mAIn Street #299: One bad AI experience could run off half your customers; Meta planned to lay off 60% of its workforce before productivity imploded; Google adds pay-as-you-go to Enterprise accounts



Meta's AI workforce overhaul backfires, Google adds pay-as-you-go Gemini, and U.S. funding pushes remote robotic stroke care.
mAIn Street Daily Newsletter. AI news for people who actually have jobs to do.
Thursday, August 27, 2026
On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment produce work that feels specific to you. See the workshop details and register.
Top 5
What matters today
01
Reuters found Meta went through with a 10% layoff in May and canceled a second wave planned for November. Internal data showed code changes jumped 220% while user-facing improvements rose only 36%, and employee sentiment fell from 74% to 55%.
Source: Reuters
02
Businesses can combine seat licenses with usage billing, set project caps and get alerts before agent costs run away. Select workloads will eventually run off-peak for up to half the normal inference cost.
Source: Google Cloud
03
ARPA-H selected Philips to work with Johns Hopkins and Boston University on remote-assisted and increasingly automated endovascular procedures. The aim is to extend specialist stroke care to U.S. hospitals that cannot keep an expert on site.
Source: Philips
04
Another 38% would allow only two or three mistakes, and 47% blame the business when AI fails. Customer-facing automation puts the company's reputation on the line every time the system answers.
Source: Storable
05
Camera towers now capture SKU and lot data from incoming pallets and send it to Lineage's warehouse system. The installation replaces a receiving process built around manual scans, document checks and data entry.
Source: Kargo
AI Workflows | Thursday: Find work that was too expensive to do before
Turn buried promises into a daily follow-up list.
Let an approved AI work tool search for loose ends. You decide what becomes a task, and no message leaves without you.
Laptops and scattered sticky notes on a busy worktable

Google began rolling out Ask Gemini in Chat this week. It can search approved Gmail, Drive and Calendar information, catch you up on conversations, and help manage tasks. That makes a useful new job possible: checking the workday for promises and questions that never reached a to-do list.

Most workers handle this with memory, flagged emails and notes written wherever paper was handy. A complete check would mean rereading days of email and chat, opening linked files and comparing everything with the calendar. Few people have time for that. Hiring someone to do it would cost more than many of the missed items are worth. The win is simple: fewer forgotten commitments and less time rebuilding the day from memory.

Give yourself a loose-end check

At the end of the day, have the tool look only through work sources you are allowed to use. It should find things you promised to send, questions that may still need an answer, dates mentioned in passing, and work waiting on another person.

The result is a private review list, not an automatic command center. Each item should show what may be owed, who is involved, the date if one was actually stated, and a link to the original message or file. “No date stated” is better than a made-up deadline.

Delete false alarms. Fix anything the tool misunderstood. Approve the real items one at a time. Only then may the tool create tasks and prepare short follow-up messages. Those drafts stay unsent until you read them.

A worker checking completed items in a handwritten notebook

Give the agent this assignment

Search only my approved work email, chat, files and calendar from the last five workdays. Build a private loose-end list. Look for things I said I would do, questions to me that appear unanswered, dates or next steps, and items waiting on someone else. For each item, show a short description, the people involved, the stated date or “No date stated,” why you included it, and a link to the source. Never invent a deadline or claim that a reply is missing unless you checked the full conversation. Mark possible duplicates. Sort the results into Do, Reply, Waiting and Maybe. Ask me to approve, edit or reject every item. After I approve an item, create the task and draft a brief follow-up in my normal tone. Never send a message, change a source file or contact anyone without my approval.

Test it on last week

Run the check on the previous five workdays. Open every source link and verify every claim. The test passes if it finds at least one real loose end that was missing from your list, takes less than ten minutes to review, and invents nothing.

Stop if the tool cannot show where an item came from, repeatedly mistakes finished work for open work, or reaches information outside your approved access. A longer list is not a better result. A short, accurate list is.

