SAP Links Autonomous AI Strategy to Hard ROI at Enterprise Scale

SAP: Advancing the Autonomous Enterprise: Less Chat. More Act. Guided by Fact.with SAP — Ai4

Ai4

SAP argues enterprise AI requires grounding in trusted, process-aware data rather than basic chat interfaces to generate real financial return and avoid operational risk.

AI-assisted summary of SAP: Advancing the Autonomous Enterprise: Less Chat. More Act. Guided by Fact.with SAP — original source. Check the original for full context.

Source published: 2026-08-21

Detailed notes

## Background Published on YouTube by Ai4 on August 21, 2026, this presentation features Bob Sakalas, Innovation Strategist at SAP. The video outlines SAP's Autonomous Enterprise vision, illustrating how organizations can move beyond conversational AI chat toward reliable, process-grounded AI agents governed at enterprise scale.

## TL;DR Bob Sakalas presents SAP's strategy for advancing the autonomous enterprise by shifting AI from simple chat interfaces to actionable, fact-guided automation. He highlights hard-dollar ROI examples from major corporations like JPMorgan Chase, Saudi Aramco, and Mastercard, while addressing common pitfalls in proof-of-concept deployments, particularly around data quality. By pairing LLMs with SAP's extensive Knowledge Graph, Common Domain Model, and strict enterprise governance, SAP enables organizations to automate complex workflows safely. Real-world implementation stories, including Heineken and Cirque du Soleil, demonstrate how process-aware AI drastically reduces operational friction and drives revenue growth without expanding headcount.

## Atmosphere & style The session is lively, fast-paced, and highly engaging, with Sakalas warning early on that the presentation will move at the speed of "motion pictures." He blends pop culture references, such as quotes from Ferris Bueller, with relatable personal anecdotes and self-deprecating humor to keep the audience entertained. Directly interacting with audience members and calling out specific industry peers, Sakalas maintains an energetic, pragmatic tone that balances high-level vision with sharp, real-world business realities.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) SAP system support is critical to global commerce, powering roughly 85% of global transactions primarily across finance, supply chain, and human resources. - [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) Enterprise AI development is still in its earliest phases, characterized as being at 'stage five out of 21' on a Tour de France scale. - [02:01](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=121s) Major industry leaders are already demonstrating massive hard-dollar ROI from AI automation, including $2 billion returned at JPMorgan Chase, $4 billion in operational value at Saudi Aramco, and $35 billion in prevented fraud at Mastercard. - [03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) Deploying uncontrolled AI directly inside systems of record poses serious business risks, ranging from shadow IT data leaks to costly operational failures caused by hallucinations. - [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) Inaccurate or inconsistent data is the primary reason enterprise AI proof-of-concepts fail to transition into production environments. - [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) SAP standardizes semantic enterprise data using a Common Domain Model and Business Data Cloud to support zero-copy data sharing with ecosystem partners like Databricks, Snowflake, and Collibra. - [14:02](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=842s) To deliver reliable results, enterprise AI requires process-grounded domain context, as demonstrated by SAP's Knowledge Graph achieving 100% accuracy on operational order and recall checks compared to ~75% for standalone LLMs. - [14:02](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=842s) SAP defines its enterprise AI competitive framework around three key pillars (PDG): Process expertise, Data authority, and enterprise Governance. - [16:22](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=982s) The operational architecture of the autonomous enterprise scales from single-task cell assistants (e.g., Joule Studio) to departmental organs and multi-LOB business integration while keeping humans in the loop. - [18:34](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1114s) Cirque du Soleil reduced invoice dispute resolution times from 30 minutes to 2 minutes by deploying an SAP AppHouse AI assistant to help a 4-person team manage 70,000 multi-language invoices.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) Introduction and the Speed of AI Transformation

Bob Sakalas opens the session with a lively, fast-paced tone, warning the audience that because they only have 20 minutes, the presentation is going to run "like motion pictures." Establishing context, he asks the room how familiar they are with SAP before emphasizing its massive footprint as the largest ERP company on the planet. He highlights that roughly 85% of global commerce touches an SAP system, with core operational focuses centered squarely on finance, supply chain, and human resources. Sakalas adds a quick, standard forward-looking disclaimer: while SAP’s future direction is clear, they reserve the right to fine-tune details until features become generally available.

Transitioning to the broader landscape of AI, Sakalas quotes "the great philosopher Ferris Bueller" to emphasize that "life moves pretty fast," warning that if you don't stop and look around, you might miss it. He asks the room if everyone's job is going to change, asserting that it's a "100%" certainty. To illustrate, he points to Walmart CEO Doug McMillon, who manages 2.2 million employees—many performing manual labor like stocking shelves or buffing floors. McMillon is on record stating AI will redefine every job; for instance, if a store goes a few minutes without selling a Diet Coke or Coke Zero, intelligent systems will instantly recognize the anomaly and trigger a inventory shelf-check.

- SAP powers approximately 85% of global commerce, primarily across finance, supply chain, and HR. - According to Sakalas, enterprise AI is currently only at "stage five out of 21"—using a Tour de France analogy to show we are still in the early stages. - Doug McMillon manages 2.2 million employees at Walmart and envisions real-time AI inventory monitoring (e.g., tracking 3-minute sales cycles of Coke Zero).

---

[02:01](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=121s) Proving Real-World ROI in Enterprise AI

Addressing widespread skepticism around whether AI investments are actually paying off, Sakalas presents concrete evidence of massive hard-dollar ROI across major industries. He cites JPMorgan Chase CEO Jamie Dimon, who described AI as a technological shift as transformative as electricity. Sakalas shares a personal anecdote about his eldest daughter, who works at JPMorgan; when she comes home, he asks her how much she relies on AI, and she routinely reports using it for three to four hours a day.

Sakalas then turns to the energy and financial sectors. Referring to an audience member from Halliburton, he notes how oil and gas rely heavily on repetitive tasks like seismic data interpretation and asset maintenance. He quotes Amin Nasser, CEO of Saudi Aramco (the largest oil company on Earth), who revealed that applying AI to repetitive operational workflows yielded over $4 billion in return. Similarly, the CEO of Mastercard reported that AI-driven fraud prevention systems have helped the company prevent $35 billion in fraud to date.

- **JPMorgan Chase:** Invested $2 billion into AI and measured a direct $2 billion hard-dollar return. - **Saudi Aramco:** CEO Amin Nasser cited $4 billion in value generated through AI automation of repetitive oil and gas tasks. - **Mastercard:** Leveraged AI fraud detection to block an estimated $35 billion in fraudulent transactions.

---

[03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) The Autonomous Enterprise Vision and Modernization Challenges

Sakalas traces SAP’s vision for the autonomous enterprise back to 2017, when co-founder Hasso Plattner took the stage at SAP Sapphire to address customer demands for simpler ERP and HR operations. While the technology wasn't fully ready back then, modern advancements have finally caught up to make Plattner's vision a reality. Sakalas conducts a quick audience poll on AI adoption: while almost everyone uses consumer AI for personal tasks or email drafting, far fewer use it within core business processes. However, executive pressure is shifting rapidly because board members and CFOs are now personally experimenting with AI on their mobile phones.

This top-down push forces IT leaders to address the underlying technology stack. Enterprise systems in Fortune 500 or Fortune 1000 companies are often "creaky" and "held together with bubble gum and duct tape." Modernizing these complex architectures is necessary, but Sakalas warns that enterprise AI represents a double-edged sword—much like the Chinese symbol representing both opportunity and crisis simultaneously.

- Hasso Plattner laid the groundwork for SAP's autonomous enterprise strategy at Sapphire 2017. - Sakalas notes an informal survey showing only ~32% of enterprise audiences actively use AI inside core operational business processes. - **Risks of Uncontrolled AI:** - **Shadow IT:** Uncontrolled local data extracts expose companies to severe cybersecurity threats. - **Hallucinations in Systems of Record:** A single hallucinated data point could redirect a cargo ship to the wrong port (costing millions), take a functional airplane offline unnecessarily, or misreport quarterly financials to Wall Street.

---

[07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) The Business AI Platform and the Data Quality Bottleneck

To prevent complexity sprawl, Sakalas advises against creating disjointed, point-solution AI tools with a dozen different vendors. Enterprise AI requires a unified platform that keeps the core software clean while allowing repeatable innovation. While building a quick proof-of-concept (PoC) using public data tools like Claude or custom scripts has become trivial—taking weeks instead of six months—moving a PoC into full enterprise production remains difficult.

