Summary

Microsoft offered one of the clearest internal samples yet of whether AI is translating into revenue: after 65,000 MCAPS employees deployed Microsoft 365 Copilot, per-capita opportunity creation rose 10%, deal-closing speed improved 23%, and revenue per salesperson increased 9%. 段威之 stressed that saving work hours is not the end goal: “Ultimately, substantive business outcomes are what matter.” Copilot for Sales consolidates CRM, email, meetings, and internal product materials, but “it doesn’t make decisions; it gives you very good suggestions.”
Advertising gains have also landed in measurable metrics: after Samsung adopted Microsoft’s AI Ads solution, click-through rates on Microsoft inventory rose 2.8x, while time spent on account analysis and reporting fell 35%. Performance Max hands Bing Search, MSN display, and gradually connected inventory from Outlook, King, and Activision Blizzard to AI-powered delivery; clients mainly provide creative assets and KPIs. Creative Studio and Copilot then cover asset production, performance reviews, and optimization recommendations.
Microsoft’s enterprise AI bet is not on a single model, but on a model-neutral platform, a complete toolchain, and existing enterprise distribution. 段威之 said Azure AI Foundry has aggregated more than 11,000 models, including those from OpenAI, Mistral, Llama 3.5, DeepSeek, and xAI Grok 3, with Prompt Flow, Fine-tuning, and Evaluation supporting model selection and deployment. At the top layer, “Copilot is UI for AI” serves business users; Copilot Studio serves low-code users; the underlying layer is handed to IT and developers.
The real enterprise moat is not generating polished text, but enabling models to securely access proprietary data. 黄秀兰 said that if Microsoft 365 Copilot cannot access commonly used data in Microsoft Graph, its answers may “look polished and impressive, but actually be of little use.” If labels, classification, permissions, and data governance are not in place, employees could even ask for their boss’s salary, making AI adoption a forcing function for Modern Work cloud migration and end-to-end data governance.
Teams Interpreter shows how Microsoft turns model capabilities into workflow products: after identifying a speaker’s voice in roughly 5–10 seconds before a meeting starts, the system can output another language in real time using that speaker’s own voice. 黄秀兰 saw a Chinese team review business in Mandarin while foreign executives listened and interrupted without friction; in a meeting 段威之 attended with Orion, Chinese, English, and Korean were used simultaneously. “Isn’t that my voice? Why am I speaking English?” 段威之 said Teams Premium users can currently access the feature.
The next organizational bottleneck is not whether employees have AI licenses, but whether they can build new “muscle memory.” Microsoft uses internal boot camps, Viva Engage communities, Garage hackathons, product champions, and workload-level MAU and DAU metrics to drive adoption. 段威之 opposes a “Great Leap Forward-style” KPI exercise, but acknowledges that building new habits requires pressure and prescribed routines. In China, an advertising team’s technical-services staff taught themselves to build a customer-service Bingo tool, while 黄秀兰 used AI to code and launch a 24-point game, showing how Agents are putting execution capabilities in the hands of nontechnical employees.
Management’s discussion of GPT-5 was still forward-looking ahead of launch; the higher-conviction call is that the industry is moving from Q&A to the Agentic Era. 段威之 described the expected leap in multimodality, structured output, and reasoning as moving “from dolphins to whales,” while making clear she would not use it ahead of release. Her focus is on multiple Agents and multiple models working together on complex tasks. For investors, that also explains the financial figures cited in the episode: Microsoft’s AI business has surpassed $13B in annualized revenue, up 175% year over year, with commercialization relying on a closed loop across models, cloud, Copilot, data permissions, and organizational adoption.

Deep dive

1. Microsoft is testing AI at two revenue front lines: sales services and advertising

段威之 (Kinki) is responsible for emerging markets and SMB customers across mainland China, Hong Kong, and Taiwan, covering Microsoft’s entire B2B product line. Her team handles sales and service for products including Teams and Copilot across this customer base.
黄秀兰 (Landy) runs Microsoft China’s advertising business, including domestic advertising and Chinese companies going global. She said the team is small, but its combined domestic and outbound business is already the largest in Asia-Pacific and has maintained the fastest growth globally for the past three years. “The productivity per employee is actually very high.”
程曼祺 chose these two lines because sales, customer service, and advertising are core B2B AI use cases. The question is not just whether AI makes employees more productive, but whether it can generate opportunities, conversions, and revenue.

