Summary

袁欣 breaks AI’s impact on legacy software into 2 layers: coding makes “what used to be valuable no longer valuable,” but the deeper issue is the collapse of per-seat pricing. SAP may shift to consumption-based billing, while newer Silicon Valley companies are charging by outcome through RaaS. SAP’s stock has fallen from above $300 in mid-2025 to $160, but it is still up more than 20% from its under-$130 level at the start of the cloud era in 2021; single-product SaaS has taken a much harder hit, with Adobe falling from roughly $480 to around $220 and Zoom having “basically lost most of its gains.”
The moat is not code but business logic. S/4 has several hundred million lines of code, creating a massive gap versus reproducing it at a midpoint pace of 20,000 lines per day, but “reproducing the code itself is actually simple; the bigger challenge is figuring out what the business logic really is.” As long as “the world is still managed primarily by carbon-based organisms,” the underlying enterprise processes will not change, so there is little reason to rebuild the standardized base; the real battleground is using AI to serve personalized needs cheaply on top of standard processes.
His view of the FDE boom is that, unlike traditional on-site ERP consultants, FDEs are product engineers whose output flows back into the product. The current missing piece is “a converter”—someone who understands the business and can translate it into codable scenarios—which is why Anthropic and OpenAI are building deployment businesses while partnering with McKinsey, Accenture and PE firms. He expects the role may not reproduce the 10x or 20x consulting-industry explosion around 2000: “fewer but better,” with the engineering and business camps eventually merging into a more broadly capable professional.
The brutal implementation baseline is that McKinsey interviewed more than 2,000 companies, 28% were using AI in projects, but only 3% believed they were seeing real business returns. Most are still stuck at the POC stage. His customs example is the footnote: a major company built an agent for customs filings in 33 countries, but accuracy was only 60%-70% and “people got busier after deploying the agent”; SAP combined its GTC tax platform with country-specific Document AI templates, and accuracy “instantly rose to the 90% range,” turning a one-country pilot into a rollout that more than 30 countries competed to adopt.
The biggest enemy of scaling AI is the organization, not the technology. People who truly understand AI do not want to join traditional companies with only a few GPUs, while organizational gravity is harder to overcome than the talent shortage—“even in a high-tech company, not everyone will embrace it” when tokens are available to burn. His conclusion: “The technical challenges AI encounters when it truly scales in real-world settings are far smaller than the organizational challenges.”
The root problem behind China’s failure to produce a global SaaS giant is the pursuit of top-line growth. “A lot of the underlying systems were built very sloppily because the companies grew too fast.” AI is not a shortcut: “AI will not naturally fix the part you did not want to do,” and “you always have to pay your debts.” By contrast, overseas companies with stronger foundations can accelerate much faster once AI is applied.
The line with the most upside is the combination of Chinese AI applications with physical AI, with EV full-stack integration as the template: “I make everything from the chip to the car, compressing the space for Tier 1 suppliers.” The supply-chain foundation of companies that combine software and hardware could become a critical support layer—precisely where SAP believes it “still has a role to play.” SAP is partnering with foundation-model companies and opening its platform to every third-party model, while training a tabular-data vertical model, RPT-1, now at 1.5, because in finance “it may not be 99% accuracy that is acceptable; it may have to be 100%.”

Deep dive

1. The 54-Year-Old Giant: ERP Manages People, Money and Things

袁欣, who entered the industry in 1998 and has worked at Oracle, VMware and Microsoft before taking charge of SAP Greater China, opened with a portrait of SAP: founded in Germany in 1972, the 5-person startup grew into a B2B giant. Ninety-nine of the global top 100 companies use it; its cloud products have more than 300M users, it has 110,000 employees and operations in more than 150 countries, and generates more than €30B in revenue, or more than $40B. Over the past 2 years, its positioning has shifted from enterprise application software to “the world’s largest business intelligence software.” At this year’s Sapphire conference, SAP announced the autonomous enterprise—“an act of self-disruption by a traditional business software company.”
He reduced ERP to a simple skeleton: companies essentially manage people, money and things, with finance at the center. SAP’s original R/1 was finance-led, then expanded into “finance and business integration,” linking leads to cash with procurement to pay. “End-to-end processes ultimately roll up into a massive enterprise software system.”