Picking on an audience member named John, Sakalas asks why most PoCs fail to reach production. The answer is single-word and definitive: **DATA**. If underlying data does not accurately reflect reality, the AI will fail 100% of the time. Problems arise when master data entities (like a "customer") are defined inconsistently across ServiceNow, Salesforce, and SAP, or when agents operate on stale, historical data. For instance, Southwest Airlines cannot rebook a flight on last week’s data. Achieving semantic context required SAP over 5 years and more than 10,000 engineers to build a unified Common Domain Model.

- **Proof-of-Concept (PoC) Acceleration:** AI allows teams to spin up PoCs in 2 weeks compared to the traditional 6 months. - **The Common Domain Model:** Took 5 years and 10,000+ people to standardize semantic data across enterprise applications. - **SAP Business Data Cloud (BDC):** A SaaS data fabric designed to maintain SAP context (finance, HR, supply chain) without creating redundant data copies. - **Ecosystem Partnerships:** - **Zero-Copy Data Sharing:** Partnered with **Databricks** for ML/AI workflows without duplicating data. - **Governance & Lineage:** Partnered with **Collibra** for end-to-end data cataloging and lineage visibility across AWS, Databricks, and legacy systems. - **Open Hyperscaler Support:** Integrations extend across **Confluent**, **DataRobot**, **Google BigQuery**, **Snowflake**, **Microsoft Fabric**, and **Amazon Athena**.

---

[12:43](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=763s) Real-World Implementation: Heineken and Process-Aware AI

Sakalas shares a practical case study featuring Heineken, quipping that he became personally "familiar with the brand last night." Heineken initially tried delivering business insights using Microsoft Copilot alone, but quickly realized the underlying data lacked proper structure and context. By implementing SAP Datasphere (within BDC) and the Generative AI Hub, Heineken built an assistant called **"Ask Copy."** This provided employees with rapid, reliable access to enterprise data anchored in accurate business context.

Sakalas highlights a memorable quote regarding data value: "Data without data products is like crude oil delivered to your backyard." While crude oil is one of the world's most valuable commodities, a raw barrel sitting in a backyard is useless without refinement. To drive this refinement, SAP is focusing heavily on Agentic AI that orchestrates business workflows. Sakalas emphasizes that SAP’s vision is open—SAP agents will interoperate seamlessly with customer-built agents and third-party agents from platforms like Salesforce.

---

[14:02](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=842s) Demonstrating Knowledge Graph Precision and Core Architecture

To prove the necessity of process-grounded AI, Sakalas highlights SAP’s massive Knowledge Graph. Built on top of 450,000 ABAP tables, 80,000 CDS views, and 7 million data fields, the graph embeds structural domain expertise directly into AI models. He describes a head-to-head live demonstration created by his teammate, Shri, comparing a standard LLM agent against an SAP Knowledge Graph-backed agent:

1. **Delivery Status Query:** The standard LLM achieved ~75% accuracy, whereas the SAP process-aware agent reached 100% accuracy for a fraction of a cent in token costs. Sakalas asks the room if 75% accuracy is acceptable in enterprise business, answering with a resounding "No." 2. **Product Recall Check:** When queried on order fulfillment during an active recall, the standard LLM declared the order clear to ship. The SAP agent correctly identified the active product recall and blocked the shipment.

Sakalas summarizes SAP’s enterprise AI differentiation through the acronym **PDG**: - **P — Process:** 50 years of deep domain expertise spanning 28 distinct industries. - **D — Data:** Unmatched authority as the primary system of record for financial and operational data. - **G — Governance:** Strict auditing and compliance capabilities required for Wall Street-grade reporting.

This philosophy is captured in Sakalas's central mantra: *"Less chat, more act, guided by fact."*

---

[16:22](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=982s) Structuring the Autonomous Enterprise: From Cells to Organs

In the closing segment, Sakalas explains how SAP builds agentic architectures while keeping humans in the loop. Insurance constraints and operational risk mean fully automated, unmonitored systems will not happen overnight. Instead, SAP constructs AI capabilities from the ground up using a biological metaphor:

- **The Human Cell (Task Level):** Represents a highly specialized AI assistant focused on a single outcome, such as a financial closing assistant. - **Joule Studio:** SAP's developer tool that allows customers to extend standard agents or build custom agents tailored to their unique business processes (e.g., accounting for differences between Halliburton and Coca-Cola). - **The Organ (Departmental Level):** As individual cells (task agents) are combined with complete auditability and governance, they form functional organs—such as an entire Accounts Payable, Accounts Receivable, or Tax workgroup. - **The Full Body (Line of Business Integration):** These functional workgroups scale across lines of business (LOBs)—spanning Spend Management, Supply Chain, Human Resources, and Customer Experience (CX)—to power the broader autonomous enterprise.

[18:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1084s) Delivering Reliable Results and the Path to the Autonomous Enterprise

Bob Sakalas details how SAP connects agentic AI from the ground up to ensure enterprise results that organizations can actually count on. The strategy combines the best large language models with SAP’s deep, foundational understanding of business processes, anchored by the customer’s proprietary data that remains fully protected and never escapes to the outside world. This synthesis is driven by the interaction between the Knowledge Graph and Business Data Fabric (BDC). The ultimate business objective is straightforward and powerful: allowing companies to drive significantly higher revenue without being forced to scale up headcount at the same rate.

- **Process & Data Integration**: Merges LLM capability with SAP's business process domain knowledge and private enterprise data. - **Data Privacy**: Enterprise data used by the system is strictly guarded and never exposed to the outside world. - **Core Value Proposition**: Enables substantial revenue growth while maintaining flat headcount expansion.

---

[18:34](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1114s) Getting Started with Workshops and Cirque du Soleil’s Success Story

To translate this framework into reality, SAP starts with targeted customer workshops to pinpoint high-impact pain points where AI intervention yields massive cost savings. Sakalas brings this to life with a story shared by Philippe at SAP Sapphire about Cirque du Soleil ("Circus"). After selling roughly 10 million tickets in 2024, Cirque du Soleil faced a mounting administrative hurdle managed by a leader named Jessica. Her lean team of four people was inundated with processing 70,000 invoices arriving in various languages. Whenever an invoice dispute or error occurred, resolving it required extensive research.

Self-deprecatingly poking fun at himself for wearing the wrong glasses and struggling to read his slides, Sakalas explains how Jessica’s team partnered with the SAP AppHouse in Palo Alto. Using SAP's Business AI platform, they implemented an AI-enabled invoice assistant. The results were dramatic: the average time required to research and resolve a faulty invoice plummeted from 30 minutes down to just 2 minutes.

- **Cirque du Soleil Scale**: Managed over 10 million ticket sales in 2024. - **Invoice Challenge**: A small team of 4 people was responsible for managing 70,000 multi-language invoices. - **Efficiency Impact**: Reduced invoice resolution time from 30 minutes to 2 minutes using SAP's Business AI invoice assistant developed at the Palo Alto AppHouse.

---

[19:53](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1193s) Surfing the Wave of Innovation and Core Takeaways

Reflecting on SAP's history, Sakalas emphasizes that the company has continuously succeeded by "surfing change" across major tech shifts—from mainframes and client-server models to the internet explosion. However, the excitement around Agentic AI surpasses previous eras because it natively aligns with SAP's core expertise in process execution. Channeling *Ferris Bueller* with a reminder that the future comes moving fast, Sakalas challenges the audience on whether anyone can afford to wait.

He concludes with the foundational pillars of why SAP is uniquely positioned for AI: deep process context, reliable data, and robust enterprise governance. Optimizing applications, ensuring pristine data foundation, and infusing Business AI directly into daily workflows are the essential steps to building an autonomous enterprise.

- **Historical Perspective**: Tracks SAP's evolutionary tech shifts across mainframes, client-server architectures, the internet era, and now Agentic AI. - **Three Pillars of SAP AI**: Process expertise, data integration, and enterprise governance. - **Resource Link**: Sakalas directs the audience to `[prontock.com/ai4](https://prontock.com/ai4)` to download his presentation deck before wrapping up to applause.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) run like motion pictures — Bob Sakalas

> [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) life moves pretty fast — Bob Sakalas

> [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) stage five out of 21 — Bob Sakalas

> [03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) creaky and held together with bubble gum and duct tape — Bob Sakalas

> [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) DATA — Bob Sakalas

> [12:43](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=763s) familiar with the brand last night — Bob Sakalas

> [12:43](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=763s) Ask Copy — Bob Sakalas

> [12:43](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=763s) Data without data products is like crude oil delivered to your backyard. — Bob Sakalas

> [14:02](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=842s) Less chat, more act, guided by fact. — Bob Sakalas

> [19:53](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1193s) surfing change — Bob Sakalas

## Mentioned Bob Sakalas, SAP, Joule, Ferris Bueller, Doug McMillon, Walmart, Diet Coke, Coke Zero, Jamie Dimon, JPMorgan Chase, Halliburton, Amin Nasser, Saudi Aramco, Mastercard, Hasso Plattner, SAP Sapphire, ServiceNow, Salesforce, Southwest Airlines, Common Domain Model, SAP Business Data Cloud, Databricks, Collibra, AWS, Confluent, DataRobot, Google BigQuery, Snowflake, Microsoft Fabric, Amazon Athena, Heineken, Microsoft Copilot, SAP Datasphere, Generative AI Hub, Shri, Joule Studio, Philippe, Cirque du Soleil, Jessica, SAP AppHouse

## Video description Bob Sakalas - Innovation Strategist, SAP

Are you under pressure to turn AI into real, visible ROI — without adding more sprawl, risk, or complexity? SAP’s Autonomous Enterprise vision points to a new operating model: people set direction, Joule coordinates work, and AI assistants and agents execute defined tasks across business processes in a reliable, responsible way. This session shows how to move beyond chat by choosing the right processes, grounding AI in trusted data, and governing agents at enterprise scale — so you can become the quiet hero who moves AI from hype to progress leadership can see, trust, and scale.

