mAIn Street #354: ChatGPT Wallet gives your agent purchasing power on your behalf; Businesses can now claim ChatGPT profiles; AI freelance projects are now on the rise


mAIn Street #354 | Approval before action
mAIn Street Daily Newsletter. AI news for people who actually have jobs to do.

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 anything.

Today’s top five

1. ChatGPT Wallet puts your approval between an AI agent and a purchase

OpenAI introduced ChatGPT Wallet, a way to let an AI agent handle checkout without handing a merchant your actual card number. The initial rollout is for eligible U.S. ChatGPT Pro users age 18 and older using dots, with supported U.S. payment cards.

The important part is the approval step. You review the items, total and payment card, then authorize the purchase through a security check. A payment token stands in for your card details. This is not a launch of unrestricted, autonomous spending. OpenAI describes broader delegated purchasing as a future direction.

For business owners, the question is practical: can a legitimate AI shopper complete your checkout without losing the information a customer needs to approve the order? Clear prices, delivery terms and cancellation rules matter just as much when software handles the clicks. For buyers, keep the distinction clear: help completing a purchase is not permission to choose what you spend.

Read the source · 2026-11-13

2. Businesses can update their information and see how customers discover them in ChatGPT

OpenAI’s Business Hub beta gives verified businesses a way to manage information used in ChatGPT, including business details and product feeds. It also offers aggregate measures of discovery, such as mentions, clicks and topics, subject to reporting thresholds.

That could help a small business answer a question that ordinary website analytics misses: what happens before a potential customer reaches the site? The reporting does not expose individual conversations, and claiming a business does not guarantee that ChatGPT will recommend it.

OpenAI is also introducing business agents, initially for eligible U.S. ecommerce and online-service businesses. Availability depends on the program, so do not assume every local business can switch one on today. The useful first step is less glamorous: check whether your business qualifies, then make sure your descriptions, products and service details are accurate. Treat discovery data as another signal about customer questions, not a complete account of what customers think.

Read the source · 2026-11-10

3. AI freelance projects are getting larger, but that does not mean higher hourly rates

Research from OpenAI and Upwork suggests AI-related freelance work is expanding in scope. The researchers found larger fixed-price project budgets and more requested hours, while posted hourly rates were similar. The range of tasks in AI-related job postings became 45% more diverse between early 2023 and late 2025.

That distinction matters if you sell writing, training, research or other professional services. A larger project can bring in more revenue without paying more for each hour. It may also require more coordination, review and responsibility than the original assignment.

The research also found that some kinds of work commonly performed with ChatGPT lost share on Upwork. These are platform-specific findings, not proof that AI caused every change or that all freelancers benefit. The practical takeaway is to define the completed result you can deliver. A checked research brief or a working staff-training process gives a buyer something clearer to evaluate than a promise to “use AI.”

Read the source · 2026-11-10

4. Healthcare AI agents take on the work between disconnected systems

OpenAI is expanding ChatGPT for Healthcare with more than 20 healthcare plugins, improved clinical search and computer-use capabilities. The aim includes a familiar administrative problem: useful information is scattered across systems that do not work together neatly.

In a companion announcement, CVS reports that an agent cleared a backlog over a weekend, while Cigna describes work involving information across more than 10 systems. Those are company-reported examples, not independent evidence that the same gains will appear in every organization.

For other industries, the lesson is to look at the handoffs. Copying information between approved systems can consume more time than writing the final answer. Healthcare adds a much higher bar for privacy, accuracy and review. The relevant protections depend on eligible services, a business associate agreement and the right settings. These announcements are not permission to put patient information into a consumer chat account or let an agent make clinical decisions without qualified oversight.

Companion source: healthcare computer use

Read the source · 2026-11-16

5. Adobe Firefly turns repeated creative steps into reusable workflows

Adobe’s Firefly expansion puts more attention on repeating good work, not just generating another image. Firefly Graph, in beta, lets users build reusable creative workflows. Creative Agent adds features including brand preferences and a planning mode.

Availability varies across the announcement. Edit Text, Retouch, Relight and Quick Actions are generally available. Firefly Builder is in private beta with a waitlist. A demonstration of a connected workflow should not be mistaken for a feature every subscriber can use today.