2. What Feishu and DingTalk Cannot Reach: Identity, Security and Compliance Beneath the Interaction Layer

The dividing line with Feishu and DingTalk is clear: IM, meetings, documents and approval workflows are everyday OA interactions, but “the actual business processes still run in the back end, inside ERP systems.” Even before AI, Feishu wanted to penetrate applications through People, but once it moves deeper into enterprise-management workflows, identity management, security management and data integration become much more demanding. Even payroll is complicated across global enterprises because tax rules and calculation logic differ by jurisdiction.
The leverage in the ecosystem is substantial: the overall industry is roughly RMB1.2T, by what he recalls as an IDC estimate, while SAP itself generates more than €30B in revenue. SAP is “leveraging an industry worth dozens of times its own revenue.” Payroll requires connections to banks, tax authorities and e-invoicing systems; every exposed API raises questions around security and data consistency. That is why B2B is hard.

3. From $300 Back to $160: Point SaaS Dies First, but the Pricing Model Is What Collapses

曼祺 asked directly whether SAP’s stock falling from above $300 in mid-2025 to $160 meant the market was repricing the impact of AI coding on legacy software. 袁欣 changed the reference point: against the under-$130 level in 2021, the “first year of the cloud era” under the Rise with SAP definition, the stock is still up more than 20%. Much of the decline reflected “panic over all SaaS, with everyone voting with their feet.”
The companies showing the clearest signs of replacement are tool-like point solutions. Adobe fell from nearly $500, or about $480, to around $220; more competition-sensitive names such as Zoom “may have basically lost most of it,” after benefiting from the pandemic.
The deeper shock is to the business model. Traditional SaaS is priced on the user base: “You count how many users you have and I charge you accordingly … if everyone becomes agents, are you still going to charge by headcount?” Even if layoffs reduce the seat base only slowly, that may already count as outperforming peers.
SAP is changing its own pricing playbook. Going forward, “more of it will be consumption-based, priced by usage, rather than simply by the number of users or seats.” He also pointed to newer Silicon Valley companies charging by outcome. “Finding an accurate pricing model is a challenge facing SaaS companies.”

4. Hundreds of Millions of Lines of Code Can Be Compared; Business Logic Is Harder to Reproduce

The scale math is straightforward: generic coding runs at 2,000-3,000 lines per hour, while complex code with validation and security runs at 200-300 lines per hour. Using the midpoint, that is roughly 20,000 lines per day. All S/4 modules together contain several hundred million lines of code, versus 2M-3M for a relatively mature startup application ready for deployment. “This is only a comparison of scale. Reproducing the code itself is actually simple; the bigger challenge is figuring out what the business logic really is.”
The deeper point is that “as long as the world is still managed primarily by carbon-based organisms,” the way enterprises operate will not change materially in the short term. Factories will still produce in factories; there will simply be less human intervention. The foundational logic SAP has accumulated around generic business support remains relevant. “It is already relatively standardized. Why wouldn’t you use it?”
That makes the real battleground serving more companies’ personalized needs more efficiently and at lower cost on top of existing standard processes. What agents can do above the basic workflows “will definitely be a more important battleground.”

5. Disposable Software Works, but Growing Enterprises Buy Trust

曼祺 cited 陈宇森 of Alibaba’s idea of “disposable software”: code temporary, bespoke tools cheaply, use them for 1-2 weeks and throw them away. 袁欣 agreed that large companies will have such tools too, and that today’s architecture lets users hand-build them in an agent studio. But the boundary is clear: small companies can run on interpersonal trust, where “I only need to make my business visible.” Once a company coordinates externally, discloses as a public company or faces audits, permissions and audit trails become enterprise requirements. Data can be consumed by others only on a broader platform where everyone has reached agreement.
Going abroad adds another layer of complexity. Something built by hand “may simply not work in Mexico,” where tax rules, regulations, currencies and languages differ. His candid assessment: “I don’t think SAP is naturally your only choice. It depends on whether we evolve faster ourselves, or whether the new players targeting this market evolve faster.”

