Abstract: The world will continue to operate on SAP, but AI will reshape it
Original Title: Why the World Still Runs on SAP
Original Author: Eric And Seema Amble, a16z
Translation: Peggy, BlockBeats
Editor's Note: While discussions about AI are still focused on new products and capabilities, a more structural change is quietly happening at the underlying level of enterprise software. This article is concerned not with how many new applications AI will create, but with how it enters a more substantial and real scenario, represented by core enterprise systems like SAP, Salesforce, and ServiceNow.
In simple terms, these three types of systems correspond to different aspects of enterprise operation:
· SAP is responsible for managing core resources such as finance, inventory, and production, serving as the company's "general ledger";
· Salesforce manages customer and sales processes, determining how the company generates revenue;
· ServiceNow supports internal processes and operational systems, allowing the organization to operate in an orderly manner. Together, they constitute the infrastructure for daily operations in enterprises.
On one hand, these systems are extremely critical; on the other hand, they are generally difficult to use, complex, and cumbersome. Enterprises have layered a large amount of customization and processes on top of them, making them not only carriers of organizational memory but also gradually evolving into technological burdens that are difficult to migrate. The more important the system is, the harder it becomes to change.
The opportunity for AI lies here.
Rather than replacing these systems, a more realistic path is to build a new layer of action systems based on them, reducing migration costs during the implementation phase, simplifying operations through co-pilots and agents during the usage phase, and replacing complex customizations with lightweight applications during the expansion phase. Therefore, the real change is not whether the system itself is replaced, but how the interaction between people and systems is being rewritten. AI will not replace SAP, Salesforce, or ServiceNow, but may make them gradually "invisible." New platforms will reconstruct the true value boundaries of enterprise software on these invisible interfaces.
The following is the original text:
As AI develops, the focus of startups and their customers is largely on entirely new capabilities and the products born from them, such as various dazzling voice agents, workflow automation tools, and platforms that generate applications from text.
These directions have indeed emerged and will continue to yield many exciting companies (we have invested in some of them). However, the real, far-reaching impact of AI may not lie in these seemingly cool areas but in a less glamorous, but more valuable direction: helping organizations better utilize the vast amount of software they already have running.
There is a question that sounds somewhat offensive, but if you've spent a week at a Fortune 500 company, you'll understand its practical significance: why are people still using SAP (as well as ServiceNow and Salesforce) to this day?
The short answer is: SAP and similar large systems record and carry the critical data necessary for enterprise operations. But more importantly, enterprises have made a significant amount of customizations on top of these systems, layering complex processes and role divisions, much of which has not even been clearly documented. Migrating away from these systems is often costly, time-consuming, and painful, usually requiring a large consulting team, taking years, and costing hundreds of millions of dollars. For example, upgrading from SAP ECC to SAP S/4HANA can cost $700 million, take 3 years, and involve a team of 50 from Accenture. Even after completing the migration, this software is often only used to generate read-only reports, making flexible operations nearly impossible.
But this situation is changing.
AI is opening a new possibility space, allowing enterprises to upgrade, customize, and replace these systems, and more importantly, to access and use the data embedded within them more efficiently.
Ultimately, the goal of AI may not be to replace SAP/ServiceNow/Salesforce, but to make them more programmable and easier to use. The real winners will be the platforms that can achieve two things: first, penetrate the enterprise digital transformation budgets by quantifiably reducing risks and shortening cycles; second, gradually integrate into daily operations, becoming the control hub for work, disassembling traditional cumbersome interfaces into operations and lightweight applications that are composable, governable, and AI-assisted.
In other words, the record systems themselves will not disappear; the real transformation will occur in the upper interaction interfaces, automation capabilities, and extensibility layer, which is the forefront of software competition in the next phase.
SAP is difficult to use, but we still cannot live without it
To lay the groundwork for this problem, let's briefly explain what SAP is and what it does. On the surface, these systems are difficult to navigate, complex to operate, and costly to modify, which makes them quite painful to use; yet at the same time, they remain the core pillars of operation for large organizations worldwide. Imagine what the daily experience of using SAP would be like.
