This time, the company is updating its Covid-era ‘Work from Anywhere’ policy, which allows employees to work remotely from any location away from their main office for up to four weeks per calendar year. Now Google will count even one WFA work day as a full week, according to an internal document obtained by CNBC on Wednesday.

“Whether you log 1 WFA day or 5 WFA days in a given standard work week, 1 WFA week will be deducted from your WFA weekly balance,” the document, which was circulated over the summer shortly before the update went into effect, said, according to CNBC.

The WFA policy is distinct from Google’s regular hybrid schedule, which grants employees permission to work from home two days per week. The hybrid schedule, which was also established during the pandemic, won’t be altered. WFA days give employees the flexibility to work remotely, but not “from home or nearby,” according to the leaked internal document.

The WFA policy update doesn’t apply to all Google staffers and may exclude those required to be in physical offices and data centers. Violations will result in disciplinary action or termination, according to CNBC.Google did not immediately respond to Fortune’s request for comment.

In April, Google warned remote employees in certain divisions that their roles could be eliminated if they did not comply with hybrid schedules, specifically requiring attendance in the office three days a week. Some teams also offered voluntary buyouts to remote workers living within 50 miles of an office and unwilling to relocate to work in person. 

“As we’ve said before, in-person collaboration is an important part of how we innovate and solve complex problems,” a Google spokesperson told Fortune in April. 

According to the WFA policy update, workers aren’t allowed to work from a Google office in a separate state or country during their WFA time due to ““legal and financial implications of cross border work,” according to CNBC.

Tech remote work

JobLeads, a Germany-based job search platform, CTO Jan Hendrik von Ahlen told Fortune of the more than 12 million U.S.-based job postings in its database, just under 6% of positions are fully remote, under 7% are hybrid, and almost 88% are on-site. “That’s comparable to pre-pandemic numbers,” Hendrik von Ahlen said.

Across the sector, companies increasingly track attendance via badges and other tools and present RTO as necessary for culture and productivity, even as many maintain hybrid rather than fully remote models.

Apple moved to a three‑days‑in‑office hybrid and tied compliance to badge tracking and potential discipline, framing the shift around collaboration benefits despite employee pushback. Meta reinstated a three‑day requirement in September 2023 with enforcement up to performance hits or termination for noncompliance, citing internal data that in‑person work boosts engineering outcomes and keeping exceptions for originally remote hires. 

Amazon escalated from a three‑day mandate to tighter enforcement against “coffee badging,” explored minimum daily in‑office hours, and signaled or set moves toward five days in 2025, sparking internal dissent and phased implementation challenges. Microsoft has shifted toward a formal three‑day on‑site baseline for employees near offices, reflecting a broader trend of codifying hybrid expectations after looser post‑pandemic flexibility. 

The tightening of Google’s policy is in line with the larger trend among big employers that is seeing them slowly dial back pandemic-era flexibility, Kelsey Szamet, partner at Kingsley Szament Employment Lawyers, told Fortune.“Such a policy change can take a tangible toll on morale and retention,” Szamet said, adding that changing expectations of more flexible in-office work policies may frustrate employees that were hired under the assumption of semi-or-fully remote work.

“That can contribute to disengagement or increased turnover, particularly among high performers who recognize that they can obtain flexible options elsewhere,” Szamet said.
Fortune Global Forum returns Oct. 26–27, 2025 in Riyadh. CEOs and global leaders will gather for a dynamic, invitation-only event shaping the future of business. Apply for an invitation.


Source link

Menlo Park-based Felicis led the round, valuing Mercor at $2 billion — eight times its previous valuation, The Wall Street Journal previously reported. Existing investors Benchmark and General Catalyst, as well as DST Global and Menlo Ventures participated.

General Catalyst led the company’s $3.6 million seed round in 2023, while Benchmark backed its $32 million Series A in 2024 at a $250 million valuation.

The round makes CEO Brendan Foody, CTO Adarsh Hiremath, and COO Surya Midha, some of the youngest founders of a billion-dollar startup. The two-year-old platform, which counts Peter Thiel, Jack Dorsey, and Adam D’Angelo as backers, says the latest funding will help “accelerate its ability to match billions of people with their calling, applying human talent to its highest potential.” 