6. The Self-Disruption Lever: Joule Hides the Complexity of 400M Lines of Code

SAP recognizes that it is “on the side being disrupted, so it should work harder to disrupt itself.” The strategy is to retain its business and industry understanding while addressing the complaints users make most often: complexity, long implementation cycles and poor interactivity.
The architecture is concrete: Joule serves as the front-end work assistant; after intent recognition, an orchestration agent dispatches instructions and execution agents carry them out. The system “hides the complexity of the 400M lines of code underneath” and automates the process of finding and clicking through functions one by one. At the May Sapphire conference, SAP also launched its Business AI platform, the Joule Work interface and more than 50 intelligent assistants.

7. Why Companies Exist: 2 Invisible Hands

First, he corrected a common assumption: 80% of SAP’s customers are actually small and midsize companies, defined here as those with annual revenue below RMB1B. The “99 of the global top 100” line is a branding tactic.
He then went back to the underlying theory of the firm: companies exist because transaction costs within the boundary are higher than transacting externally. “As AI breaks the existing pricing model, if the cost of transacting inside a company is no longer lower than outsourcing to an agent, the form in which companies exist will inevitably change.” At the same time, there is a force toward aggregation: under uncertainty, adopting a federation model beneath a trusted umbrella makes it easier to build trust. “An invisible hand pulls things apart into the umbrella, and another invisible hand makes them aggregate.” It is too early to say what the final form will be.
History supplies the time frame. It took roughly 70-80 years for steam engines to raise productivity across society, 30-40 years for electricity and roughly 20 years for the internet. AI’s spillover is not absent; history says it takes time, with the trajectory accelerating and the spillover certain to arrive. For now, it is still a “celebration in AI-related fields.”

8. The Autonomous Enterprise: From a System of Record to an Executable System, with Human-in-the-Loop for 5-10 Years

The goal in one sentence is to turn SAP “from a system of record into an executable system.” But he was explicit that for the next 5-10 years, there will still be a human in the loop, because complex operations require human intervention.
Intercompany reconciliation illustrates the boundary. An agent can reconcile 70%-80% or even 90% of the entries, but “whether a difference caused by foreign exchange is an FX difference or a data error requires human judgment.” The same applies to the method and amount used for bad-debt provisions. Agents automate 70%-80% of routine work; people perform the final review and confirmation. Only then is the end-to-end process complete.

9. Working with Foundation-Model Companies: Maximum Model Access, Plus SAP’s Own Tabular-Data Model RPT

The division of labor follows the speed of model development. “Foundation models are evolving extremely fast—Cursor was acquired too. Something you once thought could be built in a niche can be absorbed as foundation models improve.” SAP’s strengths are enterprise operating processes, process data and accumulated knowledge around user interaction. It is therefore opening itself to every third-party model and offering the broadest selection of usable models, matching the best model to each use case.
SAP is also training its own vertical model on tabular data: RPT-1, now at 1.5. A typical use case is receivables forecasting. An external LLM captures customer goodwill and repayment pressure, while SAP’s internal system draws on structured historical order-fulfillment data. For a simple customer question, SAP aggregates different models and assembles the answer that best fits the requirement.
Zero tolerance for hallucinations is the anchor for model selection. “With the current technical paradigm of large models, there will be hallucinations to some degree. In financial management, 99% accuracy is not necessarily acceptable; it may have to be 100%.” Accuracy above 90% is already a high benchmark on leaderboards, but in the real world the data is unusable. “Some technology does not necessarily need to use a model.”

10. Putting a Model into B2B Is Not Enough: The CIO’s Biggest Headache Is the Boss Forwarding a WeChat Post

He avoided an absolute definition: if an entity like the one in the movies really becomes omnipotent, “even reaching that day may not mean” the current conclusion holds. But over the visible 5-10-year horizon, with existing enterprise forms largely intact, “simply throwing a model into B2B is definitely not enough.” Aggressive large companies are setting up AI task forces and pairing digital employees with people, which shows that humans still have to fill in the last mile.
CIOs who deal with this every day offer a vivid complaint: “The boss reads something on WeChat every day about what AI can do, forwards it to him and says, ‘I want that too.’” In general, “many enterprise leaders at the top do not know that much about technology.”
The biggest mismatch is around data. Many companies assume that once AI arrives, they no longer need to do the hard, dirty and exhausting work associated with data. That is a major misunderstanding. AI will accelerate data cleaning, but the work itself will not disappear. The real goal is for data to carry sufficient business context at the moment it is generated, rather than spending huge amounts of effort and tokens cleaning it later.

