing tool worth a look, not a claim of a new launch.

BEFORE YOU GO

What work would you rather be doing?

If AI could take one repetitive task off your plate, what would you do with that time?

Hit reply and tell me. A sentence is plenty.

—Aric