2. M365 Copilot’s first source of value is eliminating platform switching

段威之’s most-used tool is M365 Copilot, built on GPT-4o and embedded in Word, Outlook, PowerPoint, and Teams. What matters to her is not point-solution generation, but being able to access email, meeting, and document context “on the same work platform.”
After returning from a week of leave, she used to reserve most of a day to work through her inbox message by message. Now she asks for the most important items from the previous week, and Copilot produces a table of key events, action items, deadlines, and relevant people—turning the inbox directly into a task list.
Japanese was one of her undergraduate majors, and writing English emails at a US company still felt difficult. Now she writes the first draft herself and asks Copilot to improve the language. The benefit goes beyond efficiency: the emails are “more professional,” and she feels more confident sending them.
黄秀兰’s workflow has also flipped. When information is missing, she used to ask an assistant or team member to look it up; now she asks Copilot first. “Rather than asking a team member next to me, I’ll search with Copilot first,” reducing the team’s low-value information relay work.

3. Products become indispensable through iteration, not a single launch day

段威之 felt that GPT-4o brought a clear step change, but refused to identify a single inflection point: “It keeps iterating.” The real change came when she suddenly realized the tool had become embedded in her workflow—and that returning to the old process would be difficult.
When M365 Copilot first launched, Simplified Chinese was one of the earliest supported local languages, but quality lagged far behind English. Early customer feedback was poor. At that point, the stronger value proposition was enterprise security, compliance, and access to Microsoft Graph, SharePoint, and Outlook data.
Simplified Chinese improved first, followed by major gains in Japanese and Traditional Chinese. 段威之’s view is not that the product suddenly “exploded,” but that language quality and usability gradually crossed the threshold for everyday adoption.

4. Teams Interpreter has turned multilingual meetings into a single conversation

In one complex business review, a Chinese team spoke entirely in Mandarin while management from a foreign partner listened online and interrupted with questions, with almost no language barrier. After 黄秀兰 turned on Teams Interpreter, she heard a colleague speaking English in his own voice in real time. Afterward, he was surprised: “Isn’t that my voice? Why am I speaking English?”
During a customer discussion with Orion, 段威之 connected Asian Global Black Belt technical experts. The China team spoke Chinese, while English, Chinese, and Korean were used simultaneously. Each participant selected an input and output language, and the meeting proceeded with “not much communication friction.”
Both executives said the system may need roughly 5–10 seconds before the meeting to identify a speaker’s voice. It then translates not only the content, but also preserves the speaker’s vocal tone and cadence.
On commercialization, 段威之 did not offer a simple answer on whether the feature would be fully open to all users. She only said that the pricing structure is complex and that Teams Premium users can currently access it.

5. Advertising AI now covers the full chain from accounts to creative, delivery, and review

黄秀兰 breaks Microsoft’s advertising AI into four parts: ad-account creation and management, Creative Studio for asset production, Performance Max for intelligent delivery, and post-campaign data analysis and optimization recommendations. “The most obvious gains are in two areas: improving efficiency and improving advertising performance.”
Microsoft’s advertising inventory includes not only Bing Search and the MSN display network, but also gradually connected ecosystems including Outlook, King, and Activision Blizzard. Clients do not need to plan channels and timing upfront. They provide creative assets, objectives, and KPIs, then let Performance Max allocate delivery intelligently.
Asked whether intelligent delivery includes creative generation, 黄秀兰 drew a clear line: Creative Studio, embedded in the ad platform, can assist with asset generation; Performance Max handles delivery; AI handles the rest. After the campaign runs for a period, Copilot analyzes performance and proposes optimization plans.
AI is therefore not replacing a single production step. It is creating a closed loop across generation, distribution, measurement, and iteration, with clients mainly responsible for providing assets and objectives.