For managers: Start with volunteers and one week of low-risk work. Count real items found, review time and false alarms. Do not use the list to judge employee performance. Its job is to help workers keep commitments, not monitor them.
AI Tool Spotlight
Keep your script near the camera while the teleprompter follows what you actually say.
Tellie is a Mac teleprompter that sits beside the camera and tracks the exact words you say. It can follow skipped lines and ad-libs, check off must-cover points, warn when time is running short and review your pace after a take.
For webinars, training videos, remote briefings and recorded updates, that means better eye contact without pushing a script to a cloud service. Speech recognition runs on the Mac, the prompt stays hidden from Zoom and screen recorders, and more than 40 languages are supported. Tellie requires macOS 14 or later; Pro costs $29 once after a 10-day trial.
Official site: Tellie
Headlines
And that was only the Top 5
Work, skills & management
Reuters found Meta went through with a 10% layoff in May and canceled a second wave planned for November. Internal data showed code changes jumped 220% while user-facing improvements rose only 36%, and employee sentiment fell from 74% to 55%.
A Futurum Group study says factories, utilities and supply chains are using digital workers to monitor routine work and send judgment calls to people. That offers relief for thin teams while changing which tasks remain human jobs.
The integration lets hiring teams evaluate every candidate inside their main recruiting system as AI-written resumes flood the pipeline. Employers still need a defensible way to review how the system scores people.
The Coursera specialization covers meeting facilitation, backlog work, predictive analytics and team sentiment alongside coaching and conflict resolution. It gives project leaders a job-specific route into AI.
The connector can bring current country-by-country hiring, payroll and compliance information into approved work assistants. HR teams can ask a question where they already work, with the supporting rules close at hand.
Business & operations
Businesses can combine seat licenses with usage billing, set project caps and get alerts before agent costs run away. Select workloads will eventually run off-peak for up to half the normal inference cost.
Camera towers now capture SKU and lot data from incoming pallets and send it to Lineage's warehouse system. The installation replaces a receiving process built around manual scans, document checks and data entry.
Two connectors bring validated company identities, relationships and risk data into Perplexity Computer. Finance, procurement, sales and compliance teams can ground research in D-U-N-S-linked records.
Rowan searches past satellite images, orders new captures and runs geospatial analysis from plain-language requests. A user can ask how a construction site changed over a year without learning mapping jargon.
The service gives Mindbody and Booker customers a ready-to-use site with scheduling, pricing and booking already connected. Playlist says it can launch and maintain the site within days at no extra cost.
Health & public service
ARPA-H selected Philips to work with Johns Hopkins and Boston University on remote-assisted and increasingly automated endovascular procedures. The aim is to extend specialist stroke care to U.S. hospitals that cannot keep an expert on site.
The new agents work across healthcare payments using a platform that processes 7.5 billion transactions and touches about 60% of U.S. patients. Hospitals may see faster administrative work, with automated actions requiring clear review and exception rules.
New features show staff who has arrived, who is ready and what each patient still needs. That links paperwork completed at home with the front desk and clinic floor.
Aetna and other health plans use the platform to review behavioral-health care and costs. The funding will expand a system working in a field that affects more than 23% of U.S. adults each year.
RTX-117 is an investigational oral treatment created by a company combining AI with RNA biology. The designation supports development for a rare inherited nerve disorder that often begins in childhood.
Trust, security & consumers
Another 38% would allow only two or three mistakes, and 47% blame the business when AI fails. Customer-facing automation puts the company's reputation on the line every time the system answers.
The platform can surface relevant camera activity and turn scattered clips into one connected case. Security teams get faster review, while organizations still need rules for retention, access and false alarms.
The report found 74% of AI environments already pull outside data through APIs, connectors or plug-ins. Confidence depended heavily on leadership understanding, clear ownership, budget and security staffing.
The service maps which tools employees use and what company information reaches them before policies are written. That gives smaller organizations a practical starting point for controlling shadow AI.
The state House passed bills covering deceptive campaign media and disclosure labels, but they are unlikely to become law before the midterms. Courts have already struck similar rules in California and Hawaii on First Amendment grounds.
Around the world
Gates wants international monitoring for AI systems capable of hacking or designing dangerous biological material. His proposals also include taxes on AI-company revenue for safety nets and preserving some jobs for people.
The Hanover Institute looked independent, but a contractor funded by Israel built it and registered under U.S. foreign-agent law. The operation shows how organizations can flood the web with authoritative-looking material meant for AI systems to absorb.
The 61-member unit plans to recruit 26 private-sector specialists. Its projects include agents that notify people about services they qualify for and help with applications.
AI-generated phishing and impersonation led the list, followed by prompt injection and deepfake or voice-cloning attacks. Only 49% said they monitor access to AI tools and their outputs.
The companies plan to add ALLAM models to Microsoft Foundry and the Microsoft 365 Copilot ecosystem. Joint engineering teams will help organizations identify and deploy projects on Microsoft's stack.
mAIn Street gives nontechnical readers useful AI news they can put to work.
mAIn Street #299 · Thursday, August 27, 2026

Read All mAIn Street Back Issues Here


mAIn Street

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

Read more from mAIn Street
Calculator, charts and planning materials arranged on a desk

AI is moving deeper into medical care, professional work, personal finance and the infrastructure debate. Wednesday, August 26, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment...

Colleagues discussing work around a laptop during a meeting

AI pay packages, school rollout caution and supervised clinical agents lead Tuesday’s AI news. Tuesday, August 25, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment produce work that...

A professional mapping a workflow with sticky notes

Nvidia server price hikes, Kaiser triage concerns and OpenAI’s safety-law reversal lead Monday’s AI news. Monday, August 24, 2026 On September 9, I’m hosting ChatGPT for Marketing: Using AI So It Doesn’t Look Like AI, a 90-minute workshop for marketers and business professionals who are tired of generic copy and look-alike visuals. We’ll work with writing, images, fonts and web pages, then use real examples to show how references, clear instructions, repeated edits and human judgment produce...