But this so-called incomprehensibility is precisely where the opportunity lies.
A somewhat uncomfortable but more realistic answer is that beneath those cumbersome interfaces and endless configurations, these systems are extremely powerful. They carry the most core data models of the enterprise, define the permissions and control mechanisms necessary for compliance, embed workflows that support scalable operations, and connect the integration relationships of dozens, or even hundreds, of downstream processes. They are not applications in the sense of consumer internet but are organizational memories condensed in the form of data tables, role systems, approval processes, accounting logic, and anomaly handling.
Replacing such systems is not only expensive but also fraught with risks. The more enterprises invest, such as in customized fields, processes, pricing rules, and reporting logic, the more this system becomes a moat built on switching costs and even becomes part of competitive advantage. This is also why scalability is so important: every enterprise is unique, and changes are ubiquitous, whether from new regulatory requirements, new products, or new organizational structures, these platforms can exist long-term because they can be continuously adjusted to adapt to reality.
But the problem is that this scalability, which makes them powerful, also makes them vulnerable. Every customization is a potential minefield for future upgrades; every process evolves into a complex labyrinth; every interface is a continual drain on users.
This vulnerability is almost ubiquitous. Although CRM has been widely adopted, user satisfaction remains uneven; the high customizability of ERP is almost always associated with project delays and budget overruns. Employees are overwhelmed by disconnected workflows, needing to switch between different applications around 1200 times a day, resulting in an estimated wasting of about 4 hours per week; 47% of digital workers struggle to find the information they need to complete their work. Large digital transformation projects also frequently face setbacks, with estimates suggesting that about 70% fail to achieve their intended goals. The spending around these frictions is immense, with the software implementation and system integration market alone reaching approximately $380 billion in 2023.
It is in these processes and pain points that AI brings the opportunity to reshape the way software is implemented and used. One simple way to understand this opportunity is to look at the lifecycle of enterprise software: first, implementation or migration, then daily usage, and finally continuous accumulation and construction during business changes. At each stage, the essential work is to translate chaotic human intentions into correct actions that are executable and auditable in system records.
Next, we can look at how AI improves the usage of traditional software systems at each stage.
Implementation Phase
Starting with the implementation phase, this is the riskiest, most budget-sensitive part, and also offers the clearest returns. Specifically, it involves converting scattered research information, such as meetings, documents, and work orders, into structured requirements and automatically generating the workflows needed for implementation, including process and field mappings, configurations and code, test scripts, switch plans, migration manuals, and the data cleansing and verification required before going live. This process is extremely complex and prone to errors. The German retail giant Lidl, after investing $500 million, ultimately abandoned its SAP transformation project.
Around this phase, a number of companies are building tools to assist migration and implementation, such as various co-pilot systems and project management tools. Here are some typical examples:
· Axiamatic offers an AI assurance layer for ERP, building project knowledge graphs and alerting potential issues in requirements and change management through Slack or Teams, thus mitigating risks and accelerating S/4HANA project progress, and has integrated with SAP Build, embedding in consulting processes with KPMG, EY, IBM, etc.
· Conduct is a co-pilot tool for code and process mapping that generates a semantic layer and technical documentation during the migration from ECC to S/4, supporting Q&A related to custom tables and APIs, accelerating internal takeover within enterprises.
· Auctor provides agent-based implementation delivery capabilities for system integrators and professional services teams, automatically turning the research process into structured requirements and further serving as the system's records for managing SOW, design documents, user stories, configuration, and testing plans.
· Supersonik focuses on product enablement, teaching within real interfaces through visual and voice agents, reducing the manpower required from solution engineers, and supporting channel- and customer-driven implementation and expansion.
· Tessera builds AI-native system integration capabilities, which can directly connect to existing ERP systems to assess their state of implementation and automatically identify and remedy issues during the migration process, achieving end-to-end transformation management.
The value of these companies lies in making the transformation faster, cheaper, and more controllable. This is reflected in several aspects: detecting problems early in the requirements and change management phase, avoiding later amplification; compressing time cycles, as even a month's delay can incur millions of dollars in costs; transforming scattered project data into structured knowledge, enabling internal teams to take over more quickly; and reducing reliance on large system integration teams through automated mapping, documentation generation, testing, and training.