For a small marketing team, the useful test is consistency: can you turn an approved starting asset into several versions without changing the message, losing the brand’s look or creating fresh errors? Start with a recurring task such as adapting a promotion for different placements. Keep a person responsible for the wording, claims and final approval. Saving a workflow is valuable only after you know its output survives that review.

Read the source · 2026-11-10

23 more stories worth your time

6. Adobe previews AI work that ends in editable creative files

Adobe is extending Creative Agent across more Creative Cloud products and previewing connections that could let outside AI tools work with native creative documents. Editable layers, vectors and sequences would make revisions easier than a flattened image. The connections were previewed, not broadly released; several individual app features also remain in beta.

Read the source · 2026-11-10

7. Adobe’s campaign agents aim to carry a brief through production

Adobe describes a coming CX Enterprise Coworker and other GenStudio tools for work involving briefs, audiences and asset variations. Brand Intelligence supplies context, while enterprise Firefly Graph capabilities are in beta. The appeal is fewer handoffs between planning and production. This is a mix of announced and beta capabilities, not a finished campaign department in a box.

Read the source · 2026-11-10

8. Adobe and Google plan closer document and creative connections

The companies say Acrobat will become available within Google Drive and Gmail, alongside additional Gemini connections and Google models in Adobe products. The proposed benefit is fewer trips between apps to read, edit or create. Check the specific release and account requirements before planning work around it: the partnership announcement includes future availability.

Read the source · 2026-11-10

9. Google positions Gemini 4 Carbon for harder, longer tasks

Google’s Carbon release emphasizes deeper reasoning for difficult work. Access is rolling out through selected products and plans, including Gemini Ultra, with introductory API pricing of $2 per million input tokens and $10 per million output tokens. Google still points to Argon for most work. Test whether the harder task actually needs the more expensive model.

Read the source · 2026-11-10

10. Anthropic reserves its most demanding models for more demanding work

Anthropic introduced Claude Fable and Mythos 5.2 for difficult technical tasks, with organization verification required for Mythos access. The company reports speed improvements over the previous generation, but recommends starting with Opus 5.5 for everyday knowledge work. Higher capability does not automatically make a model the right default for routine drafts, summaries or office questions.

Read the source · 2026-11-09

11. Anthropic launches a proposed IPO roadshow, not a completed offering

Anthropic announced a roadshow for a proposed offering of 825 million shares at an expected $70 to $78 per share and said it had applied to list on Nasdaq as ANT. At the announcement, the registration statement was not yet effective. Those are proposed terms, not a confirmed sale price or evidence that the offering has closed.

Read the source · 2026-11-10

12. OpenAI gives outside assessors closer access to its safety work

OpenAI describes an initial phase of embedded assessments involving Apollo Research, METR, Transluce and Goodfire. Selected access to models, internal systems and records allows more sustained examination than a short demonstration. The scope differs by assessor, with some follow-on work still being arranged. Outside access is useful evidence about scrutiny, not a blanket safety certification.

Read the source · 2026-11-16

13. LVMH’s media division pairs AI tools with attributed publishing content

OpenAI’s partnership with LVMH’s media division includes tools for tasks such as research, comparison and preliminary translation, plus attributed summaries and excerpts linking to publications including Les Echos and Le Parisien. For publishers, the questions remain editorial: who checks the work, where the original reporting is credited and whether readers can reach the source.

Read the source · 2026-11-16

14. OpenAI plans another 50,000 researcher grants starting November 23

The next round of ChatGPT grants for eligible academic researchers is scheduled to open November 23 on a first-come basis. It provides Plus access for recipients and up to two collaborators, unlike the earlier Pro pilot. Reported time savings from participants are self-reported, not controlled proof of an effect. Eligibility and the application window matter more than the headline total.

Read the source · 2026-11-16

15. Veterans’ ChatGPT benefit passes 20,000 participants

OpenAI says more than 20,000 transitioning service members and veterans claimed a free year of ChatGPT Plus during the program’s first year. This is an anniversary update to an existing benefit, not a new offer for every veteran regardless of circumstances. Check the current eligibility rules before counting on access for career preparation or study.