Founded in 2023, Mercor uses AI to streamline hiring. Its platform automates resume screening and candidate matching, and offers AI-powered interviews and payroll management. Employers upload job descriptions and Mercor’s system recommends the best candidates.

Mercor claims its automated system not only streamlines hiring but also removes bias from the process. That claim alleges that AI systems are less biased than humans, which hasn’t always proved to be true. Nevertheless, tech companies such as OpenAI are already using Mercor’s automated tools, which the company claims can find better human candidates than, well, other humans.

Job seekers complete a 20-minute AI interview that evaluates their skills and creates a profile. The platform then matches them with relevant full-time, part-time, or hourly roles.

“We collect performance data on candidates and use it to refine our predictions on who will perform best in the future,” Foody said.

Mercor initially focused on hiring software engineers and tech professionals in operations, content creation, product development, and design. Software engineers are still the most in-demand talent on Mercor today, Foody said. But AI labs are increasingly seeking other professionals — consultants, PhDs, bankers, doctors, and lawyers.

To meet rising demand, Mercor has expanded its talent pool, helping HR teams evaluate 468,000 applicants. India remains its largest talent source, followed by the U.S., while Europe and South America are seeing rapid growth.

Revenue skyrockets as companies embrace flexible work

This momentum has driven a sharp increase in Mercor’s revenue, which it generates by charging hourly finders’ fees to its clients.

Last September, the startup was growing 50% month-over-month, with an annual revenue run rate (calculated by multiplying its latest monthly revenue by 12) in the “tens of millions.” Maintaining that pace, it now stands at a $75 million ARR, most of which comes from AI labs. Mercor says it now works with the world’s top five AI labs, including OpenAI.

Mercor’s $2 billion valuation gives it a 27x ARR multiple, a reasonable figure compared to the more inflated valuations seen today. Some investors are willing to pay up to 50 times ARR for the fastest-growing generative AI companies.

Aside from concerns about hiring bias, another debate surrounding Mercor’s technology is its potential to accelerate job displacement as AI advances.

Foody, however, argues that rather than displacing workers, Mercor is automating large parts of the economy, making workers even more valuable in the areas where they are still needed.

According to the chief executive, Mercor helps identify jobs humans should be doing in an AI-driven economy or jobs AI can’t perform — such as training AI models, managing complex decisions, or filling creative and strategic roles.

“If AI automates 90% of the economy, then humans become the bottleneck for the remaining 10%. So there’s 10x leverage on every unit of economic output that humans contribute because the rest has been automated,” Foody explains. “That means the way people work is changing as we move toward a more fractional, gig-like work model.”

That’s why the founder believes Mercor will remain relevant in the long run, as more companies prioritize expertise over tenure and hire specialists for short-term projects instead of relying on full-time staff.

“I think work becomes more efficient through smarter job matching,” he said. “Every project should be handled by the best person for the job, not just whoever is available on staff.”

As for its own hiring, Mercor, with an average team age of 22, recently hired the former head of Human Data Operations at OpenAI and the previous head of Growth at Scale.

Today, former OpenAI CTO Mira Murati announced her new venture: Thinking Machine Labs, a public benefit corporation that aims to build accessible and broadly capable artificial intelligence systems. After leaving AI juggernaut OpenAI last September, Murati has brought together a team of engineers and researchers who have worked at buzzy startups like Character AI, Mistral and unsurprisingly, OpenAI. Additionally, Thinking Machine Labs said it will publish its technical blog posts, code and papers and collaborate with the broader community, indicating it plans to open source its work.

Now let’s get into the headlines.

BIG PLAYS
The cluttered world of AI reasoning models just got its newest addition. On Monday Elon Musk’s AI company, xAI launched a new AI model called Grok-3 that can process and answer complex questions across domains like science and math. In a livestream on X, Musk said the company’s mission is to “understand the universe…to figure out what’s going on, where are the aliens and what’s the meaning of life.”

The billionaire claims the model is built using 10 times more compute than Grok 2, likely from its “gigafactory of compute” in Memphis, and that it has been trained on public data from sources including social media platform X and legal documents. xAI also rolled out an AI-powered search engine called DeepSearch. OpenAI cofounder and former Tesla executive Andrej Karpathy, who tested the model, says Grok-3’s capabilities are largely on par with OpenAI’s best models but gets some questions wrong. Others have noted the model is lacking in its coding abilities compared to others. Grok-3 has yet not been independently evaluated and is only available to paying users.