6. Samsung anchors the advertising gains at 2.8x and 35%

黄秀兰 said that after Samsung adopted Microsoft’s AI Ads solution, click-through rates on Microsoft’s platform rose 2.8x, while time spent on account analysis and reporting fell 35%. She also qualified the result: “Every client may be different, and the data may not perform the same way.”
程曼祺 asked whether the click-through figure included distribution through social media. 黄秀兰 explicitly denied that: the number refers to Microsoft-side delivery, primarily Bing Search ads and MSN display ads, rather than Samsung’s consolidated marketing activity.

7. The 65,000-employee deployment sample is already producing sales results

段威之 cited data Microsoft had disclosed at AI Tour: roughly 65,000 employees across customer service, marketing, sales, and partner functions in MCAPS had received M365 Copilot licenses, with adoption continuously driven internally.
After deployment, per-capita opportunity creation rose 10%, opportunity-closing speed improved 23%, and revenue per salesperson increased 9%. She said the figures must be viewed together; “time saved” cannot substitute for business outcomes.
Copilot for Sales pulls together customer profiles, business strategy, CRM data, Teams meetings, email exchanges, and internal product materials, then highlights key points for pre-meeting agendas. After meetings, it automatically generates minutes and action items, allowing salespeople to complete more prescribed actions—and at higher quality—in the same amount of time.
For deals that remain open, an internal Agent can recommend Microsoft programs and resources to leverage based on deal size, product, and opportunity stage. 段威之 said internal programs are complex, and salespeople previously wasted time figuring out “where to find the resource.”

8. Copilot offers recommendations but does not make enterprise decisions

程曼祺 summarized the mechanism as “retrieving the old knowledge base, then depositing new meetings and documents into a new knowledge base.” 段威之 immediately corrected that framing: the system first understands the data and processes the current question, then returns to the existing data to generate an answer or recommendation. It should not simply be treated as automatically accumulating knowledge.
She repeatedly emphasized the boundary: “It doesn’t make decisions, but it gives you very good suggestions.” A person still has to prepare the content based on those suggestions, bring it to ready-to-send quality, and then send it to the customer or partner.
This also explains the improvement in closing speed. The Agent is not closing deals automatically; it is reducing friction in customer research, meeting preparation, note-taking, and internal resource matching, redirecting sales time toward judgment and deal progression.

9. Enterprise AI requires data accumulation, permissions, and governance upfront

黄秀兰 believes the key dividing line in customer outcomes, beyond training and adoption, is data. If a model cannot leverage Microsoft Graph and the data people use every day, the answer may simply “look polished and impressive, but actually be of little use.”
Relevance and accuracy improve when enterprise data is combined with customer information available online. But classification, labels, permissions, and data governance must come first. More data is not automatically better; the model must know who can see what.
She used an extreme example to illustrate the risk: if governance is poorly implemented, an ordinary employee could ask Copilot for the boss’s salary and retrieve it. Microsoft 365 Copilot relies on security permissions set through Microsoft Entra ID, with the goal of enabling AI within the company’s existing permission framework.
For customers that have not previously used Microsoft’s B2B suite and have a weak data foundation, 黄秀兰 said Microsoft provides end-to-end transformation. AI demand also “forces customers to migrate many applications to the cloud.” Once Modern Work is in the cloud, the Copilot experience is typically more complete.

10. Microsoft chose to be model-neutral rather than bet on one model

After 13 years at Google, 黄秀兰 “voted with her feet” to join Microsoft during the AI cycle. She was drawn to two things existing simultaneously: the willingness to make a massive investment in OpenAI, and the insistence that technology quickly become productivity and commercialization. “There isn’t a project that doesn’t need to consider a business model or a closed loop.”
Faced with customers asking why Microsoft does not have its own large model, 段威之 said Microsoft has moved beyond focusing on a single model. Its goal is to build an Azure AI Foundry that is “open and free, growing together.”
She said the platform already has more than 11,000 large and small models, including models from OpenAI, Mistral, Meta Llama 3.5, DeepSeek, and xAI Grok 3. Customers can use one model or combine multiple models.
Tools including Prompt Flow, Fine-tuning, and Evaluation handle model selection, optimization, and deployment. 段威之 compressed the strategy into one line: “For models, we are model-neutral; what we provide is a complete toolchain.” The final choice belongs to the customer.