[Source](https://www.youtube.com/watch?v=xuIh5U1l0Vg)

中文简报

SAP将自主AI战略与企业级硬性投资回报挂钩 SAP认为，企业AI必须扎根于可信且具备流程感知能力的数据，而非基础聊天界面，才能产生真实的财务回报并规避运营风险。 Bob Sakalas介绍了SAP推进自主型企业的战略，主张推动AI从单纯的聊天界面转向可执行、以事实为指导的自动化。他着重列举了JPMorgan Chase、Saudi Aramco和Mastercard等大型企业的硬性投资回报案例，并剖析了概念验证部署中的常见陷阱，尤其是数据质量问题。通过将大语言模型与SAP庞大的知识图谱、通用领域模型以及严格的企业治理相结合，SAP助力机构安全地实现复杂工作流的自动化。包括Heineken和Cirque du Soleil在内的实际落地案例表明，具备流程感知能力的AI能够在不增加人员编制的情况下大幅减少运营阻力并推动收入增长。 SAP 系统支撑着全球约 85% 的商业交易；创新战略专家 Bob Sakalas 指出，当前企业级 AI 发展仍处于极早期阶段（相当于环法赛程的第 5 赛段）。 (0:04) 主流企业已展现出显著的 AI 硬投资回报率（ROI），包括摩根大通获得的 20 亿美元直接回报、沙特阿美创造的 40 亿美元运营价值，以及万事达卡防范的 350 亿美元欺诈交易。 (2:01) 底层数据质量与不一致性是企业 AI 概念验证（PoC）走向生产的核心瓶颈；SAP 历时 5 年构建通用领域模型，并通过业务数据云支持与 Databricks 和 Snowflake 等伙伴实现零复制数据共享。 (7:25) 依托内嵌业务流程上下文的知识图谱，SAP AI 代理在发货与产品召回核查中实现了 100% 的准确率，显著优于独立 LLM 约 75% 的准确表现，有效规避核心业务系统中的幻觉风险。 (14:02)

Reading summary

In this presentation, SAP’s Bob Sakalas argues that realizing an autonomous enterprise requires moving beyond conversational AI to deploy process-aware agents grounded in governed business data—a strategy he summarizes as "less chat, more act, guided by fact." Because SAP systems touch roughly 85% of global commerce across finance, supply chain, and human resources, Sakalas contends that integrating AI directly into core systems of record is essential to drive revenue without proportionally increasing headcount. Proven ROI Versus Core Process Adoption (2:01) Sakalas contends that enterprise AI is still in its early stages—comparing it to stage five of the 21-stage Tour de France—yet it is already reshaping every job. He cites Walmart CEO Doug McMillon, who oversees 2.2 million employees and anticipates AI monitoring real-time sales anomalies, such as triggering a shelf check if a store goes three minutes without selling a Diet Coke or Coke Zero. Across other industries, Sakalas points to substantial hard-dollar returns: JPMorgan Chase invested $2 billion in AI and measured a direct $2 billion return, Saudi Aramco CEO Amin Nasser reported over $4 billion in value from automating repetitive oil and gas workflows, and Mastercard’s CEO cited $35 billion in prevented fraud. Despite these benchmarks and widespread personal use of consumer AI, Sakalas notes an informal survey showing only about 32% of enterprise audiences actively use AI within core operational business processes. However, as board members and CFOs experiment with AI on their phones, executive pressure is forcing IT leaders to modernize complex, aging Fortune 1000 technology stacks to fulfills the autonomous enterprise vision first outlined by SAP co-founder Hasso Plattner in 2017. Operational Risks and the Data Bottleneck (7:25) While public AI tools allow teams to spin up a proof-of-concept in two weeks rather than six months, Sakalas warns that uncontrolled deployment introduces severe risks. Unmonitored local data extracts create shadow IT and cybersecurity vulnerabilities, while hallucinations inside systems of record could redirect a cargo ship to the wrong port, ground a functional airplane, or misreport quarterly financials to Wall Street. Above all, Sakalas stresses that proofs-of-concept fail to reach production when underlying data is stale or defined inconsistently across platforms like ServiceNow, Salesforce, and SAP. Comparing raw data without data products to unrefined "crude oil delivered to your backyard," Sakalas explains that SAP spent over five years and 10,000 engineers building a Common Domain Model to standardize semantic context. Through the SAP Business Data Cloud (BDC), organizations can preserve SAP business context without creating redundant copies, supported by zero-copy data sharing with Databricks, governance and lineage via Collibra, and integrations with Confluent, DataRobot, Google BigQuery, Snowflake, Microsoft Fabric, and Amazon Athena. For example, after Heineken found that standalone Microsoft Copilot lacked sufficient data structure and context, the brewer combined SAP Datasphere and the Generative AI Hub to build "Ask Copy," giving employees reliable, context-grounded enterprise insights. Grounding Agents in Enterprise Knowledge Graphs (14:02) To illustrate why general-purpose models struggle with operational execution, Sakalas details SAP’s Knowledge Graph, which maps relationships across 450,000 ABAP tables, 80,000 CDS views, and 7 million data fields. In an internal comparison evaluating a standard LLM agent against an SAP Knowledge Graph-backed agent, the standard LLM achieved roughly 75% accuracy on a delivery status query, whereas the SAP agent reached 100% accuracy at a fraction of a cent in token costs. In a second test involving order fulfillment during an active product recall, the standard LLM cleared the order to ship, while the SAP agent recognized the recall and blocked the shipment. Sakalas attributes this precision to three differentiators, which he terms PDG: Process (50 years of domain expertise across 28 industries), Data (serving as the primary system of record for financial and operational data), and Governance (the auditing and compliance controls required for financial reporting). Scaling Agentic Workflows with Human Oversight (16:22) Because operational risk and insurance constraints prevent immediate, unmonitored automation, SAP designs agentic architectures that keep humans in the loop and scale like a biological system. At the task level, an individual "cell" acts as a specialized assistant, such as for financial closing, which customers can customize using Joule Studio. Combining these task agents under strict auditability forms departmental "organs"—such as Accounts Payable, Accounts Receivable, or Tax workgroups—which then connect across lines of business and interoperate with customer-built or third-party agents from platforms like Salesforce. To identify high-impact starting points, SAP conducts customer workshops focused on specific operational bottlenecks. Sakalas highlights Cirque du Soleil, which sold roughly 10 million tickets in 2024 but relied on a four-person team to process 70,000 multi-language invoices. By partnering with the SAP AppHouse in Palo Alto to deploy a Business AI invoice assistant, the team reduced the average time required to research and resolve a disputed invoice from 30 minutes to 2 minutes. Sakalas concludes that just as SAP navigated previous shifts from mainframes to client-server architectures and the internet, agentic AI aligns directly with its core focus on process execution. By keeping proprietary enterprise data protected while uniting large language models with a governed data foundation, organizations can scale operations without matching headcount growth—though Sakalas notes via a standard disclaimer that specific future product details remain subject to refinement until general availability.