ETHICS + LAW
Condé Nast, Vox, The Atlantic and a group of publishers have sued $5.5 billion-valued AI company Cohere for copyright and trademark violations. (Forbes is part of the group suing Cohere.) The lawsuit alleges that the Canadian AI startup scraped 4,000 copyrighted articles from the internet and used them to train its family of large language models called Command, which reproduced sections or entire works (at times word for word), allowing users to get information without visiting the publishers’ websites. It’s not the first time an AI company has faced publishers’ scrutiny. Last year, AI search engine Perplexity came under fire for republishing copyright works from multiple publications including Forbes. (In response, Forbes sent a cease and desist letter to Perplexity, accusing it of copyright infringement.)
AI DEALS OF THE WEEK
Humanoid robotics company Figure AI is in talks to raise $1.5 billion in venture capital at an eye-popping $39.5 billion valuation, Bloomberg reported. The news comes as the company is reportedly in talks with Meta to make robots for household chores.
AI legal company Luminance, which helps customers like AMD and National Grid generate, negotiate and analyze contracts, has raised $75 million in series C funding.
Chip startup Encharge AI has raised $100 million in a series B funding led by Tiger Global. CEO Naveen Verma started the company out of a lab in Princeton, where he worked on designing architecture for hardware that would help run large language models more compute and energy efficiently. Verma says the chips allow AI models to run locally on devices such as personal computers.
DEEP DIVE
Elon Musk’s surprise bid for the nonprofit controlling artificial intelligence behemoth OpenAI did exactly what he wanted it to. Announced as OpenAI CEO Sam Altman and other business and world leaders convened in Paris for a global AI summit, the unsolicited $97.4 billion offer for the nonprofit refocused the world’s attention on Musk and his efforts to block OpenAI’s transition to a for-profit company.
An irked Altman quickly dismissed Musk’s offer and sources close to OpenAI say it’s hard to imagine it going anywhere. But even if that’s the case, Musk has likely caused a headache for Altman, who is orchestrating the company’s transition to a for-profit venture. He’s attempted to forcefully raise the nonprofit price – which would make it harder for OpenAI to justify paying anything less.
Musk’s bid is the first hard number that values the nonprofit that controls OpenAI; that entity has to be bought out and become a minority shareholder for OpenAI to successfully transition to a for-profit company. Previously, The Information had reported the nonprofit was worth around $40 billion, citing a 25% stake and the company’s valuation at time. But with his $97.4 billion bid, Musk has backed Altman into a corner; now, as a board member, Altman faces pressure to sell the nonprofit for at least what Musk is asking. If he were to sell for anything less, it’d be a bad look, making it seem like he’s lowballing his own company to reduce share dilution.
“The important part here is that if [the board] doesn’t take it, which they almost certainly won’t, then they’ve made clear that they think the assets Musk is trying to buy are worth more than $97 billion,” a person familiar with the company told Forbes. “So if the for-profit tries to buy them later, the nonprofit will have to get more than that — otherwise the board is likely in breach of their fiduciary duties.”
Read the full story on Forbes.
MODEL BEHAVIOR
Generative AI is making it easier for fraudsters to carry out romance scams at scale, Wired reported. AI chatbots are being used to generate hundreds of deceptive scripts and generate fully fake profiles on dating apps. AI has already made a foray into the dating world. Last year, we wrote about a man who programmed ChatGPT to reply to his matches on Tinder and set up dates for him.

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More

Speaking for myself, interacting with any merchant’s AI-powered chatbot on their website is often an exercise in frustration. Phone trees with robot voices are typically worse. I’d wager I’m hardly alone in my assessment. Who amongst us hasn’t experienced long hold times, slow responses, lack of updated information and awareness of the customer’s own account history, a granting faux politeness and a host of other inefficiencies?

A new startup called Palona debuted last week that aims to fix this sorry state of affairs. It equips direct-to-consumer enterprises — think pizza shops and electronics vendors — with live, 24/7 customer support sales agents that are uniquely reflective of each business’s brand personality, voice, inventory stock, and value proposition. The electronics vendor has a “wizard” agent made by Palona, while the pizza shop gets a surfer dude agent personality — naturally.