11. Full Stack gives people with different technical thresholds their own entry points

段威之 describes the model as “Copilot is UI for AI.” Business users can click into it from familiar interfaces such as Word, Excel, and PowerPoint. Copilot Studio is designed for people who do not know how to program, enabling them to build mini-apps, workflows, and Agents through low-code and no-code tools.
More complex Agent platforms are handed to IT and developers, who gain greater flexibility through Azure AI Foundry, GitHub Copilot, and other tools. She calls this layering the AI full stack, designed to prevent AI from remaining trapped at the conceptual stage because of usage barriers.
Security sits at the base layer: “security as foundation.” Tools such as Microsoft Purview handle data governance, while Responsible AI standards cover auditing, data interpretation, and content filtering. The goal is to let enterprises “use AI without running into problems.”
The episode cited Microsoft’s fiscal 2025 second-quarter earnings: AI business annualized revenue had surpassed $13B, up 175% year over year. 段威之 grouped the product lines into three pillars: ABS (AI Business Solutions), CAIP (Cloud & AI Platform), and standalone Security.

12. GPT-5 was still a pre-launch expectation; the Agentic Era is the more certain direction

This episode was recorded shortly before GPT-5’s release. 段威之 said she would not use it ahead of time, so her comments were expectations rather than firsthand experience: multimodality, structured output, and reasoning could see a step-change improvement. The metaphor circulating among friends was that models had moved from small fish and shrimp to dolphins, while GPT-5 would be a whale.
She believes the Agentic Era is more important to watch. The discussion should no longer stop at Llama, DeepSeek, or any single Large Language Model, but focus on how multiple Agents and multiple models can be linked to handle complex tasks.
She divides the evolution of interaction into three stages: first, chatbot Q&A; second, taking action after receiving an instruction; and third, interactive work, in which multiple Agents collaborate to complete longer and more complex chains of work between people and machines.

13. Nontechnical users can already turn ideas into products with Agents

黄秀兰 has a liberal-arts background and no coding experience, but began asking tools to generate code and deploy it directly to understand AI. She and her husband enjoy playing 24 points, so she built a game called My Play. After encountering several minor issues, she eventually published it on the Apple App Store.
What impressed her most was not the app itself, but the shift in development from code syntax to natural language: “Someone with a liberal-arts background who can’t write code can also develop an app.” In her view, the key feature of an Agent is its ability to execute tasks for the user, not merely answer questions.
程曼祺 took that idea further: coding is becoming a form of creation. The episode’s closing formulation was that “coding is to an Agent what hands and feet are to a human”—it gives the Agent the ability to enter the networked digital world and act within it.

14. Microsoft is spreading AI capabilities through training, communities, and local innovation

Microsoft’s internal AI boot camps are taught by experts from different fields. Teams removes the physical limit on class size, while Interpreter lowers the barrier to cross-language participation. 黄秀兰 said team members voluntarily discuss recent courses over meals; participation is not driven entirely by mandatory enrollment.
“Growth mindset” requires both personal development and helping others develop. Employees record training and knowledge sharing in their half-year top priorities and reviews. In the M365 Copilot community on Viva Engage, one employee in China demonstrated how to build Snake using natural language.
Garage runs hands-on workshops and hackathons, giving employees from different departments a chance to test AI innovations and compete with one another. 段威之 believes an active, continuous community creates a stronger adoption environment than a one-off training session.
A technical-services colleague on 黄秀兰’s team taught himself to build an AI customer-service Bingo tool using public advertising documentation and routinely organized materials as its knowledge base. It answered frequent questions, improved efficiency, and proved popular. The team later demonstrated it to global leadership in hopes of turning the local creation into a global tool.