11. Enterprise Memory: Distilling the Master Craftsman’s Know-How

The substance of the ontology FDEs build is organizing vast historical data around business objects so agents can operate. But it has 2 parts. Fixed processes do not need to be reprocessed with tokens: “In SAP, steps 1, 2, 3, 4 and 5 are very clear. Just load the process and use it.” That is SAP’s biggest differentiator.
AI is needed for the external loop: breaks between systems, and human decisions formed in meetings and emails. This is where AI can process unstructured historical data and “distill the master craftsman’s know-how,” such as the judgment behind a provisioning ratio. Standard processes combined with extracted know-how become company memory. As that memory compounds layer by layer, the share of work that can be executed automatically will continue to rise.

12. The Fundamental Difference Between FDEs and Traditional On-Site Consultants: Their Output Becomes the Product

曼祺 asked what fundamentally separates FDEs—first popularized by Palantir—from the pre-sales and on-site engineers at SAP and Salesforce. 袁欣’s breakdown: traditional ERP consultants are industry or finance experts who start from the business side, work backward from financial-analysis outputs to design the account architecture, and ultimately return to a standard product plus customized extensions. Being on-site is not the essence: “Even if someone does development, he is still a consultant, not a product.”
FDEs work in the opposite direction: “Deploy the people who actually build the product to your site in advance, and what we build with you will ultimately become my product.” Because the question of what AI can solve and in what form it can solve it remains far from settled at the model layer, it requires engineers with deeper technical instincts to investigate on-site.
The market is signaling the same evolution. In May, Anthropic and OpenAI each announced an independent FDE organization or deployment company, partnering with consulting firms such as McKinsey and Accenture and with PE firms that open their portfolio companies to hands-on projects. The industry has realized that “deploying product engineering to the customer does not mean the code will naturally appear. There is no converter”—someone who understands the business and can turn it into a codable scenario.
Microsoft has also recently formed an FDE team, and SAP has one of its own. “Everyone agrees on the implementation approach, but there is not yet consensus on what should be retained from the FDE process. That is still being explored.”

13. The Customs Case: A Self-Built Agent Reached 60%-70% Accuracy and Made “People Busier”

The case was presented without embellishment: a highly innovative company handled customs declarations in 33 countries, where customs documents differ completely by country. It used low-code to build a digital employee that reviewed customs filings, but accuracy was only 60%-70%. “After deploying the agent, people got busier. They had to investigate where the remaining errors came from. There was no efficiency gain; people were even more frustrated.”
SAP’s difference was its existing GTC global tax platform, which has a baseline understanding of tax requirements in 33 countries, combined with Document AI configured through country-specific templates. “By slicing the data against a standard, you don’t have to build a long context every time, burn excessive tokens and suffer low accuracy.” Accuracy “instantly rose to the 90% range,” and the project expanded from a one-country pilot to a rollout that more than 30 countries competed to launch.
The methodological lesson is also the central theme of the episode: “In exploring where AI actually creates business value, you do not need to discard everything that came before. Most of your current business still follows established rules. You only need to build on a solid foundation and fill in the parts that were previously nonstandard or expensive to collect.”

14. A Grand Coopetitive Melee: Fixing the Business Is Harder Than Fixing the Technology, and Organizational Inertia Is Harder Than Both

Will SAP compete head-on with Anthropic and OpenAI? “I think it definitely will. The entire AI world is one large coopetitive ecosystem, a grand melee. No one can do without anyone else, but everyone wants to carve out a piece of territory.” The contest will be determined by how quickly each side fills its gaps: SAP must add engineering capability, while Anthropic must build business understanding. “It depends on who adjusts faster.”
Asked which gap is harder to fill, his personal view was unequivocal: “Technology is definitely a little easier to add.” Capabilities such as Cloud Code make it easier for business experts to work with foundation models; getting a model expert to understand how accounting is done is still much harder.
The more realistic challenge is organizational gravity. Some people lack the will, some have the will but not the ability, and some have neither. “Even in a high-tech company, not everyone will embrace and use it when you open the taps and burn tokens—let alone a large traditional enterprise with lower talent density.” His conclusion was categorical: “The technical challenges AI encounters when it truly scales in real-world settings are far smaller than the organizational challenges.” The talent gap is equally concrete: CHROs call him asking for recommendations, while people who truly understand AI ask traditional companies, “How many GPUs can you give me? Where is the data? I need to bring several people with me.”