In all cases, Palona focuses on creating AI agents that have high “EQ,” or “emotional intelligence / emotional quotient,” building them from a combination of open source and proprietary AI models and training some of their own using human sociology research.

“A kind of fundamental thesis that the company is that we can create something an experience that is delightful and feels genuine, like a real human conversation,” said Palona co-founder and chief technology officer (CTO) Tim Howes, in an in-person interview with VentureBeat recently. “ChatGPT is a hugely useful tool, but it does not feel like a human conversation.”

Palona claims its system can be easily implemented by a non-techie brand on their website, mobile app or phone lines — with responses uniquely tailored to each brand and each communications environment. And in fact, its agents are already at work handling orders, answering questions and complaints, and suggesting products and upsells to customers.

Strong founding background

In addition to Howes, Palona is co-founded and led by a team of engineers from some of the top tech companies in the world, among them: Maria Zhang, Palona’s CEO, is a former VP of Engineering at Google, VP/GM of AI for Products at Meta, and CTO of Tinder. She also founded Alike, acquired by Yahoo in 2013.

Palona’s chief scientist Steve Liu, PhD, was formerly chief scientist at Samsung AI Center and Tinder. A tenured professor at McGill University, Liu is also a Fellow of IEEE and the Canadian Academy of Engineering, with over 390 research papers to his name.

And Howes himself is the co-inventor of the industry-standard, open source Lightweight Directory Access Protocol (LDAP) online data storage system, as well as co-founder of LoudCloud and OpsWare (the latter acquired by HP for $1.65 billion). He was also previously the chief technology officer at Netscape, HP Software, and led Developer Productivity at Meta AI Infrastructure.

“We’re building fully autonomous sales agents—not tools for salespeople, but actual AI salespeople,” said Zhang in the same live interview, adding that it was “the employee of the century.”

24/7 polite, distinct, personable sales agents

Palona AI positions itself as a solution for companies looking to improve their sales performance, customer engagement, and brand loyalty.

The Palona Agents act as customized virtual sales employees, combining soft sales skills with 24/7 availability, unlimited capacity, and advanced memory recall, and can interact with customers in an online chatbot format, an SMS/text number they can correspond with, or even AI-powered voices over the phone.

“100%—we support voice,” Zhang explained, “For example, in pizza ordering, voice is still a major user pattern. In the Midwest, about 50% of people still call to order. On the East and West Coasts, it’s around 20%, but it’s still significant.”

Palona’s voices are licensed, but the company has the ability to train and deploy custom ones, even voice cloning of authorized customer reps or a CEO, let’s say.

The company realized through testing that the voice version of Palona’s AI sales agents would need to have distinctly different interaction styles from the text chatbot.

“We tested different voice interactions, and for pizza ordering, for example, customers wanted efficiency,” Zhang related. “They didn’t want a chatty AI—they just wanted to get their order done as fast as possible. So we optimized for that, making it have less personality, less verbosity, more efficiency.”

Unlike traditional chatbots that serve as assistants to human representatives, Palona AI is designed to handle entire sales cycles without human intervention.

“There’s a big gap between lifelike AI models like ChatGPT and what businesses actually need—an AI agent that can fully sell, convert, and upsell,” Zhang explained.

Palona claims to minimize errors and reduces AI hallucinations by up to 98%, ensuring reliable interactions.

Zhang and Howes said that for even the most analog businesses, it was a short lead team to get going with Palona, only several days for a simple implementation.

The enterprise provides Palona with “FAQs, employee training manuals, policies, and procedures,” said Howes.

Then, they define with Palona what actions the agent should take — be it processing orders, answering inquiries, or handling support issues.

“The biggest factors affecting setup time are: how much integration is needed with the customer’s existing systems (e.g., POS, CRM, ordering platforms). If we already support their system, it’s plug-and-play,” Howes explained. “If it’s a system we already support, the agent can be ready in a couple of days. If they’re using a new, unfamiliar system, that requires additional engineering work, which could take longer.”

In addition, Zhang said that Palona was “actually in the process of automating agent setup. Eventually, businesses will be able to use a Palona agent to configure their own Palona agent!”

