AI won't break SAP, but it will shatter seat-based pricing
AI won't flatten what enterprise software has accumulated, but it will shatter seat-based pricing. SAP's defense isn't training its own foundation model — it's open partnerships plus business understanding. The real battlefield is the last mile: process, data and organizational inertia.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The real disruption is the pricing model
AI's impact on SAP isn't at the code layer, it's in the business model. SAP's stock fell from a high of $300 to $160, but stretched out to 2021 it still grew more than 20%; Adobe dropped from about 480 to around 220, and Zoom has almost vanished. What's actually being replaced is point-solution SaaS. The core of the shock is the traditional seat-based pricing model, which will shift toward consumption-based or even outcome-based pricing — a challenge every enterprise software company will have to face.
— Yuan XinDon't train a foundation model — stay open
SAP doesn't train its own general-purpose large model, because foundation models evolve too fast and what SAP is good at is enterprise process, data and interaction. SAP's approach is to open up to all third-party models, and on model selection, pair the most suitable model with the most suitable scenario. At the same time it trains its own proprietary small model, RPT 1.5, on its own structured table data, for financial scenarios like receivables forecasting where accuracy needs to be close to 100% — because in finance, 99% accuracy is still unusable.
— Yuan XinThe boss wants AI, but the dirty data is still there
What big-company CIOs dread most is the boss reading an article on WeChat and forwarding it with "I want that too." The gap is that everyone assumes once AI arrives, the dirty, grubby work no longer has to be done. AI can accelerate data cleaning, but it can't make that process disappear. To land AI, you still need FDEs to comb through historical data and organize it by business object, and at the same time to "distill" the veteran's unstructured decisions into enterprise memory, so that the share that can be automated keeps growing step by step.
— Yuan XinFDE and consultants will eventually merge
FDE is not a rebranded on-site engineer. Traditional ERP consultants are industry experts who enter from the business side and ultimately land requirements into a standard product; FDEs deploy the people who build the product to the customer ahead of time, and the capability they distill flows back into the product. Early FDEs leaned toward low-level engineering and lacked a converter to turn business needs into coding requirements, which is why Anthropic and OpenAI started partnering with consultancies like McKinsey and Accenture. The two groups will merge into a hybrid role that understands both business and technology — which is exactly the necessary path for spilling out of a small circle into the enterprise market.
— Yuan XinThe customs agent's accuracy flipped
A large innovative company built its own agent to handle customs declarations for 33 countries, and accuracy was only 60-70%, so people had to review the remaining errors anyway, to loud complaints. After SAP came in, relying on the GTC global tax platform's basic understanding of 33 countries' tax rules plus document AI's template configuration — replacing long-context tokens — accuracy jumped instantly to the 90s, and what started as a pilot in one country had thirty-plus countries clamoring to go live. This shows that exploring AI's business value doesn't require discarding old accumulation; standard process plus incremental value is the effective path.
— Yuan Xin28% attempt it, only 3% see returns
McKinsey interviewed more than 2,000 companies worldwide: 28% have already used AI on projects, but only 3% believe there is real business return. So very few projects actually go to production; most are still in POC and prototype stage. Big customers have money to burn on trial and error and move fast; small companies run fast with open-source tools; those struggling most are mid-sized companies in the few-hundred-million to roughly 10-billion range, where whether to embrace AI is strongly correlated with the founder's mindset — and Chinese companies are more willing to put money into the product side than the management side.
— Yuan XinGoing global rides on a localization base
SAP's global platform itself covers operations in more than 150 countries, and its product organization includes a localization team dedicated to tax compliance localization for each country. Chinese companies going global can borrow this base directly and only worry about their business. Lenovo has worked with SAP from 20 billion to several hundred billion in scale, and now uses process mining tools to comb through more than 2,000 processes to find breakpoints and room for efficiency gains. When acquiring European companies, the other side is naturally an SAP user, so system integration has a built-in advantage.
— Yuan XinIn their own words · checked verbatim
You can't escape handling the enterprise's process, organization and data; you can't escape understanding the customer's business; you can't escape how I turn business into my code. It still needs a human in the loop to intervene.
你逃不出来处理企业的流程组织和数据，逃不掉对客户业务的理解，逃不掉我怎么把业务转化成我的代码，他还是要有human in the loop去做干预的。
Yuan Xin0:00
I think it's not that it doesn't exist — it's just that the patterns of history tell us that spillover takes time.
我觉得不是不存在，只是历史的规律告诉我们，它外溢是需要时间的。
Yuan Xin30:51
The boss often reads something on WeChat about what AI can do, then forwards it and says "I want that too" — that's the biggest headache for CIOs right now, especially at big companies.
老板经常在微信上读了一个什么什么什么AI能干啥，然后丢给他说我也要这是现在就是尤其大企业的CIO最头疼的事。
Yuan Xin44:47
Because after all, most of your current business still follows certain rules. You just need to, on top of a well-laid foundation, fill in the parts that used to be less standardized, that you couldn't read, or that were very costly to collect — that's actually a very good state.
因为毕竟你现在的业务的大部分，其实还是遵循一定之规的。 你只需要在打底打的好的基础上，把那些以前不太规范。 和你不能读懂的或者收集那样的数据代价很大的那一部分补充上，其实就是很好的一个状态。
Yuan Xin56:30
Right now the whole AI world is one big world of coopetition, a big melee. Yes — nobody can do without anybody else, but at the same time everyone wants to carve out a piece of land for themselves.
现在大家其实整个AI的世界就是一个大的竞和的世界大混战啊。 对，就是谁也离不开谁，但同时大家都想去圈出来自己的一块地。
Yuan Xin58:40
For enterprises it's actually a relatively complicated thing — you play the game, you pay the price.
对企业其实相对来讲还是比较复杂的一个东西，就是出来混，总是要还的。
Yuan Xin1:11:28
Figures
| Companies in the global top 100 that use SAP | 99 | 3:15 |
| SAP cloud users | more than 300 million | 3:15 |
| SAP global employees | 110,000 | 3:15 |
| Countries SAP's business covers | more than 150 | 3:15 |
| Share of SAP customers that are SMEs | 80% | 27:28 |
| Customs agent accuracy | from 60-70% to the 90s | 56:30 |
| Share of companies that have used AI on projects | 28% | 1:06:03 |
| Share of companies that believe there is real business return | 3% | 1:06:03 |
| Lines of SAP code | several hundred million | 18:03 |
| Number of Lenovo processes | more than 2,000 | 1:24:03 |
Glossary
- ERP
- The core system SAP was built on, managing end-to-end enterprise processes like finance, procurement and production.
- FDE
- Forward Deployed Engineer: goes to the customer site to land AI, while distilling the results back into the product.
- Human in the loop
- After AI executes automatically, human review of key judgments is still required.
- Process mining
- Using logs to analyze the real business flow and find breakpoints and room for efficiency gains.
- Physical AI
- AI combined with physical things like robots and EVs, considered a direction where China has an edge.
How to listen
Enterprise software founders, investors, heads of technology at companies going global, and engineers trying to figure out whether FDE is a good career direction.