Read the source · 2026-11-11

16. A student survey points to an AI guidance gap

OpenAI cites ThinkYoung research in which 79% of surveyed students used AI for assignments or homework, while 47% had been taught how to use it. The announcement concerns education efforts in Europe, the Middle East and Africa; the figures should not be generalized to all U.S. students. The workplace parallel is worth considering: access to a tool is not the same as instruction.

Read the source · 2026-11-10

17. The King’s Trust and OpenAI target AI training for 25,000 young people

A £1 million partnership aims to reach 25,000 young people in the United Kingdom over 18 months, with priority given to those outside education, employment or training. Work includes developing qualifications. It is a training and access commitment, not a guarantee of jobs, and it is not a U.S. enrollment program.

Read the source · 2026-11-16

18. Amazon expands shopping by text, including approved purchases

Amazon’s updated Alexa for Shopping announcement describes text-based product searches, cart management, price tracking and purchases with approval. The November 13 update builds on an earlier announcement rather than introducing an entirely new service that day. For retailers, clear product information matters when a customer’s first interaction is a text request instead of a visit to a product page.

Read the source · 2026-11-13 update

19. AWS commits $100 million in technology support to health projects

AWS’s five-year Health Impact Initiative offers cloud credits and technical support to qualifying health organizations, including nonprofits, public agencies, researchers and businesses. This is not a $100 million cash-grant fund. Applications are reviewed weekly, and credit duration and project costs deserve attention: a funded pilot still needs a plan for operating after the support ends.

Read the source · 2026-11-09

20. Amazon adds detail to its workforce training push

Amazon’s Future Ready 2030 update outlines credentials and learning routes involving institutions including Stanford, Purdue, Western Governors University and SANS. It builds on an existing $2.5 billion commitment rather than announcing another investment of that size. Some opportunities are employee benefits; public AWS learning is a separate route. Check who can enroll before forwarding a program as free training for everyone.

Read the source · 2026-11-12

21. IBM and Microsoft package AI agents around industry-specific work

IBM’s Imagine Enterprise offering brings together consulting, reusable assets and agents built with Microsoft technology, initially emphasizing financial services, insurance and retail. The point is to fit software to an industry’s actual processes. This is a consulting-led offering, not a simple self-service app that a small business can install and expect to run its operations.

Read the source · 2026-11-10

22. IBM’s agent-security plans focus on identity and access

Following its AuthMind acquisition, IBM describes planned Vault integration to help identify unregistered agents, examine access patterns and revoke permissions. The plain-English issue is straightforward: an agent with a valid login may still be doing something it should not. Keep an inventory of agents and their permissions rather than treating successful authentication as proof that every action is authorized.

Read the source · 2026-11-12

23. Fed Governor Barr flags risks in how AI infrastructure is financed

Michael Barr’s remarks discuss debt, circular financing arrangements and commitments outside conventional balance-sheet measures in the AI buildout. These are his views, not a new Federal Reserve rule. For business readers, separate two questions: whether an AI tool helps your work and whether the investments supporting its industry are financially sound. One does not settle the other.

Read the source · 2026-11-16

24. Pony.ai launches paid, round-the-clock driverless Toyota rides in Guangzhou

Pony.ai says its seventh-generation Toyota bZ4X robotaxis have entered paid, 24-hour driverless service in Guangzhou through PonyPilot and partners. A wider deployment of 1,000 vehicles across Chinese cities is a plan, not a completed fleet. The location matters: this is evidence of a specific service rollout, not a claim that driverless taxis are available everywhere.

Read the source · 2026-11-16

25. Northern Trust’s AI example starts with redesigning data preparation

A Microsoft account says Northern Trust reduced one data-preparation process from two and a half weeks to three days, with 96.5% of tests passing on the first attempt. These are reported customer results, not a promise for every organization. Analysts remain responsible. The useful lesson is to measure an entire checked process rather than the seconds needed to generate an answer.

Read the source · 2026-11-12

26. KT turns experienced technicians’ knowledge into help for network teams

Microsoft describes KT’s AIO-Net work with veteran technicians known as Meisters. Their knowledge and procedures help agents investigate network problems and present conclusions for human review. For managers, this is a useful place to look for an AI project: recurring decisions that depend on experience currently held by a few people, with those experts helping define and check the output.