精炼阅读摘要

SAP高管Bob Sakalas在Ai4大会上提出，企业AI的真正价值在于从单纯的文本对话转向以事实为导向的工作流执行，并通过融合底层业务数据、知识图谱与严格治理来实现“自主企业”架构。企业AI面临的投资回报与数据瓶颈 (2:01) 尽管摩根大通、沙特阿美和万事达卡等行业巨头已经公布了数十亿美元规模的实际硬性财务回报，例如沙特阿美通过自动化重复性流程获得逾40亿美元收益，万事达卡借助AI防范了350亿美元的欺诈，但企业在落地核心业务AI时仍面临严重迟滞。Sakalas指出，根据现场非正式调研，仅有约32%的企业受众在核心运营流程中积极应用AI。概念验证（PoC）往往可以在两周内迅速搭建，但若无法跨越底层数据质量的瓶颈，多数项目都无法进入生产环境。由于各系统之间主数据定义不一致或依赖过时历史数据，未受治理的通用大模型存在幻觉风险，这在核心记录系统（System of Record）中可能导致货运误导、飞机误检修或财务虚假汇报等高昂代价。知识图谱与语义上下文消除模型幻觉 (7:25) 为了在不需要多份冗余副本的情况下保持业务上下文，SAP历时5年组织超万名工程师构建了通用领域模型（Common Domain Model），并推出SAP商业数据云（BDC）。该架构依托包含45万张ABAP表、8万个CDS视图和700万个数据字段的知识图谱，将行业专业知识直接内嵌至AI推理流程中。在交付状态查询的实测对比中，普通大模型仅取得约75%的准确率，而结合知识图谱的流程感知智能体达到了100%准确率；在面对召回产品的发货请求时，普通大语言模型误放行，而SAP架构准确识别出召回状态并阻断了发货。喜力啤酒早期仅使用通用Copilot遭遇数据缺乏结构化的问题，随后通过Datasphere和生成式AI Hub构建了“Ask Copy”助手以获得可靠业务洞察。开放生态集成与PDG企业AI准则 (14:02) Sakalas强调企业AI平台不应碎片化，而需遵循PDG准则：深耕28个行业的业务流程（Process）、依托核心系统的权威数据（Data）以及满足财报审计要求的企业级治理（Governance）。在数据共享与管控方面，SAP采取零复制（Zero-Copy）策略，与Databricks在机器学习场景下开展无冗余数据共享，并联合Collibra实现跨AWS、Databricks及遗留系统的端到端数据目录与血缘治理，同时支持BigQuery、Snowflake、Microsoft Fabric和Amazon Athena等主流超大规模云厂商系统。此外，SAP的智能体架构采取开放互通原则，允许客户自定义智能体及Salesforce等第三方平台智能体协同运作。从细胞到机体的分层推进与人机协同 (16:22) 受限于运营风险与合规保险要求，完全自动化的系统无法一蹴而就，因此SAP主张以生物学层级隐喻构建自主企业：单个专业任务助手为“细胞”（如利用Joule Studio定制的财务关账助手），组合成应付、应收或税务工作组等部门级“器官”，最终联动供应链、人力资源与支出管理形成跨业务线的“完整躯体”。在落地案例中，太阳马戏团（Cirque du Soleil）在2024年处理1000万张门票业务时，Jessica带领的4人团队需要应对7万张多语言发票；通过与SAP帕洛阿尔托AppHouse合作开发AI发票助手，单笔争议发票的研究处理时间由30分钟骤降至2分钟，印证了在不增加雇员规模的前提下支持业务扩张的可行性。 Sakalas总结认为，企业实现自主运营的核心路径在于结合大语言模型能力与深层业务流程逻辑，以私有数据安全为红线实现增收而不单纯增加人力；但企业在迈向全面自动化过程中，仍须克服跨系统老旧IT架构升级的复杂性，并维持人机协同以防范操作风险。