15. Adoption KPIs carry a gaming risk, but still provide necessary pressure to build new muscle

段威之 appoints “product champions” for Copilot and Copilot for Sales as role models, while monitoring MAU and DAU and drilling down to workloads such as Teams, Outlook, and PowerPoint to understand where different roles use AI.
She explicitly opposes mechanical daily clicking for the sake of KPIs: “Like the Great Leap Forward, everyone has to use it.” But she also acknowledges that forcing a team to “build new muscle” often requires pressure and prescribed actions early on. Habits may take 21 days or longer to form.
For sales, those prescribed actions include using Copilot to improve emails, generating a first draft of a PPT, reviewing historical records before meetings, and preparing proposals. There is also a more practical form of pressure: customer-service teams must “use it well themselves before they can teach it and answer customer questions.”

16. Travel and education show AI moving from answers toward personalized companionship

黄秀兰’s most common personal use case is travel. She enters a request for a seven-day Osaka trip with two children and their special needs, then repeatedly optimizes the itinerary until it is printable and executable. While visiting French colonial architecture in Da Nang, she could ask questions as she walked and have AI explain the history by voice.
Homework help has also moved from checking problems one by one to photographing work for correction. 段威之 prefers AI learning tools based on “Socratic reasoning”: instead of photographing a question and receiving the answer, the child is guided through a series of questions to reason it out. That gives parents more confidence than simply providing a shortcut.
She has used Duolingo for years and likes its gamified learning, AI role play, and adaptive pathways. After a few days away or repeated difficulty with one unit, the difficulty and learning path adjust. She has studied Spanish for more than a year and has also tried Cantonese because her husband is from Guangdong.

17. Parenthood did not become a zero-sum choice for either executive, but support systems were indispensable

段威之 had her child while at EMC and said she never viewed pregnancy during a career upswing as a setback or crisis: “Everything happened naturally.” She also explicitly credited her husband with taking on more household work; when helping their child with homework, her husband’s greater patience was an advantage.
During her 13 years at Google, 黄秀兰 had two daughters two years apart and received two promotions under the same manager while taking on increasingly broad responsibilities. When she was pregnant with her first child, she had just been recommended for a more stressful role. After she disclosed the pregnancy, her manager did not change course: “Why not take it?”
黄秀兰 attributed the experience to multinational companies’ support for working mothers and diversity, a good manager, and family cooperation, while noting that she did not know the working culture at other companies in China. On younger people’s choices, 段威之 said the decision is not a simple yes-or-no question; each person can follow a different life path.

18. Long-term career advantage comes from credible choices and finding the right way to speak

段威之 has gradually developed three criteria for choosing a company: first, the direction of the industry; second, whether the product platform is worth selling to customers over the long term; and third, culture, chemistry with the manager, and trust within the team. She moved from databases and enterprise disaster recovery to AWS cloud and Microsoft AI, staying focused on technology sales because the “next opportunity” often emerges naturally from the accumulation of the previous role.
If she could advise her 15-years-ago self, she would simplify her life earlier: no drinking, bars, card games, or karaoke, and no pointless socializing. “The simpler your material desires, the richer your spirit.” Customers ultimately care less about happy time than whether you can provide valuable service over the long term; trust is accumulated over time.
黄秀兰 is an INTJ, a longtime introvert, and not a self-promoter. As the only non-native English speaker on a global team, where most colleagues were foreign men, she used to stay silent in meetings. Her manager would pause the more talkative participants, call on her to speak, and remind her that she knew the business, customers, and product best: “Your English is good enough.”
She does not interpret “being yourself” as staying permanently in her comfort zone. Instead, she pushes herself in ways that fit her: replacing large receptions with high-quality one-on-ones, preparing thoroughly before meetings, taking the front seat, and being the first to speak. 段威之 likewise set herself a flag to speak in every global meeting. AI has pushed her further from reading and typing toward listening and natural-language input: “AI will gradually change you.”

Alpha School's Disruptive Vision for the Future of Education Aaron Levie and Steven Sinofsky on the AI-Worker Future