Read the source · 2026-11-11

27. Ohio State and Microsoft set a five-year plan for care, teaching and research

The partnership covers hospital technology, data infrastructure and training. Clinical examples include Dragon Copilot drafting notes that clinicians review. Much of the announcement is a roadmap, including a planned migration of a substantial share of systems to Azure. Keep implemented tools separate from future ambitions when comparing it with your own organization’s readiness.

Read the source · 2026-11-10

28. Saint-Gobain uses AI to narrow a difficult materials-research search

Microsoft describes researchers working with a collection of 62,000 patents and 44,000 abstracts to investigate a longstanding materials problem. The reported benefit is faster organization and exploration of possible approaches. A promising candidate is not a commercially proven formula. Laboratory testing still decides whether an AI-assisted suggestion works outside the document collection.

Read the source · 2026-11-10

AI Workflows | Monday: Capability mapping

Your AI can compare supplier bids while you keep control of the purchase.

A supplier sends a low quote. Then you notice the freight charge, the minimum order and the delivery date. Comparing offers means finding the differences that change the decision.

Anthropic’s November 12 manufacturing release gives this work a useful test case. Its plugin can compare supplier bids and retrieve records through connected business systems, with a person reviewing the result. The plugin requires a Claude Team or Enterprise plan. You can use the same method with an approved agent that already has access to your business documents.

Map the job before handing it over.

Capability mapping means separating the parts an AI can perform from the parts you need to check or decide. Start with one purchase you understand well, such as replacing office printers. Tell the agent where the current quotes and requirements live. Have it test whether it can read them before promising a comparison.

Delegate: Find the approved quotes, identify their dates and revisions, and build a comparison with links to the source pages.
Review: Check quantities, taxes, shipping, service coverage, delivery promises and any calculation that changes the ranking.
Keep: Decide which tradeoffs you can accept, choose the supplier and approve any commitment.

Use your company-approved account and the narrowest useful access. A connector lets the agent retrieve material from another app; it does not guarantee that your account can see every attachment. If access fails, have the agent name the missing document. An approved, redacted export is a fallback when a direct connection is unavailable.

Give it a bounded assignment.

Prepare a comparison of the current printer quotes in the approved purchasing folder. Read the requirements there and use only that folder. List each supplier’s total stated cost, exclusions, delivery date, warranty and unmet requirements. Link every figure to its document and page. Show calculations separately. Mark missing information as unknown; do not fill gaps with estimates. Flag conflicting revisions. Return a comparison and a short list of questions for my review. Do not contact suppliers, place orders or change any records.

Make it earn the next assignment.

Check every cost and delivery term against the originals on this first run. A useful result includes all the quotes, makes its arithmetic inspectable and leaves unanswered questions visible. A neat table that quietly assumes free shipping fails the test.

Managers can turn a successful comparison into a shared review checklist. Keep a named person responsible for the supplier decision, and measure time saved after checking the work.

Pick one small purchase today. Ask your agent to verify its access and list what it can complete before you authorize the comparison.

Tool Spotlight | Granola voice typing

Granola’s voice typing turns speech into text in the app you are already using. The company lists it as free for all Granola users, including the Basic plan, with unlimited words on Mac, Windows and iOS. Android is listed as coming soon.

A useful first test: dictate a short project brief with the goal, audience, constraints and deadline. Read the result before pasting or sending it. Names, figures and commitments deserve a second look even when the prose sounds polished. This is dictation for your own writing, distinct from recording a meeting. Use an approved account and avoid confidential material until you have checked your organization’s rules.

The point is to get your thinking onto the page with less typing, not to skip the edit.

Explore Granola voice typing · Availability checked November 16, 2026

This week, pick one task with a clear finish line. Let the AI do a bounded first pass, then check the parts that could cost someone time, money or trust.

mAIn Street

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

Read more from mAIn Street
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,...

Customer discovering a neighborhood shop through an AI-assisted business listing

ISSUE #351 · WEDNESDAY, NOVEMBER 11, 2026 OpenAI is giving businesses a place to manage their information in ChatGPT. Claude is adding spreadsheets and whiteboards, and a new insurance product tackles the risks of AI taking action. Plus: a Wednesday workflow for testing whether a rush job fits your team’s week. Today’s top five 1. Businesses can now claim their place in ChatGPT OpenAI launched Businesses in ChatGPT in beta, with Business Hub for claiming listings, updating information and...