Detailed notes — zh

## Background 该内容由行业峰会频道 Ai4 于 2026 年 8 月 21 日在 YouTube 平台发布。视频收录了 SAP 创新策略师 Bob Sakalas 的主题演讲，主要探讨企业如何在不增加系统复杂性与风险的前提下，通过 SAP 的自治企业愿景、可信商业数据与智能体治理将 AI 转化为切实可见的业务投资回报（ROI）。 ## TL;DR 演讲直击企业级 AI 从概念验证（PoC）步入实际生产环境的痛点，强调核心业务记录系统与大模型幻觉绝不能共存。SAP 提出“少闲聊、多行动、以事实为指引”（Less Chat, More Act, Guided by Fact）的核心理念，基于其覆盖全球约 85% 商业交易的沉淀，打造了整合流程、数据与治理的 PDG 支撑体系。通过商业数据云（BDC）与企业知识图谱打通私有业务上下文，SAP 展示了如何借助智能体（Agentic AI）实现从单一业务任务到端到端自治运营的演进，并以太阳马戏团将异常发票处理耗时缩短 90% 以上等案例印证了实际商业价值。 ## Atmosphere & style 整场演讲风格沉稳务实却不失生动幽默，主讲人善于用通俗贴切的隐喻（如环法自行车赛赛段、未经精炼的原油）拆解复杂的企业架构与数据治理难题。现场穿插了自嘲戴错眼镜看小抄等轻松花絮，营造出亲切而真诚的互动氛围，有效中和了企业软件议题的严肃感。整体叙事层层递进、论据扎实，既直面行业落地中的现实挑战，又洋溢着对技术变革浪潮的充沛激情，给予听众兼具战略视野与实战落地信心的听会体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) SAP 系统承载全球约 85% 的商业交易，核心覆盖财务、供应链与人力资源三大企业支柱。 - [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) 企业级 AI 的巨额投资回报已在领军企业中兑现，例如摩根大通测算出 20 亿美元硬性回报，万事达卡成功拦截 350 亿美元欺诈损失。 - [03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) 多数员工仍停留在用 AI 润色邮件等个人辅助层面，企业现有 IT 架构复杂脆弱，亟须向自治企业（Autonomous Enterprise）深度现代化转型。 - [03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) 核心记录系统（System of Record）绝对无法容忍大模型幻觉，供应链误派或财报失真将直接导致千万级损失和法律风险。 - [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) 大量 AI PoC 难以投入生产环境的瓶颈在于数据语义割裂与时效性差，必须依托统一的数据编织与数据产品打破壁垒。 - [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) SAP 不重复造底层基础大模型，而是推出商业数据云（BDC），与 Databricks、Collibra 等伙伴合作实现零复制共享和端到端数据血缘治理。 - [13:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=818s) SAP 构建了融合 45 万张 ABAP 表、8 万 CDS 视图及 700 万字段的企业知识图谱，使智能体业务问答准确率从通用模型的 75% 提升至 100%。 - [13:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=818s) SAP 提出 PDG 三要素架构（流程 Process、数据 Data、治理 Governance），贯彻“Less Chat, More Act, Guided by Fact”的行动准则。 - [16:22](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=982s) 自治企业采用自底向上的生物学演进路径（细胞到器官），当前仍需人机协同（Human-in-the-loop）以满足合规与监管要求。 - [16:22](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=982s) 借助 Joule Studio 等工具，企业可在严密审计跟踪下针对特定业务场景定制智能体，逐步扩展至 AP、AR 及全业务领域的自主运营体系。 - [18:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1084s) 落地企业 AI 应优先通过针对性研讨会锁定高 ROI 业务痛点，以“顶尖模型 + 流程认知 + 私有数据”实现营收规模化且无需扩张人力。 - [19:00](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1140s) 太阳马戏团利用 SAP 商业 AI 发票助手，将 4 人团队处理 7 万张多语种发票的异常排查时间从 30 分钟骤降至 2 分钟，效率提升超 90%。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) 自治企业的愿景与企业级 AI 的真实回报 SAP 作为全球最大的 ERP 厂商，全球约 85% 的商业交易都会触及 SAP 系统，其核心业务领域牢牢扎根于财务、供应链与人力资源。在谈及未来战略时，SAP 指出世界正在以前所未有的速度演进。借用经典电影台词 "Life moves pretty fast. Don't blink. You might be able to miss it"，当下的各行各业工作模式都在发生剧变。以拥有 220 万员工的沃尔玛为例，首席执行官 Doug McMillon 曾公开表示 AI 将改变每一个人的工作方式——从擦地板、理货等日常操作，到系统智能感知销售动态（如发现某种零度可乐在几分钟内没有销售记录，系统就会自动意识到需要去检查货架库存）。如果把当前的 AI 发展比作 21 个赛段的环法自行车赛，目前全行业大约仅处于第 5 赛段，没有人能确切预知 2030 或 2035 年 AI 的终局，但眼前的推进方向已极为清晰。针对外界对 AI 投资回报率（ROI）的疑虑，诸多行业巨头已经拿出了过硬的实际成果。摩根大通（JPMorgan Chase）CEO Jamie Dimon 评价 AI 的变革意义堪比电力发明，他们在投入 20 亿美元后迅速测算出了等额的硬性财务回报；石油巨头沙特阿美（Aramco）CEO Amin Nasser 两年前就指出，AI 在地震数据分析、资产维护等高度重复性工作上为他们创造了至少 40 亿美元的价值；万事达卡（Mastercard）则借助 AI 进行欺诈防御，迄今已成功拦截了高达 350 亿美元的欺诈损失。 - **SAP 商业覆盖规模**：全球约 85% 的商业交易流经 SAP 系统，专注财务、供应链及人力资源三大支柱。 - **摩根大通（JPMorgan）**：投入 20 亿美元用于 AI 建设，迅速实现 20 亿美元等额硬性回报。 - **沙特阿美（Aramco）**：在重复性运维及地震分析场景中实现超 40 亿美元效益。 - **万事达卡（Mastercard）**：依靠 AI 累计预防超过 350 亿美元的欺诈交易。 ### [03:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=218s) 企业现代化的现实挑战与幻觉风险在 SAP Sapphire 大会上，SAP 将“自治企业”（Autonomous Enterprise）确立为核心战略。早在 2017 年，SAP 联合创始人 Hasso Plattner 便在演讲中描绘过这一愿景——每位客户都渴望让 ERP 和 HR 系统变得更容易运行，而如今的技术终于跟上了这一构想。尽管董事会和高管们都在手机上体验过 AI 的魅力，但目前多数企业员工仍仅在个人事务或润色邮件中使用 AI，真正将 AI 深度嵌入业务流程的比例仍处于低位。企业现有的 IT 系统往往由于历史沉淀而错综复杂，甚至处于“靠口香糖和胶带勉强粘合”的状态，现代化改造势在必行。企业级 AI 既蕴含着巨大机遇，也伴随着深重危机。一方面，缺乏管控的 Shadow IT（影子 IT）可能导致员工在本地分散拉取数据、自行搭建报表，带来巨大的数据泄露与安全隐患；另一方面，核心记录系统（System of Record）与大模型幻觉（Hallucinations）绝对无法共存。一旦 AI 出现幻觉将货轮指引到错误港口、为处于盈利状态的客机安排不必要的停飞维护，或向华尔街发布了错误的财务数字，都将带来数以千万计的巨额损失甚至引发法律与人事地震。 ```mermaid flowchart TD accTitle: 企业级 AI 落地挑战与系统风险 subgraph S1["现有企业 IT 困境"] A["复杂老旧架构 多系统割裂"] --> B["影子 IT 泛滥 本地数据失控"] end subgraph S2["幻觉对核心系统的破坏"] C["大模型幻觉 缺乏企业上下文"] --> D["供应链调度错误 货轮送错港口"] C --> E["资产维护误判 客机不必要停飞"] C --> F["财务数据失真 财报披露违规"] end B -.-> C ``` ### [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) 打破 PoC 僵局：数据编织与商业数据云（BDC）打造企业级 AI 方案绝不能靠堆砌不同厂商的离散工具来进一步加剧复杂性，而必须依托统一、安全且可复用的平台。当前阶段，利用公开数据在几周内做一个 PoC（概念验证）演示非常容易，但绝大多数 PoC 难以走入生产环境，其根本瓶颈在于“数据（Data）”。如果企业在 ServiceNow、Salesforce 和 SAP 中对“客户”的定义相互冲突，或者只能获取上周的过时数据（如航空公司无法依靠上周数据处理改签），AI 智能体就无法做出准确决策。SAP 耗时 5 年、投入超过 10,000 名人员构建了通用领域模型（Common Domain Model）以统一业务语义。在底层架构上，SAP 自身不重复研发底层基础大语言模型，而是聚合业界顶尖模型，并推出商业数据云（Business Data Cloud, BDC）。通过将应用层与数据层解耦，BDC 联合 Databricks、Collibra、Confluent 及 DataRobot 等合作伙伴构建数据编织（Data Fabric）： - 引入经过严密定义的**数据产品（Data Products）**，避免数据在下游被反复复制 17 到 18 次； - 与 **Databricks** 合作实现“零复制数据共享”（Zero-copy Data Share），直接赋能机器学习场景； - 与 **Collibra** 合作建立跨越整个企业全局的数据目录与数据血缘（Data Lineage）； - 保持全面开放，先后原生接入 **Google BigQuery**、**Snowflake**、**Microsoft Fabric** 以及 **Amazon Athena** 等主流数据湖与计算引擎。喜力啤酒（Heineken）便是这一架构的典型实践者：他们最初仅使用通用助手（如 Copilot）遇到了底层数据不准确的问题，随后采用 BDC 中的 SAP Datasphere 进行数据整合，结合 Generative AI Hub，成功上线了面向全员的智能数据助手 "Ask Copy"。正如业界名言所比喻的那样：“没有数据产品的数据，就像直接送到你家后院的原油桶”——虽是宝贵资产，但未加精炼便无法直接使用。 ### [13:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=818s) 智能体 AI、知识图谱与三要素架构（PDG） SAP 全面押注智能体 AI（Agentic AI），并致力于实现开放互联，不仅提供原生智能体，还支持接入客户自定义及第三方（如 Salesforce）的智能体。为了保证智能体的绝对可靠，SAP 构建了庞大的企业知识图谱（Knowledge Graph），其背后融合了 450,000 张 ABAP 数据表、80,000 个 CDS 视图以及 700 万个字段的业务元数据与逻辑。在现场演示的对比测试中，普通大模型与挂载 SAP 知识图谱的智能体处理相同业务请求： - **交货状态查询**：通用大模型仅达到约 75% 的准确率，而具备完整业务流程与数据认知的 SAP 智能体准确率达到 100%（在企业级场景中，75% 的准确度等同于不可用）； - **产品召回检测**：SAP 智能体能立刻识别该产品处于召回状态并阻断发货，而通用大模型因缺乏上下文判定“可以发货”。 