# The AI agents are coming for Microsoft Office > Source: > Published: 2026-07-10 16:39:15+00:00 # The AI agents are coming for Microsoft Office ## Today, we're diving into the state of the memory ecosystem, a few key executive departures, AI costs, and why everyone wants to eat Microsoft's productivity revenue. **Friday. **What a week, yeah? Worry not, stocks are still rising, the Iran War is somewhere off its kinetic peak, and we’re drowning in new AI models. Today, we’re diving into the state of the memory ecosystem, a few key executive departures, AI costs, and why everyone wants to eat Microsoft’s productivity revenue. To work! **— Alex** ## 📈 Trending Up [Thinking Machines Lab](https://x.com/yilin_yang721/status/2075345236587610490)? …[SK Hynix, which begins trading on the Nasdaq today](https://www.cnbc.com/2026/07/09/meet-sk-hynix-the-trillion-dollar-chipmaker-debuting-on-us-markets-.html)…[reusable rocket boosters from China](https://x.com/XH_Lee23/status/2075446068134007281)…[competition in stablecoins](https://www.bloomberg.com/news/articles/2026-07-10/stablecoin-firm-circle-gets-approval-for-coveted-us-bank-charter)…[the EU](https://x.com/magyarpeterMP/status/2075540230002622624)…[inflation](https://www.cnbc.com/2026/07/10/delta-air-lines-dal-q2-2026-earnings.html)…[GOP anti-datacenter agitation](https://financialpost.com/pmn/business-pmn/gop-candidates-break-with-trump-on-data-centers-to-boost-midterm-odds) **Memory chip makers:** SK Hynix raised $26.5 billion in a US-listing that priced its shares at $149 apiece, which [RenCap describes](https://www.renaissancecapital.com/IPO-Center/News/120386/SK-hynix-prices-US-IPO-at-$149-a-3-percent-premium-to-Korean-close) as a “3% premium to [the company’s] as-converted close on the KRX KOSPI.” This will bear a decent amount of capital for the company, which is building a $4 billion facility in the United States that will “establish HBM production lines and advanced packaging R&D facilities with operations scheduled to commence by the end of 2028.” - Chip giant TSMC and memory titans Samsung, Micron, and SK Hynix are trillion-dollar companies today. AI demand better not slip, or a host of companies are going to get cut in*all**half*. **Change of guards: **CEO of Apps at OpenAI, Fidji Simo, is [stepping down](https://x.com/fidjissimo/status/2075353170927304861) due to a medical condition. I only know Fidji a tiny bit from her days as Instacart’s CEO, but I hope she recovers quickly. Elsewhere, Coinbase’s chief legal officer, Paul Grewal, is [departing his role](https://x.com/iampaulgrewal/status/2075312992514924642?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5Etweet) after more than five years at the company (he’ll stick around as an advisor). That’s not a bad run at all, considering the state of the U.S.’ crypto regulation before he joined the company and where we are today. **Cutting your AI bill:** One of the best things about AI development is how quick it is. AI models got really good at coding over the past year, kicking off a war to build better models and better harnesses. We saw a step-change in models’ capabilities, particularly Anthropic’s Mythos, which led to a slew of developments across the sector. Then, after many AI products moved from flat-rate to usage-based pricing, a cost meltdown ensued. As a result, we heard the following this week: **Mark Zuckerberg**“Today we’re releasing Muse Spark 1.1 — a strong agentic and coding model at a very low price […] Our focus is on delivering strong agentic and multimodal models at very low cost.[on](https://x.com/finkd/status/2075218444056707458)Meta’s new Muse Spark 1.1 model:**SpaceX’s Elon Musk**“Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster. The combination of capability, faster speed and lower cost is what makes it competitive.”[on](https://x.com/elonmusk/status/2074911038286295049)his new Grok 4.5 model:**OpenAI’s Sam Altman**“[W]e have heard enterprises on their concerns about AI costs, and 5.6 sol is a huge step forward for dollars-per-task, as are [its smaller versions] terra and luna.”[on](https://x.com/sama/status/2075267201058426944)his new GPT-5.6 model family: Cost, cost, cost. The narrative has quickly shifted from a pure focus on capability to building smarter *and *more efficient models. Well, everyone but Anthropic. [📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html) Trending Down [📉](https://finance.yahoo.com/news/servicenow-pledges-1-5bn-investment-110000403.html) [American democracy](https://www.notus.org/trump-white-house/election-assistance-commission-firings)…[task force comity](https://x.com/boes_/status/2075297179783057484)…[American household financial health](https://www.bloomberg.com/news/articles/2026-07-09/polymarket-seeks-license-to-offer-margin-trading-legally-in-us)…[free speech](https://reason.com/2026/07/06/77-year-old-florida-veteran-investigated-for-sending-an-official-a-postcard-saying-you-lack-values/)…[regulation-free