15. Is FDE a Good Career Direction? Fewer but Better, with 2 Camps Converging

For job seekers, the implication is that a startup may move from needing 50-200 people to having 20-30 people do the work of 200-300. “Fewer but better—the statement that fewer, better people will do this work is valid, but it does not mean most people will have no jobs.” The meaning of FDE has evolved dramatically in less than a year, and “you still do not know what its final form will be.”
The part that will not change is the dirty work: “You cannot escape the dirty, exhausting work.” That means mapping enterprise processes, organizations and data; understanding the customer’s business; and translating business requirements into code. AI increases coding productivity, but does not automatically accelerate the translation process. Combining business and technology is “a necessary route for AI to move from a small circle’s celebration into the broader enterprise world,” and that work will be needed for some time.
Expectations for scale should be tempered. The role may not reproduce the 10x or 20x explosion of consulting firms around 2000, because labor productivity is higher. The 2 camps are also competing for the same jobs: McKinsey and Boston consultants moving downstream, and FDEs with stronger AI and coding backgrounds. “The 2 forces will merge into a more broadly capable person who continues working in this role.”

16. The Real-Money Stratification: 28% Are Doing It, Only 3% Are Seeing Returns

The macro backdrop is a broad cut in IT budgets alongside rising AI budgets. Traditional IT projects are increasingly difficult to approve unless they are linked to AI. But McKinsey’s interviews with more than 2,000 companies globally found that 28% were using AI in projects to some degree, while only 3% believed they were seeing real business returns. Few projects make it into production; POCs and prototypes are more common, while consulting firms are beginning to sell AI-scenario planning, prototype design and workforce-restructuring services.
The adoption map is split. Large customers move quickly because they have enough data, acute pain points and the financial capacity to experiment. Small companies can move fast on open-source software and AI tools. The companies struggling most are midmarket enterprises with revenue ranging from hundreds of millions of RMB to RMB10B, where progress depends heavily on the CEO’s understanding. Many companies with innovative, cool branding are willing to spend on AI-enabled products—putting large models into robot vacuums or having top talent optimize an edge model—but far fewer are willing to spend on AI productivity gains in enterprise management.

17. Why China Has Not Produced a SaaS Giant: The Bill for Chasing Top-Line Growth Always Comes Due

His diagnosis is coherent. Fast-growing companies chase top-line growth, with pressure concentrated on acquiring markets and customers. “A lot of the underlying systems were built very sloppily because the companies grew too fast,” leaving no breathing room to plan and design properly. Domestic SaaS also lacked pricing models and recognition of value strong enough to support helping companies build a solid foundation.
AI is not an exemption from the bill. “Do you think AI can naturally make a twisted, unstructured mess stand up and run immediately? It cannot. AI will not naturally fix the part you did not want to do … it is not a logic of taking a shortcut.” In the end, “you always have to pay your debts.” By contrast, overseas companies whose foundations are better—though he sometimes mocks how cumbersome it is to create a simple approval workflow—can accelerate much faster once AI is applied.
McKinsey’s cost accounting is the most consequential figure: every $1 spent on software iteration generates $3-$4 of change-management costs. “Simply switching to an AI-native platform is still old wine in a new bottle if nothing substantive changes.” Much of today’s so-called AI infrastructure is actually paying down the unfinished work of past digitization.