SAP 将自身在企业级 AI 中的核心优势总结为 **PDG 三要素**，即 **"Less Chat, More Act, Guided by Fact"**： 1. **流程（Process）**：深耕 50 年、覆盖 28 个细分行业的端到端业务流程沉淀； 2. **数据（Data）**：掌握最核心、最权威的业务事实与财务记录系统（System of Record）； 3. **治理（Governance）**：具备面向资本市场与审计机构的严密追踪、审核与合规能力。 ```mermaid flowchart LR accTitle: SAP PDG 业务 AI 支撑体系 subgraph PDG["PDG 核心支柱"] P["流程 Process 28 个行业深耕经验"] D["数据 Data 业务核心记录与真实语义"] G["治理 Governance 可审计与合规保障"] end subgraph Core["核心底座"] KG["SAP 知识图谱 45万 ABAP 表 / 8万 CDS 视图 / 700万字段"] end subgraph Action["业务动作"] Agent["自主智能体 Agentic AI Less Chat, More Act, Guided by Fact"] end KG --> PDG PDG --> Agent ``` ### [16:22](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=982s) 从细胞到器官：自治企业的构建路径在迈向自治企业的演进过程中，“人机协同”（Human-in-the-loop）在相当长的时间内仍是刚需，无论是监管要求还是保险合规，都不允许直接切换到完全无人看管的自动化运作。SAP 采用了一种自底向上的生物学演进视角： - **细胞层（Cell）**：针对特定的业务成果（如财务结账助手）定义细粒度的执行单元与智能体组件； - **可扩展性与定制**：通过 **Joule Studio** 工具，企业（如哈里伯顿或可口可乐）可以根据自身独特的流程规则对智能体进行定制与扩展开发； - **透明度与审计**：全面追踪记录智能体在系统内的每一步操作轨迹； - **器官与系统层（Organ & Body）**：由离散的智能体单元逐步聚合成完整的业务职能域，如应付账款（AP）、应收账款（AR）、税务管理，进而形成贯通财务、支出管理（Spend）、供应链、人力资源（HR）及客户体验（CX）的完整自治运营体系。 ### [18:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1084s) 落地自主企业的路径：数据、知识图谱与业务研讨会将业务从底层逐步重构并释放大量高回报的 Agent（智能代理），核心目标在于交付真正值得信赖的业务成果。这不仅需要融合业界领先的 LLM（大语言模型），更必须结合 SAP 对业务流程的深度理解，以及绝不会泄露至外部环境的企业私有数据。通过知识图谱（Knowledge Graph）与业务数据架构（Business Data Fabric / BDC）的紧密协作，企业能够在不大幅增加人员编制（Headcount）的前提下，实现营收的规模化增长——这正是“做大业务规模，无需同步扩张人力”的底线逻辑。面对庞大的智能化转型蓝图，企业的切入点应当务实且聚焦。最行之有效的方式是开展针对性的业务研讨会（Workshops）：由双方团队共同挑选出一个“只要修复就能直接节省大笔资金”的关键痛点环节，坐下来深入梳理业务细节，并针对该场景引入 SAP 提供的 AI 助手、Agent 以及专属技能库。 ```mermaid flowchart LR accTitle: SAP 业务 AI 赋能价值闭环 subgraph Foundation["底层基础支撑"] A["企业私有数据 (绝对安全不外流)"] B["业务流程理解 (SAP 核心沉淀)"] C["前沿 LLM 模型 (顶尖大模型能力)"] end subgraph Core["核心连接与协同"] D["知识图谱 & 数据底座 (Knowledge Graph / BDC)"] E["Agent & AI Skills (智能代理与技能组合)"] end subgraph Outcome["业务成果交付"] F["营收显著增长 (Scale Revenue)"] G["人力保持精简 (Zero Headcount Sprawl)"] end A --> D B --> D C --> D D --> E E --> F E --> G ``` - **核心公式**：顶尖大模型 + 业务流程认知 + 严密隔离的私有数据 = 值得信赖的业务输出。 - **启动方法论**：通过 Workshop 锁定高 ROI 业务痛点，匹配 Agent 与特定技能组合，小步快跑验证价值。 --- ### [19:00](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1140s) 太阳马戏团实战案例：发票处理从 30 分钟骤降至 2 分钟全球知名娱乐品牌太阳马戏团（Cirque du Soleil）在拉斯维加斯及全球各地取得了巨大的商业成功，仅在 2024 年就售出了约 1000 万张门票。演讲者在此处幽默地插话自嘲道：“顺便说一句，我今天戴错眼镜了。戴错眼镜看这上面的字真的很费劲，我刚才才发现”，引来全场笑声。而在庞大业务光环的背后，财务团队的 Jessica 面临着严峻的运营挑战：她手下仅有一个 4 人的精简小组，却需要处理多达 70,000 张涉及各种不同语言的发票。过去，每当有人对发票提出异议或投诉时，团队都需要投入大量时间跨系统搜集信息和调研排查。为了解决这一痛点，团队前往位于帕洛阿尔托（Palo Alto）的 SAP AppHaus 寻求解决方案，并基于 SAP Business AI 平台构建了一款 AI 驱动的发票助手（AI-enabled Invoice Assistant）。在引入该方案后，处理一张问题发票的平均耗时从过去的 30 分钟大幅压缩至仅需 2 分钟。这项在 SAP Sapphire 大会上由管理层 Philippe 分享的标杆案例，展示了 Agentic AI 如何干脆利落地消除业务流程中的真实痛点。 - **业务规模与压力**：年售票 1000 万张，仅凭 4 人团队管理 70,000 张多语种复杂发票。 - **效率飞跃**：单张异常发票的排查与修复时间从 30 分钟缩短至 2 分钟，效率提升超过 90%。 - **实施路径**：借助 Palo Alto 的 SAP AppHaus 进行场景共创，基于 SAP 商业 AI 平台快速上线落地。 --- ### [19:54](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1194s) 驾驭 Agentic AI 浪潮：迈向自主企业的四大支柱回顾科技演进历程，SAP 一直擅长“在技术变革的浪潮中冲浪”，成功跨越了大型机（Mainframe）、客户机/服务器（Client-Server）以及互联网大爆发等各个技术周期。然而，团队对当下的 Agentic AI 时代展现出了前所未有的兴奋，因为智能代理的核心逻辑与 SAP 沉淀数十年的业务流程管理高度契合。正如经典电影台词所言（Ferris Bueller 曾说）“生活节奏飞快”，没有人能承受在当下止步观望的代价。在演讲的最后，演讲者总结了企业选择 SAP 构建 AI 的核心价值支柱：必须在应用层（Application Level）深度优化流程，夯实高质量的数据基础以确保 AI 的稳定运行，并将商业 AI 无缝嵌入端到端业务流中，从而稳步迈向真正的“自主企业”（Autonomous Enterprise）。 - **四大核心支柱**：流程（Process）、数据（Data）、治理（Governance）与顶尖 AI 技术的紧密结合。 - **关键行动路径**：在应用层优化业务流程，治理企业核心数据，将 AI 注入日常运营全生命周期。 - **资料获取**：演讲幻灯片可通过 prontock.com/ai4 进行下载获取。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=4s) Life moves pretty fast. Don't blink. You might be able to miss it — Bob Sakalas > [07:25](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=445s) 没有数据产品的数据，就像直接送到你家后院的原油桶 — Bob Sakalas > [13:38](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=818s) Less Chat, More Act, Guided by Fact — Bob Sakalas > [18:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1084s) 做大业务规模，无需同步扩张人力 — Bob Sakalas > [18:04](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1084s) 只要修复就能直接节省大笔资金 — Bob Sakalas > [19:00](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1140s) 顺便说一句，我今天戴错眼镜了。戴错眼镜看这上面的字真的很费劲，我刚才才发现 — Bob Sakalas > [19:54](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1194s) 在技术变革的浪潮中冲浪 — Bob Sakalas > [19:54](https://www.youtube.com/watch?v=xuIh5U1l0Vg&t=1194s) 生活节奏飞快 — Bob Sakalas ## Mentioned Bob Sakalas, SAP, Joule, Doug McMillon, Walmart, Jamie Dimon, JPMorgan Chase, Amin Nasser, Aramco, Mastercard, SAP Sapphire, Hasso Plattner, ServiceNow, Salesforce, Business Data Cloud (BDC), Databricks, Collibra, Confluent, DataRobot, Google BigQuery, Snowflake, Microsoft Fabric, Amazon Athena, Heineken, SAP Datasphere, Generative AI Hub, Joule Studio, Halliburton, Coca-Cola, Cirque du Soleil, SAP AppHaus, Ferris Bueller ## Video description Bob Sakalas - Innovation Strategist, SAP Are you under pressure to turn AI into real, visible ROI — without adding more sprawl, risk, or complexity? SAP’s Autonomous Enterprise vision points to a new operating model: people set direction, Joule coordinates work, and AI assistants and agents execute defined tasks across business processes in a reliable, responsible way. This session shows how to move beyond chat by choosing the right processes, grounding AI in trusted data, and governing agents at enterprise scale — so you can become the quiet hero who moves AI from hype to progress leadership can see, trust, and scale. [Source](https://www.youtube.com/watch?v=xuIh5U1l0Vg)

