PRODUCT

Microsoft tests the good-enough AI thesis

Microsoft Excel and Outlook can be among the most frustrating apps to use for professionals. Many have turned to AI for help, but the models powering that help may be quietly changing.

On Tuesday, Microsoft announced it is replacing OpenAI and Anthropic models with its own model in products like Excel and Outlook, Bloomberg reported, citing anonymous sources. Tens of thousands of AI prompts in those apps are now handled each week by internally built MAI models, according to the report.

The motivation behind this is simple: cutting costs. Using in-house models is much less expensive than licensing them from third parties. Just last month, Microsoft AI Chief Mustafa Suleyman stated Microsoft's intention to reduce spending by slowly weaning off Anthropic models in favor of MAI models, which span reasoning, image, voice, and coding.

While the move to proprietary models will cut costs, it carries some risk. Working professionals and enterprises are supercharging the race to deploy AI, and they want the fastest and most effective models, such as the ones Anthropic has become known for in the enterprise.

Anthropic just released its report on how people are using Claude Cowork, highlighting how much knowledge workers are leaning on AI. 

The report found that, from a sample of 1.2 million anonymized and aggregated Claude Cowork sessions, a third of Claude's Cowork sessions involve business operations (33.4%), with the next top use cases including content creation and copywriting (16.4%), software development (8.7%), and DevOps and infrastructure (7.0%). 

Microsoft's move to its own proprietary models isn't the first cost-cutting move this week. It eliminated 4,800 jobs (2.1% of its workforce) on Monday. While there is no denying that most companies are feeling the squeeze of AI due to expenses related to building, running, and serving AI services, Amy Coleman, EVP and Chief People Officer, made it clear in communications with employees that "the roles eliminated today are not being replaced by AI." However, the company is shifting resources to expand its AI division. 

If the MAI models prove highly efficient for users and less costly for Microsoft to run, they could help the company pull audiences away from Anthropic and other enterprise competitors. Microsoft already has a built-in edge here: decades of Office 365 dominance mean it has users locked into its walled garden, and those users trust Microsoft in a way that's hard to replicate. Even if Microsoft's models are only 80% to 90% as good as the latest frontier models from Anthropic, OpenAI, and Google, many enterprises may be willing to make the trade-off if the models are less expensive and viewed as being safer. 

TOGETHER WITH GENERAL ASSEMBLY, AN LHH BRAND

There Are AI Leaders, And There Are AI Followers

Which one do you want to be? If your answer is leader (and it probably should be) then you need to check out the latest from General Assembly. Their “AI For Leaders” program has everything current – and future – executives and decision makers need to build for the future. You’ll learn how to drive real AI transformation across your organization, apply AI skills in real business context, and shift your company culture to one of positive AI adoption.

A new era of leadership requires a new approach. See how General Assembly can set you on the path to success right here.

GOVERNANCE

AI adoption is outrunning enterprise security

Companies are deploying AI faster than they can secure it.

Nearly eight in ten enterprises reported encountering AI-related security issues over the past six months, according to new research from security firm DigiCert, which surveyed 1,001 IT and cybersecurity leaders across the US, UK, and Australia.

Approximately 50% of respondents experienced at least one security incident caused by an unauthorized or misconfigured AI agent, while another 28% identified AI-related vulnerabilities without experiencing an incident, a spokesperson from DigiCert told The Deep View. Science and technology organizations reported the highest rate of AI-related incidents or vulnerabilities, followed by banking, financial services and insurance, telecommunications and media, and retail.

While companies weren't asked to identify specific incidents they experienced, Brian Trzupek, DigiCert’s senior vice president of product, said that unauthorized or misconfigured AI agents can pose a range of security risks, including prompt injection attacks, data poisoning, unauthorized access to sensitive systems, and a lack of traceability.

“AI agents are non-human actors operating at machine speed, and most enterprises haven't extended the identity, authentication, and audit controls they already require of every user, device, and application to cover them,” Trzupek told The Deep View.

The risks are outpacing companies' ability to manage them. While 90% of organizations have discussed AI governance, only half have dedicated budgets and formal governance programs in place, according to the study. Nearly half also lack full visibility into how AI systems arrive at their outputs, making it difficult to trace which model produced a decision or investigate what went wrong.

The potential for unintended consequences is becoming more urgent as companies accelerate AI adoption. Over the past six months, 75% of organizations deployed four or more AI-powered systems, while 35% deployed more than 10. As AI becomes embedded across business operations, every new model, agent, and integration introduces new security risks. Without stronger governance, companies could face legal, regulatory, operational, and reputational repercussions.