tech](https://www.reuters.com/business/retail-consumer/uk-regulate-cloud-service-providers-microsoft-google-others-protect-financial-2026-07-10/)…[going public](https://www.cnbc.com/2026/07/09/anduril-ceo-ipo-defense.html)…[agentic browsers](https://techcrunch.com/2026/07/09/openai-is-shutting-down-atlas-but-its-ai-browser-ambitions-are-still-growing/) **Google’s AI chops, right before earnings:** What’s wrong with this chart? The [Artificial Analysis intelligence leaderboard](https://artificialanalysis.ai/#intelligence) (arguably one of the most important AI model rankings today) indicates that Anthropic, OpenAI, SpaceXAI, and Z.ai are ahead of Google. Cast your mind back to late 2025, and Google was *on top* of the entire industry. Today, the *smallest* model from OpenAI’s GPT-5.6 release is better than *any* Google AI model. Woof. Perhaps Mountain View is waiting to release new models right before it reports Q2 results (Gemini 3.4 Flash was released in mid-May) on [July 22](https://abc.xyz/investor/news/news-details/2026/Alphabet-Announces-Date-of-Second-Quarter-2026-Financial-Results-Conference-Call-2026-2h_R0kzZHY/default.aspx). I simply didn’t expect to see *Google* fall so far behind its American rivals, and even a Chinese lab, in 2026. ## Everyone is coming for Microsoft’s enterprise perch I don’t think Microsoft gets enough credit for successfully transitioning Office from a one-time purchase you could use forever to a subscription service. Office 365 launched in [June 2011](https://news.microsoft.com/source/2011/06/28/microsoft-launches-office-365-globally/), was released for consumers in [January 2013](https://news.microsoft.com/source/2013/01/29/microsoft-releases-office-365-home-premium-2/), received [business upgrades shortly afterward](https://news.microsoft.com/source/2013/02/27/microsoft-releases-next-generation-office-365-for-business-2/), and by July 2013, had [reached ](https://news.microsoft.com/source/2013/07/18/microsoft-reports-fourth-quarter-and-full-year-results/)annual run-rate revenue of $1.5 billion. Looking back, in its fiscal 2011, when Office 365 had been available for just a few weeks, the encompassing “Microsoft Business Division” [reported revenue of $22.2 billion](https://view.officeapps.live.com/op/view.aspx?src=https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/SlidesFY11Q4). In the company’s [most recent quarter](https://view.officeapps.live.com/op/view.aspx?src=https://cdn-dynmedia-1.microsoft.com/is/content/microsoftcorp/TranscriptQandAFY26Q3), Microsoft 365 had 95 million subscribers, with “M365 commercial cloud” revenue up 19% from a year earlier. Meanwhile, the “Productivity and Business Process” unit (the old Microsoft Business Division that has other products like LinkedIn) reported revenue of $35 in the *quarter*. Microsoft has proved it can survive major changes in the tides of technology — after all, it famously missed the mobile wave and it’s still going strong. Today, it faces another evolution in one of its core cash cows, as late-stage unicorns and AI labs alike push deeper into Office territory. AI labs are building general tools that compete with the work that people have traditionally used Office for. Meanwhile, software makers that used to focus tightly on one area are broadening the scope of their AI tools. What do we mean? **Anthropic’s Cowork has greatly expanded its scope** with a blizzard of plugins that tackle specific tasks, and [recently got a Web and mobile app, too](https://claude.com/blog/cowork-web-mobile). Anthropic claims about 90% of the time, Cowork wasn’t used for software development. Instead “most of it was everyday knowledge work, and the largest categories were business operations and content creation.” Smells like a generalized agentic AI platform aimed at everyday tasks! **OpenAI released ChatGPT Work this week**, [combining ](https://openai.com/index/chatgpt-for-your-most-ambitious-work/)its chatbot with its coding and automation tool, Codex. The tool can “gather information across your apps and workflows to create finished materials like sheets, slides, docs, and web apps, and stay with complex projects for hours,” and OpenAI mentioned sales, marketing, finance, bizops, data, and engineering use cases. Again, isn’t that just an AI version of Office? **Cursor is reportedly developing a ** that would greatly expand its total addressable market by shifting its historical focus away from the coding market. *generalized*AI agent The company’s upcoming corporate home, SpaceX, has already stated its goal of fully recreating Microsoft in AI form ([Macrohard](https://www.reuters.com/business/autos-transportation/musk-unveils-joint-tesla-xai-project-macrohard-eyes-software-disruption-2026-03-11/)), which makes Cursor’s general-purpose AI agent an obvious product choice. **Databricks unveiled Genie One last month**, a “data-smart AI coworker designed to help users move from insight to action” by connecting business data and business apps. If you’re thinking, “That’s not really