Source quotes

[0:04] Bob Sakalas: Life moves pretty fast. Don't blink. You might be able to miss it [7:25] Bob Sakalas: 没有数据产品的数据，就像直接送到你家后院的原油桶 [13:38] Bob Sakalas: Less Chat, More Act, Guided by Fact [18:04] Bob Sakalas: 做大业务规模，无需同步扩张人力 [18:04] Bob Sakalas: 只要修复就能直接节省大笔资金 [19:00] Bob Sakalas: 顺便说一句，我今天戴错眼镜了。戴错眼镜看这上面的字真的很费劲，我刚才才发现 [19:54] Bob Sakalas: 在技术变革的浪潮中冲浪 [19:54] Bob Sakalas: 生活节奏飞快

Source quotes

[0:04] Bob Sakalas: run like motion pictures [0:04] Bob Sakalas: life moves pretty fast [0:04] Bob Sakalas: stage five out of 21 [3:38] Bob Sakalas: creaky and held together with bubble gum and duct tape [7:25] Bob Sakalas: DATA [12:43] Bob Sakalas: familiar with the brand last night [12:43] Bob Sakalas: Ask Copy [12:43] Bob Sakalas: Data without data products is like crude oil delivered to your backyard. [14:02] Bob Sakalas: Less chat, more act, guided by fact. [19:53] Bob Sakalas: surfing change