The report argues that enterprises need stronger identity and governance controls, including the ability to verify AI identities, audit AI decisions, and trace outputs back to the underlying models and data. Companies appear to be moving in that direction, with 86% reporting that they have formal or informal processes for revoking access to compromised AI systems.

“Without governance, organizations lose the ability to answer basic questions: what AI is running, what it can access, and who is accountable when it fails,” Trzupek said.

Enterprises adopting AI have stepped into uncharted territory. Before AI, companies were already dealing with a slew of vulnerabilities and cyberattacks. Now, AI has only added to the challenges as the threat landscape evolves. Companies must not only prevent their AI systems from leaking proprietary data, but also protect them from exploitation by bad actors. And that only scratches the surface. As AI advances, threat actors are becoming more sophisticated, potentially opening the door to new attacks that haven't yet been discovered. If companies are serious about baking AI into their workflows, security and governance will have to become a higher priority.

TOGETHER WITH MERCURY

Banking* that does financial work for you

You've seen what AI can do when it's built into the right system. Mercury Command is that for your finances.

Built directly into your Mercury account, Command turns plain language into executed financial work — categorize transactions, adjust a card limit, send an invoice — with full account context, explicit approvals, and a traceable record behind every action. 

Try Mercury Command →

HARDWARE

Why ex-Tesla scientist bet on robotics in Europe

Tesla staked its claim on humanoid robots with Optimus before AI breathed new life into the market. Now, one of the scientists who helped build it is working on a rival.

On Tuesday, UMA, the physical AI startup founded in 2025, unveiled the design of its first humanoid robot. Remi Cadène, CEO and co-founder of UMA and former staff scientist at Tesla, sees humanoid robots as the key to unlocking greater productive capacity.

 "We believe intelligent robots will become part of the solution, not as a substitute for people, but as a new class of tools that enables them to devote more time to what machines will never replace: creativity, judgment, innovation, and caring for others," said Cadène in the release.

UMA is developing two systems: a dual-arm mobile industrial robot that could operate in warehouses or on assembly lines, and a compact humanoid robot designed to navigate human-centric spaces and collaborate directly with people. 

While many companies are currently tackling physical AI and have built humanoid robots, UMA's unique appeal is its use of the Real-Time Learning architecture, in which robots can learn through demonstration rather than manual programming. 

According to the company, this is meant not only to help them learn more efficiently, directly in the physical world as a human would, but also to improve their performance over time. Another distinctive characteristic is that most of the robotics buildout is happening either in the US or China, while UMA is building physical AI in Europe.

The idea of physical AI is simple: give AI access to your actual environment so it can take action and learn tangibly and firsthand. Though humanoid robots may have first been made mainstream by Tesla with Optimus as a futuristic idea, they now have buy-in from big companies and leaders, as AI has broadened humanoid robots' applications and ability to learn. UMA is a prime example, with a team comprising experts with experience at NYU, Google DeepMind, and LeRobot, as well as backing from leading AI figures and investors, including Yann LeCun and Olivier Pommel.

Europe has been falling behind in AI and robotics development compared to the US and China. However, one unique advantage the region has is actual legislation in place to standardize AI operations, making companies more likely to trust its use. For robotics specifically, there's a lot of public mistrust, since it not only looks like something out of a sci-fi movie but also involves cameras and cohabitating with humans in a way that can inherently compromise privacy. As a result, a regulated robotics industry in Europe could be exactly what it takes to bring the technology mainstream, or at least give Europe a much-needed early advantage.

LINKS

• Chinese AI labs could restrict access to their most powerful models 

• DeepSeek is reportedly developing its own AI chip 

• AI law startup Norm raised $120 million in Series C funding round

• AI actor Tilly Norwood is set to make big-screen debut in feature film

• Research shows people trust AI faces more than those of real people

• Samsung stock soars fueled by AI chip business 

• AI labs hand out free tokens to captivate startup market

• Claude Fable 5: Anthropic extends access to Fable 5 through Jul 12 on paid plans

• Google Managed Agents: Get support for remote MCP and background tasks

• Muse Image and Muse Video: first media generation models by Meta Superintelligence Labs

• SIMBA 3.2: Speechify's newest model topped the Artificial Analysis Speech leaderboard 

• Grok: The Grok API has 21 new flagship voices. 

• OpenAI GPT-Realtime-2.1-mini: Now available in the API 

• Google Finance: The recently released Android app has AI research tool and features 

• Coreweave: Staff AI Security Engineer

• Paramount: Sr. Director, Technical Program Management AI/ML

• Transformworks: AI deployment engineer

• Anthropic: Account Executive, AI Native

GAMES

POLL RESULTS

Do you think social media platforms should use AI to combat bots?

Yes (75.37%)
No (14.92%)
Other (9.71%)