competing with *Office*,” chew on this: “Genie One provides a full suite of agentic-cowork capabilities, including schedules and alerts, monitoring, document creation, custom skills, and custom MCP support.” **Legal AI unicorn Harvey is putting together** a [team to train AI models](https://x.com/gabepereyra/status/2074895016619847806) as it seeks to “expand from the application layer into the model layer and from legal into high end knowledge work more broadly.” That’s clear enough! AI labs and other modern software companies want a piece of the immense market Microsoft Office owns. Not only is there a lot of revenue at stake, but the data generated by the thousands of enterprises using Microsoft 365 is more valuable than gold if you want to improve your AI tools for everyday work. We also have companies pushing further into *general* work. Like its rivals, payments company Gusto has expanded its remit over time to include benefits management, cash flow tools, and the like. But the recent [release of its new AI tool, Gusto Cofounder](https://gusto.com/company-news/cofounder), expands that footprint further. Here’s a fraction of the official verbiage: “Gusto Cofounder [is] a new AI teammate purpose-built for small business. [It can] automate repetitive tasks across your connected tools, catch potential issues, and flag what’s coming up. [The product can] monitor your business proactively, like surfacing a scheduling conflict before it becomes a staffing crisis, flagging unusual payroll activity [and thanks to underlying Gusto data] Cofounder already knows your team, your payroll cadence, your benefits, and your compliance calendar.” Gusto is great at staying on-message. Still, it’s building a tool that aims to help its customers do much more than just run payroll. I doubt that Gusto will restrict Cofounder to the SMB market, no matter how it’s branded today. I spoke with Gusto CTO Edward Kim the other week about Cofounder. He’s an OpenClaw fan, but like * CO*, he thinks the open-source tool has a cold-start problem, being complex enough that most users fail to get significant value from it. The solution to those problems is to build an agent that is grounded in customer data. And if you manage it for users, the chances of the agent constantly breaking itself (looking at you, OpenClaw) fall to zero(ish). **An interesting detail:** When I spoke with Kim, Cofounder had only been in the market for a few days, but the executive said that early traction was looking good. Pressed on the unavoidable AI cost question, Kim said that his team is not worried about cost just yet. If many customers begin to use Cofounder heavily and its inference costs become material, that would be a good problem to have, the CTO said. Perhaps it’s easier to envision Box (general business data repository with automation) building a general AI agent that could compete with Microsoft, but I expect vertical AI tools to broaden their scope in time. - People use Office for all kinds of work: accounts and finance folks can’t work without Excel; Word is a mainstay for legal work; MBAs love PowerPoint… You get the picture. - AI companies don’t need to map agents 1:1 to specific Office tools; they merely need to compete on the quantum of work in question. Not every software company can build that, however. Gusto has a bit of an advantage, here: Ages ago, when I [wrote about the company opening up its services via API](https://techcrunch.com/2021/06/30/2171608/), its co-founder and CEO Josh Reeves said Gusto was built using “internal APIs to connect abstracted front and backends,” which enables it to expose individual products for external consumption. That very API-first setup has now helped Gusto build Cofounder. But the real reason why I think Gusto (or Box) can create generalized AI work agents is not merely their ability to build software, nor their data stores. Instead, they have customer *trust* that they can lean against to get companies off the benches and into the AI game. And one great way to do so is to use their existing products and information to create something that speeds up core workflows and extends capabilities without needing new software. Summing up, Microsoft transformed Office from an on-prem software suite to a profitable SaaS product, and now wants to get it over the agentic hump. Copilot Cowork appears to have steam, and Microsoft is building its own foundation models to handle in-house work inference. As a result, it may send less revenue and data to AI labs like Anthropic and OpenAI. Enterprise productivity is one of the biggest prizes in tech. Just as no AI lab wants to cede the application layer to incumbents, no startup building a vertical agent will want to skip the opportunity to make something more broadly saleable down the road. Databricks, Gusto and Harvey are, to varying degrees, examples of this in action today. The winners in this race will be decided by boldness and speed, so I expect the next few quarters to be goddamn electric.