We believe there is still room for more startups in this area, especially tools that collaborate with existing partners rather than compete against them. Specific directions include:
· Implementation agents that can be tied to project outcomes and risks, for example, for requirements tracking, configuration comparisons, switch simulations, code generation, and deviation detection;
· Semantic documentation tools that ensure knowledge is always up-to-date and easily accessible;
· Empowering agents that turn training and channel promotion into reusable productized capabilities.
As startups can effectively ease the burdens faced by enterprises, they can price according to the savings from delayed costs and directly tap into the transformation budgets that CIOs and CFOs are already investing in, while replacing those cumbersome systems integration projects in the process.
Usage and Maintenance
Next, once a software system has been implemented, the real challenge has just begun. Daily usage means constantly navigating the complicated, chaotic interfaces of these systems. Daily work often spans dozens of interfaces, and staff turnover can continually reset experiential accumulation, with many long-tail ancillary processes lacking good product-level support. Users spend time searching for fields, manually syncing data between different systems, or frequently requesting the operations team to "run this report for me." The result is slower process cycles, frequent errors, and long-standing training costs.
Here, the opportunity for AI lies in building a more user-friendly and powerful action system around these traditional systems.
This type of company focuses on helping teams derive more value from existing systems. In practical terms, it usually exists as a co-pilot within Slack or a browser sidebar, capable of answering questions like where to find certain data or how to complete a certain operation through semantic search, and executing secure actions where APIs are available, such as creating work orders, entering entries, and updating vendor terms. These tools can also link multiple systems to form cross-application composite workflows, such as pulling last quarter's purchase orders from SAP, verifying contract terms in Coupa, drafting a discrepancy explanation in ServiceNow, and integrating human approvals, audit records, and fine-grained permission controls along the process. Excellent products will also track usage, saving time, error rates, and other metrics.
But the reality is that a lot of critical work in enterprises is not exposed through standardized APIs but exists within various interfaces, such as traditional client programs, virtual desktop environments, and ill-documented management backends. Hence, modern computer operation agents have become an important complement to API-driven co-pilots. They extend the reach of automation to the last 30% to 40% of processes that cannot be called through interfaces.
Their core capability is not just clicking buttons but the ability to execute stably in a chaotic environment. These agents need to understand interface structures, locate stable elements, resume execution during pop-ups or layout changes, and log progress at critical nodes to safely recover after interruptions. When these capabilities align with verification mechanisms (such as difference comparisons, reconciliations, and sandbox testing) and enterprises' control means (single sign-on, key management, principle of least privilege, audit mechanisms), they can transform work that originally relied on human completion into governable, repeatable automation processes, such as work order sorting, month-end settlement steps, customer updates, price adjustments, etc., even in parts of SAP, ServiceNow, and Salesforce that were not originally designed for automation.
This can be understood this way: APIs make standard paths more efficient, while computer operation capabilities make long-tail processes automatable.
Companies like Factor Labs and Sola have deployed these agents into production environments, replacing traditional business process outsourcing expenditures and helping large organizations achieve scalable task automation.
Expansion Layer
Finally, even if you make SAP, ServiceNow, and Salesforce easier to use, the enterprise itself continues to evolve, which means that system records must also evolve. New products, new policies, new mergers and acquisitions, new regulatory requirements, and a large number of long-tail processes that will never warrant separate core module development are continuously pushing software to adapt to the real state of business. In the past, teams typically had two choices: either deeply customize the system, bearing the accompanying cost of fragility; or develop scattered standalone applications while facing the challenges of integration, governance, and maintenance.
AI provides a third path: building small, governable application experiences on top of core systems at a faster pace, without compromising them.
Building new tools and automation capabilities on traditional systems can be seen as adding a more "usable" experience layer on top of a set of unfriendly software. The basic model is to first create a unified data and action plane: reading data from system records through APIs and events (supplemented by secure interface scraping if necessary), standardizing it into semantic models of business objects, such as orders, vendors, work orders, etc., and then providing a set of operational interfaces with permission controls, approval mechanisms, and auditing capabilities based on this.