Original transcript

[0:04] This is going to go like motion pictures because it's 20 minutes. Uh we're going to talk a little [0:08] bit about autonomous enterprise. First and foremost, is everybody familiar with SAP? [0:14] Everybody in any anybody not familiar with SAP here? There we go. Okay. Well, we're we [0:21] we were the largest ERP company on the planet. We run about 85% of the commerce in the world [0:28] touches an SAP system. So, uh, our main areas of focus are right there. Finance, supply chain, [0:35] human resources. So, now you're all caught up. You know, everything that everybody else knows. [0:39] Okay. So, I'm going to talk a little bit about the future and when I do, SAP always reserves the [0:45] right to change it when until it becomes generally available. So, this is directionally correct, but [0:50] there's always some fine-tuning that can happen. Um, first and foremost, I think we're all in AI, [0:55] so we realize this, right? The world's been moving incredibly fast. As the great philosopher Ferris [1:01] Buer put it, life moves pretty fast. Don't blink. You might be able to miss it. Right. Uh, our jobs [1:08] going to change. What do you guys think? 100%. Right. Doug McMillan's got a front seat to this [1:14] because he employs 2.2 million people. And a lot of those people do manual stuff. They buff floors, [1:22] they stock shelves, they do everything else. But he's on record talking about the fact that [1:26] it's going to change every single person's job. If they don't sell Diet Coke or Coke Zero for a [1:33] few minutes, they realize the system will realize that we better go check the stock on that shelf [1:39] because we know we sell Coke Zero every 3 minutes. It's going to change every single person's job. I [1:45] do think we're at an early phase though. We're probably only in stage five out of 21. If you [1:51] put this in tour to France's perspective, I'm not sure that any of us in this room really know [1:56] where AI is going to be in 2030 or 2035, but it's getting pretty obvious as to where it's [2:01] going right now. A lot of people are talking about ROI and are things paying off. Well, [2:07] there's companies that have actually proven quite a quite a bit of stuff here. Jamie Diamond at JP [2:13] Morgan has talked about it being as transformative as electricity. It's going to change everything. [2:19] They invested $2 billion and they measured a $2 billion hard dollar return right away. I happen [2:27] to have a front seat to this by the way. My eldest daughter works at JP Morgan and I when she comes [2:31] home I ask her questions. You're like, "How much did you use AI today?" And it's mindboggling. I [2:36] use it three hours, Dad. I use it four hours today. So it is absolutely impacting jobs at [2:42] JP Morgan. It's also, you know, working in other industries as well. Your industry, right? You're [2:47] Hallebertton, right? Yeah. Oil and gas has a lot of repetitive work like seismic work, [2:55] like asset maintenance work. This is a mean. He is the CEO of Aramco, the largest oil company on the [3:02] planet. And two years ago, he was talking about the fact that it paid off for at least $4 billion [3:08] for them with a lot of this repetitive work. So yeah, there is RO R ROI out there, but we just [3:14] have to figure out where to find it within our own businesses. Here's a really great example. This [3:18] is the CEO of Mastercard. He's talking about the fact that one of their biggest savings is in fraud [3:25] prevention. They believe AI has contributed to helping them prevent 35 billion with a B [3:31] dollars in fraud so far. So yeah, there's lots of room for AI to make the world more efficient. [3:38] One of the things that we talked about at Sapphire, our biggest show this year, is all [3:42] about the autonomous enterprises, our strategy. Does everybody want to be a more autonomous [3:47] enterprise? What do you think? Yes, absolutely. Was that you? That's my boss. He said, [laughter] [3:56] "Yeah, you know, Hasso Platner, our CEO, was on stage in 2017 talking about the autonomous [4:03] enterprise. It wasn't called that then, but he was talking about the fact that every customer he [4:07] talked to wanted their ERP to be easier to run. They wanted their HR to be easier to run. Now [4:14] the technologies finally caught up. We can get to that point where we do become an autonomous [4:18] enterprise. So a pretty simple question is will AI become part of business processes? I think we all [4:25] know it. But if I how many people used AI as part of their business process this week? You guys are [4:31] an AI crowd. I can tell you I I I talk at a lot of places where it's less than 20%. I think my [4:36] my scientific study there was about 32%. We're all using it for personal reasons. We're all using it [4:42] for rewriting emails. We're not all using as part of process yet. Why is this going to happen? Well, [4:48] I can tell you the board of directors has all played with AI on their phones. They [4:55] are finally feeling it. AI is not that new, but all of a sudden it's accessible to everybody, [5:00] including the CFO. Things change when the CFO is using AI, right? What does that mean for us [5:06] technologists? Well, transformation is necessary. Our systems are kind of creaky sometimes, [5:12] held together with bubble gum and duct tape. There's a lot of complexity out there. We have [5:16] to modernize our systems so they're ready for this future. And when I say that modernization [5:23] isn't easy, this isn't a typical chart for a company. It might actually look too simple. [5:30] There's a lot of systems at the average Fortune 500 or Fortune 1000 company and how do you update [5:35] those is a crucially important topic. So let's take a look at the reality for today's [5:41] enterprises. Enterprise enterprise AI which is the whole purpose of this show is kind of like that [5:48] Chinese symbol for both opportunity and crisis at the same time. There's a lot of opportunity but [5:54] there's also consequences if we get it wrong. Those consequences are a real problem. One of [6:00] them is shadow IT. We've all talked about shadow IT for years. No one's ever put a real positive [6:05] spin on that. Now all of a sudden, everybody's enabled. They all can now create systems and [6:11] put data under their desk like never before. If we're worried about hacking our main database, [6:17] it's different. But if all of a sudden I'm able to pull down all these numbers, create my great new [6:20] report from Matias, and all of a sudden somebody can hack my machine, that's a real problem. So, [6:26] how do we make sure that we inspire people to be creative but at the same time have the governance [6:31] in place to take care of that? The real problem is that hallucinations and systems of record cannot [6:36] coexist. If our system has a hallucination and sends this cargo ship to the wrong port, it is [6:43] a multi-million dollar mistake. Even if you only send a few of those containers at the wrong place, [6:50] it could potentially be a really big deal, right? Well, what if you take this airplane offline for [6:58] maintenance that it didn't need yet? The only way that airplane makes money is when it's in the air. [7:04] And this goes on and on and on. If you have a hallucination that reports the wrong numbers to [7:09] Wall Street, somebody's getting fired, for sure. No doubt about it. And people might be losing [7:15] their jobs because all of a sudden the stock doesn't support it. So, there's huge consequences [7:19] of not doing it right. So, you know, what does it take to leverage AI? Well, in the enterprise, [7:25] first and foremost, please do not add more complexity to the mess. It's already complex [7:32] enough. So, it's important you use a platform and do things the same way. Don't take every single [7:38] problem and create a brand new solution with 12 different vendors saying, "Oh, this is this one, [7:42] and this is this one. This is this one." All of a sudden, it's way more complex than it's ever been. [7:46] You got to use a platform that keeps that core software clean, but at the same time let you [7:50] innovate in a repeatable fashion that that you can trust. The interesting time that we're living in, [7:57] though, is that it's become easier than ever to create a PC or a demo. Every vendor downstairs in [8:06] that hall can show up and do an amazing demo using public data for you. It's really pretty. Claude [8:12] does great things or anti-gravity or whatever you're using. But when you want to actually make [8:18] that an enterprise solution, you suddenly have to apply process. You got to apply data. You probably [8:23] got to apply governance. You got to make sure it's not going to give you those bad consequences [8:27] that result in heads will roll, right? So it's a pretty big deal. So we believe we've got the [8:34] right platform to help you succeed. It's called the business AI platform. It'll help you increase [8:39] momentum. A lot of people have talked about this. Not every PC's making it to production. Guess [8:46] what? Not every PC ever made it to production. That's not new news. This is actually really [8:52] really cool because all of a sudden we can do a PC in a couple of weeks that used to take us six [8:55] months and a lot of money and a lot of people. So we can be a lot more innovative now than [9:00] we've ever been able to be before. Why do a lot of PC's not make it to production? What do you think, [9:09] John? I told you I was going to pick on you. You sat in the middle. Starts with a D, ends at [9:15] an A. Data. It doesn't reflect reality data. No, it's data. The data is the if the data is wrong, [9:23] the AI is going to be wrong 100%. Right. So, a lot of PC's don't knit together the [9:30] right data fabric to support what you're trying to accomplish with those agents. [9:34] So, kind of a rhetorical question to you, but can great business AI happen if your [9:39] data isn't of great quality? What I mean by great quality is if you've got customer defined in your [9:44] Service Now system and your Salesforce system and your SAP system and they're defined differently, [9:49] can you succeed? The answer is no. Stuff's got to work well together. What if you're working on [9:55] last week's data? If you're Southwest Airlines, can you rebook your flight on last week's data? [10:01] Absolutely not. It's got to be current data. or the agents aren't going to do a good job. It's [10:05] got to be customer I mean company specific general available LLMs are not going to rebook that flight [10:12] correctly. Right? We talk a lot about semantics and context and I can tell you it's not easy. It [10:18] took us 5 years and well over 10,000 people to create this common domain model where you can [10:23] have your business all work well together. Our vision is that the autonomous enterprise is going [10:28] to be all these applications that are augmented by AI. So they're smart, they're they're they're [10:33] they quickly react, they do all the wonderful things that we talk about every single day, [10:37] but the platform underneath is how we're going to knit that together. SAP is not creating its [10:42] own large language models. We're using the best of everything that's out there. We're trying to [10:46] make sure that we knit that together with your customer specific data and our knowledge of [10:51] the process so things work well. We've made amazing progress in the last 18 months. The [10:57] way we started looking at the world was that the applications are at the bottom but we split that [11:01] data out. So we needed a data fabric that could then support AI at the top layer. We launched a [11:07] software as a service called business data cloud or BDC. It basically worked we worked with these [11:13] partners data bricks calibbra confluent and data robot to create this this software as a service [11:19] to make sure you get SAP data which is usually the finance stuff the HR stuff the supply chain stuff [11:25] and you get that context so it'll work for you right it introduced the idea of data products so [11:30] they're well described so you can reuse them for multiple use cases we included intelligent apps [11:35] so you could have analytical apps that gave you immediate value it delivers delivers this idea of [11:41] less copies. You know, a lot of companies copy the data 17 18 times downstream trying to get [11:47] that analytical value. We are doing zero copy data share with data bricks so you can use your ML and [11:53] AI use cases and data bricks and still get the best of SAP data. We partnered with Calibra to [11:59] give you a catalog and lineage of the data across your enterprise. So it's not just the data catalog [12:04] that came with Amazon or the data catalog that came with data bricks. you can see across your [12:08] enterprise and see where that data came from. It gave you a great path for BW systems. We [12:13] combined the best of that to create this world for ML and AI. Then the phone rang. Google was [12:20] on the phone. They're saying we want in a lot of our customers using BigQuery for their data [12:25] links. The phone rang again. It was Snowflake. We let them in as well because the new the new [12:29] and improved SAP is much more open than we've ever been in our entire lives. Phone rang again. [12:35] Microsoft Fabric joined and recently we just announced that Amazon is also on with Athena. So a [12:43] good story here is Heineken. Anybody familiar with the brand? I was familiar last night. [laughter] [12:51] They wanted to deliver better information to their users that needed business data. They tried [12:55] to do it with co-pilot alone. They realized the underlying data wasn't perfectly right. [13:00] They used the components that we have in BDC called datasphere to knit that data together. [13:06] They used the generative AI hub to make sure that we included all the AI components to make it easy [13:10] to use. And ultimately they rolled out ask copy. So now everybody's got access to their data much [13:14] faster than ever before. I think this quote is incredible because it's so true, right? People [13:22] have talked about, oh, data is a new oil. Well, you know, data without data products is like [13:26] crude oil delivered to your backyard. Can you do much with that barrel of crude in your backyard? [13:31] But it is one of the most uh important commodities that we have in the entire world. SAP's all in on [13:38] AI and specifically Agentic AI because it's going to help with workflows. We're not going to only [13:44] deliver our own agents. We're going to incorporate all customer written agents as well as everything [13:50] that other vendors do as well. So our agents are going to work with Salesforce's agents. this we we [13:56] see a future that the enterprise just runs much better. How do you encourage reliable AI is an [14:02] incredibly important topic. We've been working for a long time on our knowledge graph. This is not a [14:07] small piece of work. We're talking about 450,000 ABOP tables, 80,000 CDS views, 7 million fields [14:14] to pull this together. A buddy of mine, Shri, that works on my team put together this demo. [14:20] But basically, it this shows the knowledge graph is on the right. the regular LLM's on the left [14:26] and I think I I selected delivery status and it runs that delivery status against both of those [14:32] agents and what you discover is that you know 10 cents of tokens later the SAP version that knows [14:40] the process that knows the data is 100% accurate the other one comes back at about 75% accurate is [14:46] 75% accurate good enough in big business or in small business in any business no Right. It's a [14:55] big deal. The second thing that I ran, I think it was the product recall check. And yes, this was a [15:00] softball. Our system knows that there was a recall on that product. So, yeah, it's going to come back [15:06] 100% accurate, while the other side won't realize there's a recall. It's going to say this order is [15:11] clear to ship. So, there's a a very big difference in understanding the process and understanding the [15:17] data and not doing that. So, why SAP for AI comes down to this. What are the key differences? We are [15:24] 100% aware of your process. We specialize in 28 different industries and we understand how that [15:30] industry works. We get the data right. We we are the system of record for finance and everything [15:36] else. So that's an in crucially important part. But more than anything else, and I think one of [15:40] the keynotes talked about this, governance matters a lot. You've got to be able to audit it if you're [15:45] going to go stand at Wall Street and say this is what our results were. So it kind of comes down [15:49] to this process. SAP's absolute wheelhouse for the last 50 years has been process data. We've got we [15:56] don't have the the vast majority of your data, but we have the data that really matters when it comes [16:00] to finances especially and governance. Yes, we're very very focused on making sure that that system [16:06] of record mentality makes it to your data. Comes down to less chat, more act guided by fact. If [16:13] you want to remember PDG, I think it spells out pretty darn great business AI. This slide might [16:18] have been Heinekenfueled, but we're not going to talk about that. So, how does this become [16:22] the autonomous enterprise? Well, we think the humans going to stay in the loop for a very long [16:27] time. Why is that? Because insurance companies aren't going to say, "Hey, put these agents in [16:32] and just run an entirely automated world." That's not how it works, right? Uh we it starts from a [16:38] groundup approach. That's a human cell that I try to create with AI. But the whole idea is if we're [16:42] trying to create a financial closing assistant, the cell is all the components of that financial [16:47] assistant. And we definitely look at outcomes first. How do we make those outcomes work out? [16:52] Right? We're trying to figure out at each one of these levels. What is the expected value of [16:57] that? We then make these agents extensible by design because we realize that Hallebertton is [17:03] not exactly the same as Coca-Cola. There's going to be differences in each one of these things. [17:07] Juul Studio is our product to help you create those those customizations and they will let you [17:13] create custom agents as well. Now all that needs the transparency to make sure that we can audit [17:19] and figure out what that agent actually did within the enterprise and that's how that cell turns into [17:24] an organ like the lungs. Okay. Now when you start looking at across the rest of finance there's lots [17:29] and lots of components and we're doing the same thing. We have a work group for each one of these [17:32] areas that rolls up the results across accounts payable or accounts receivable or that kind of [17:38] stuff or tax. And then that of course turns into a much larger body of work in finance. And yes, [17:46] we work in a lot of different LLBs. So we're doing the exact same thing. So for example, under spend, [17:51] we're looking at all the components of spend that agents can help in. Same thing's true of [17:55] what we're doing with supply chain. Same thing's true of what we're doing with human resources. [18:00] Same thing's true of what we're doing with uh customer experience. And that's how it all grows [18:04] up from the bottom up into this idea of how we're going to release this many agents for this many [18:10] areas that are all going to have great payback. Our goal is that we give you results that you [18:15] can count on. And that means using the best of LLMs with the best of our understanding of the [18:20] process plus your data that never gets out to the outside world. That's the whole idea. [18:25] That's how that knowledge graph and and BDC work together to get you to the point where [18:29] you can do a lot more revenue without having to add a lot more headcount. That's the bottom line, [18:34] right? Gee, that's a lot. How do you get started? First and foremost, I think it [18:41] comes down to workshops. You pick an area that you know you got to that if we could fix this, [18:45] it would actually save us a lot of money. We sit down with you. We work through that specific area. [18:51] But we're going to have agents that we're offering or assistance that we're offering that include [18:55] agents and skills. Anyway, you know, and then that workshop leads you to a great story like [19:00] this one. Circus has made a heck of a living here in Vegas and everywhere else, right? They actually [19:06] have grown up pretty big. They sold something like 10 million tickets, I believe. Yeah, 10 [19:11] million tickets in 2024. I wore the wrong glasses, by the way. It's really hard to read this when you [19:16] got the wrong glasses on. I just found that out. [laughter] And Jessica had a problem. Jessica had [19:22] a group of four people that had to handle 70,000 invoices in all kinds of different languages. And [19:26] every time somebody complained about an invoice, it took a lot of research, right? They went to [19:30] the app house in PaloAlto. They kind of figured out how do we get this problem solved. They used [19:35] the business AI platform to implement this idea of a AI enabled invoice assistant. The [19:42] average fix on a bad invoice would take them 30 minutes before they got it down to two minutes. [19:47] Phipe talked at Sapphire this year. This is where I snag the story. It's a great story, right? It [19:53] absolutely solved a true pain point for them in a great way. SAP's done a great job of surfing [19:59] change. We went from mainframes to client server to the internet explosion to everything along the [20:05] way. We've never been more excited than we are today because Agentic AI fits exactly what we do. [20:11] I've asked a lot of audiences this. Can you afford to wait? Nobody can afford to wait. The future's [20:16] coming on really, really fast as Ferris Wheeler put it. So, key takeaways from today. First and [20:23] foremost, why SAP for AI? It's the process. It's the data. It's the governance. Pretty darn great [20:28] AI. You got to optimize the processes at the application level. You got to make sure your data [20:32] is right so that your AI can run well. You got to infuse business AI as part of the process if you [20:38] want to succeed. That's how you build up to the autonomous enterprise. If you want to download my [20:42] deck, that link right there, prontock.comi4 is how you do that. And thank you very much. [applause]