Three language models are better than one

Palona does all this by combining three different models: the first is a custom, fine-tuned large language model (LLM) that serves as the basis for every distinct business sales agent — the pizza shop gets a different tone and personality from the electronics vendor, and each one is customized to each customer out of the box.

There’s also a supervisory model that detects, catches, and removes hallucinations from the main model before it outputs them to the customer.

Finally, the system also incorporates a real-time memory tracking small language model (SLM), allowing it to build deep customer profiles based on previous interactions—leading to personalized conversations and stronger customer relationships.

“AI is very good at either remembering absolutely everything, which leads to a terrible experience,” Howes said, due to it bringing up unimportant details. “Or it remembers nothing, which, again, terrible experience,” because it won’t know what happened previously in the conversation.

To get around this, Palona trained its own small model to help its base LLM manage its memory.

It’s that small model’s “job to figure out, ‘okay, what’s important here, what’s not important,’” in every conversation, Howes said.

Overall, the company included an interesting dataset for all its models that reflects the wide range of human emotional responses all sales and customer-facing humans (and agents) have to deal with on a regular basis.

“We trained our models to have higher EQ [emotional intelligence] by surveying the literature in psychology, and identifying what actually makes someone emotionally intelligent,” Howes told me. “It turns out there’s an eight-dimensional definition of EQ, and we trained our AI against those benchmarks.”

Palona also trained its models to perform “gentle persuasion,” and reply to customers with humor, emoji, and sensitivity.

Early customer testimonials: Wyze, MINDZERO, and Pizza My Heart

Palona AI is already working with several consumer brands, including Wyze, MINDZERO, and Pizza My Heart, to enhance customer interactions and increase sales conversions.

Wyze, a smart home camera company, has integrated Palona-powered AI sales assistants to personalize customer support.

Yun Zhang, CEO of Wyze, said in a statement: “I’ve always wished that I could personally connect with every customer to share why they should choose our camera and CamPlus plan. Palona has made that possible with the Wyze Wizard AI Agent, delivering the curated, personalized buying journey I wish I could give every customer myself.”

MINDZERO, a wellness studio specializing in contrast therapy, uses Palona AI Agents to answer customer questions naturally and engagingly. David Semerad, CEO of MINDZERO, highlighted the impact of the technology: “Our Palona agent, Jen, has become our new best friend—she’s able to answer customer questions in such an authentic, kind, and helpful way, it really blows our minds.”

Pizza My Heart, a well-known West Coast pizza chain, has transformed its brand persona into an AI-powered interactive experience. The company’s Palona AI Agent, “Jimmy the Surfer”, allows customers to order pizza through voice or text while engaging in lighthearted conversations.

Chuck Hammers, CEO of Pizza My Heart, commented on the experience: “Palona AI brought Jimmy to life in a way that just makes me smile—which I never expected from an AI. Now, Jimmy helps my customers order pizza by talking to them directly—via voice or text.”

Three pillars of Palona’s success

Palona’s AI solutions are built on three core principles:

1. Brand Consistency – AI Agents are trained on a company’s knowledge base to ensure conversations remain true to the brand’s voice and identity.

2. High EQ (Emotional Intelligence) – Palona Agents dynamically adapt to conversations and outperform both human representatives and competing AI models in emotional engagement.

3. Persuasiveness – The AI identifies key sales opportunities, proactively recommending upsells and cross-sells at the right moments to increase revenue and customer lifetime value (LTV).

I saw all these in action in a short demo of a “Wizard” customer sales chatbot agent, in which even when I complained to it about my dogs barking at my neighbors (sorry guys!), it managed to smoothly redirect the conversation to a home camera I could buy to help keep an eye on them.

What’s next for Palona?

“We intentionally went after the hardest problem first—sales and conversion,” Zhang noted. “It requires more technical sophistication, but if you can solve that, expanding into other areas like customer support and loyalty programs is easier.”

As AI continues to shape the future of commerce, Palona’s approach suggests that customer relationships don’t have to be sacrificed for automation—instead, AI can enhance personalization, drive conversions, and build lasting brand loyalty.

Daily insights on business use cases with VB Daily
If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI, from regulatory shifts to practical deployments, so you can share insights for maximum ROI.

Read our Privacy Policy

Thanks for subscribing. Check out more VB newsletters here.

An error occured.