On this basis, teams can quickly build application experiences focused on specific scenarios, making these experiences more modern and closer to actual needs. For example, not making procurement personnel go through a dozen steps in SAP to complete vendor onboarding, but providing a single vendor onboarding lightweight application that collects information, performs duplicate checks, runs approval flows, and finally writes data back to SAP. Or not requiring the revenue operations team to switch back and forth between multiple interfaces in Salesforce to modify renewal terms, but providing a spreadsheet-like high-speed editor that allows for bulk modifications, compliance checks, previews of impacts, and ultimately submitting changes with a complete audit trail. Or, instead of repeatedly building new portal systems, providing a unified operational entry for frontline teams, enabling them to complete daily high-frequency operations across systems, such as creating returns, extending credit limits, initiating secondary fault tickets, accruing expenses, etc., without having to jump between numerous pages.
These expansion layers can also bridge cross-system workflows and automation capabilities, which are difficult for any single vendor to prioritize coverage. For example, implementing automated processes through event-driven action: when an invoice is posted and discrepancies exceed 3%, automatically generating explanations and submitting for approval; or when a work order is reopened twice, automatically creating issue records, assigning responsible persons, and syncing customer statuses, and introducing human review at key nodes.
Over time, the most valuable practices will gradually solidify into reusable intent modules, such as from quote to collection, vendor onboarding, month-end settlements, etc. These modules define not only what to do but also how to complete these operations in a secure, compliant manner within a specific enterprise context.
Products like Cell, launched by General Magic, make the foundational capabilities for building such customized workflows tangible and usable: you can upload OpenAPI specifications, turning every interface into a callable operation; then embed the raw command line through a simple script, directly executing real API calls, supported by analytical capabilities, multi-tenant architecture, secure controls, and permission management mechanisms. Thus, the focus of work shifts from rebuilding an interface to combining appropriate actions and strategies on top of existing, trusted systems.
What Will the Endgame Look Like?
Our judgment is that these traditional systems will mostly continue to exist, but they will no longer be the primary interface where work occurs. ERP, CRM, ITSM, and other systems are deeply embedded in enterprises and cannot be replaced at the pace of ordinary software; they will evolve slowly and continue to exist as system records. What will truly change is the user-facing action systems above them: AI will become the default entry point, used to understand how systems operate, execute workflows between systems, and build lightweight modern applications that bypass traditional interfaces. In other words, the layer that originally served as a bridge will become the real main thoroughfare.
Under this paradigm, software that can thrive long-term will resemble an operating system more than a chatbot: a unified data and action plane built on semantic models of business objects, equipped with comprehensive security and governance mechanisms, enabling AI to run reliably in production environments. For end users, there will no longer be a need to learn which interface to use, which field, which transaction code to learn repeatedly after interface or process changes; they simply need to describe the results they wish to achieve, and the system will help them carry out. Throughout the process, the system will ask for necessary clarifications, display execution previews, and then complete the actions under suitable approval and auditing mechanisms.
For example, you might issue such commands: create a return and notify the customer, create a secondary fault ticket and pull the three most recent related events, or complete the vendor onboarding process, including gathering information, going through approval flows, and setting payment terms. These operations today often require switching back and forth between SAP, Salesforce, ServiceNow, and spreadsheets to accomplish. However, under the new paradigm, they will be integrated into a unified execution process.
The results of this transformation are fewer errors and rollbacks, reduced reliance on experience, faster processing cycles, and significantly lowered training costs because the entire interaction is driven by intent, perceived by roles, and primarily supports self-service completion.
The moat will also accumulate continually in real use: every successful execution of a workflow will solidify into reusable intents; every anomaly handling will transform into new security constraints; every product of the migration process will become a continuously updated system context; every integration will deepen the understanding of how the enterprise truly operates. Over time, this layer of AI will become the core entry point for teams to understand the impact of changes, prevent system drift, measure input-output ratios, and construct new workflows, even if the underlying systems themselves remain unchanged.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