18. Industry Order and Physical AI: The Supply-Chain Foundation Matters

AI will penetrate light-asset and service sectors such as financial services and the internet more easily. In traditional manufacturing, the production form has changed relatively little because robots are not yet a mainstream replacement. The first impact is showing up at both the market and after-sales ends, while on the production side it is currently more concentrated in supply chains.
His forward-looking thesis is that “even Anthropic is investing in physical AI, and OpenAI started earlier.” If China is more likely to win in applications and physical AI by combining its supply-chain advantages, “the next wave of high-growth Chinese companies will definitely emerge in physical AI.” EVs are the template: “I make everything from the chip to the car, compressing the space for Tier 1 suppliers.” Is that a traditional company or an innovative one? It is a combination. The conclusion returns to SAP: strengthening the supply-chain foundation is critical to supporting the next wave of industrial growth, and “from this perspective, SAP still has a role to play.”

19. China’s Playbook: Built on Alibaba Cloud, Qwen in MaaS, 曼森’s Closed Loop and Lenovo’s Global Expansion

Because “China’s AI ecosystem is relatively unique,” SAP first deploys its products on Alibaba Cloud infrastructure and embeds Alibaba’s Qwen model into its own MaaS, or Model as a Service, offering AI services externally. The 2 sides are also exploring China-specific AI use cases through FDE, post-training and shared customers; their FDE teams may serve customers together.
The small-customer example is specific. 曼森, a clothing and down-jacket company that owns a brand and also does contract manufacturing, runs Public Cloud ERP on Alibaba Cloud and uses DingTalk day to day. The boss’s pain point was that the company served large brands and spent every day taking calls about delayed orders and follow-up shipments. SAP used an AI agent to close the loop between ERP data and order-chasing information in DingTalk, then brought production scheduling into the process to address the core problem of order fulfillment.
Overseas expansion is another area where SAP has a distinctive position. Its product architecture includes a dedicated localization team—Brazilian tax is “the most complex in the world”—and supports tax compliance in more than 200 countries. Customers only need to focus on their own business. When Chinese companies acquire European businesses abroad, those companies are naturally SAP users, creating an inherent advantage for system and process integration.
Lenovo is the large-customer example. SAP has worked with it from RMB20B in revenue to its current scale of several hundred billion RMB. More than 80% of Lenovo’s business is overseas, and SAP has supported system integration through each overseas acquisition. Lenovo now also uses Signavio process mining to examine the breaks and efficiency opportunities across more than 2,000 internal processes from a global perspective.

20. The Foreign-Multinational Employee’s Challenge: Process Discipline Plus the Superindividual

He was candid about the difficulty of SAP’s internal transformation: “Foreign-enterprise employees in China have not started businesses themselves. They are very professional when following processes, but when asked to think independently and act entrepreneurially, the existing high-quality workforce will diverge.” The task is to create an atmosphere that encourages people to take risks and allows more superindividuals to emerge. The company’s mantra is “return to a startup mindset and do it again”—after all, SAP itself was a 5-person startup in 1972.
The details are vivid. Internally, employees can use Cloud without limits. Those with ideas buy 2-3 Mac Minis, set them up at the office and use company resources to build different implementations; one of those experiments might become a superindividual. The combination he favors is the process discipline, systems familiarity, international perspective and ability to manage complexity developed in a multinational, combined with AI-enabled individual initiative. Together, they become “more powerful and more effective.”

21. 3 Cycles in 30 Years: Only the Speed Changes; People Do Not

From on-premise deployment to the cloud, and from informatization to intelligentization, his summary is one sentence: “The difference is speed.” AI changes human-computer interaction and does an excellent job of democratizing access—everyone can use it easily—so evolution is much faster. But the underlying logic has not changed. Breakthrough technologies affect every industry through a process; each cycle still moves through aspiration, fear, learning and rejection. “As long as humans still rule this planet, that underlying logic will not change.”
The closing insight is the one that lingers: “Human judgment cannot be scaled.” AI can process massive datasets, but people cannot absorb or handle that many signals. “When implementing your ideas is no longer the bottleneck, your own ideas become the bottleneck.” He pushes the impact further, asking how education systems can produce individuals capable of harnessing AI. That may bring a redesign and disruption of education itself. The trait VCs cite most often—strong learning ability—will continue to determine who reaches the top of the pyramid, because “in every era, people with that trait are the ones who ultimately become the people at the top.”

I Still Believe in AI! — A Cross-Show Episode with Three O'Clock Off Nathan Goes to China – Part 1: Tech & Agent Setup, Chinese AI UX, WAIC, and Attitudes on AI