A16z: The most difficult enterprise software to use is the biggest opportunity for AI

Why the World Still Runs on SAP
By Eric And Seema Amble, a16z
Translated by Peggy, BlockBeats

Source:

Reprinted from Mars Finance

Editor’s Note: While AI discussions still focus on new products and capabilities, a more structural change is quietly happening at the core of enterprise software. This article is not about how much new applications AI will create, but how it is entering a heavier, yet more real scenario—namely, the enterprise core systems represented by SAP, Salesforce, and ServiceNow.

In simple terms, these three types of systems correspond to different aspects of enterprise operations:

· SAP manages core resources like funds, inventory, and production—it’s the company’s “general ledger.”
· Salesforce handles customer relationships and sales processes—determining how the company generates revenue.
· ServiceNow supports internal workflows and operational systems—enabling organized organizational functioning.

Together, they form the infrastructure of daily enterprise operations.

These systems are extremely critical yet generally difficult to use, complex, and cumbersome. Enterprises layer extensive customization and processes on top of them, making them both repositories of organizational memory and increasingly burdensome to migrate. The more vital the system, the harder it is to change.

This is where AI’s opportunity lies.

Rather than replacing these systems outright, a more practical approach is to build a new layer of action systems on top of them. During implementation, this reduces migration costs; during use, it simplifies operations through co-pilots and proxies; and during expansion, lightweight applications can replace complex customizations. The real change is not in whether the systems are replaced, but in how human-system interactions are being rewritten. AI will not replace SAP, Salesforce, or ServiceNow, but it may make them gradually “disappear” into the background. Meanwhile, new platforms will reconstruct the true value boundaries of enterprise software on this invisible interface layer.

Below is the original text:

As AI develops, the focus of startups and their clients is mostly on new capabilities and the products they enable—such as impressive voice assistants, workflow automation tools, and platforms for text generation applications.

These directions have already emerged and will continue to produce many exciting companies (some of which we have invested in). But AI’s deeper impact may not be in these flashy areas, but rather in a less glamorous, yet more valuable direction: helping organizations better leverage the vast amount of software they already run.

Here’s a question that might sound even a bit offensive, but once you’ve spent a week in a Fortune 500 company, you’ll understand its real significance: Why are people still using SAP (and ServiceNow, Salesforce) today?

The short answer is: SAP and similar large-scale systems record and hold the critical data needed for enterprise operations. But more importantly, enterprises have built extensive customizations, layered complex processes and role divisions on top of these systems—much of which is not even clearly documented. Migrating away from these systems is often costly, lengthy, and painful, typically requiring large consulting teams, years of effort, and hundreds of millions of dollars. For example, upgrading from SAP ECC to SAP S/4HANA can cost around $700 million, take three years, and involve a 50-person team from Accenture. Even after migration, these systems are often used mainly for read-only reports, with limited flexibility.

But this situation is changing.

AI is opening new possibilities, enabling enterprises to upgrade, customize, and replace these systems more efficiently—most importantly, to access and utilize the accumulated data more effectively.

Ultimately, AI’s goal may not be to replace SAP, ServiceNow, or Salesforce, but to make them more programmable and easier to use. The true winners will be platforms that can do two things: first, tap into enterprise digital transformation budgets by quantifiably reducing risk and shortening timelines; second, gradually embed into daily operations, becoming the control hub of work, breaking down traditional bulky interfaces into composable, governable, AI-assisted lightweight applications.

In other words, the core data recording systems will not disappear; what will change are the user interfaces, automation capabilities, and extension layers built on top of them—these are the frontiers of the next software competition.

SAP is hard to use, but we still can’t do without it.

To set the stage, let’s briefly explain what SAP is and what it does. On the surface, these systems are difficult to learn, complex to operate, and costly to modify, making them painful to use. But at the same time, they remain the backbone of operations for large organizations worldwide. Imagine what it feels like to use SAP daily.

However, this so-called “mysterious” complexity itself presents an opportunity.

A less comfortable but more truthful answer is: beneath the heavy interfaces and endless configurations, these systems are incredibly powerful. They carry the core data models of enterprises, define permissions and controls to ensure compliance, embed workflows supporting large-scale operations, and connect dozens or hundreds of downstream processes. They are not consumer internet applications but organizational memories embedded in data tables, role systems, approval workflows, accounting logic, and exception handling.

Replacing such systems is not only expensive but also risky. The more an enterprise customizes—adding fields, processes, pricing rules, reports—the more the system becomes a moat built on switching costs, even a source of competitive advantage. This is why scalability is so important: each enterprise is unique, and change is constant—new regulations, products, organizational structures—these platforms persist because they can be continuously adjusted to fit reality.

But this powerful scalability also makes them fragile. Every customization is a potential risk for future upgrades; each process can evolve into a complex maze; every interface is a continuous drain on users.

This fragility is pervasive. Despite widespread CRM adoption, user satisfaction remains uneven; highly customized ERP systems often lead to project delays and budget overruns; employees are overwhelmed by fragmented workflows, switching between about 1,200 applications weekly—wasting roughly four hours. Nearly half (47%) of digital workers struggle to find the information they need. Large-scale digital transformation projects frequently fail—about 70% do not meet their goals. The costs generated by these frictions are enormous; in 2023, the software implementation and system integration market alone reached approximately $380 billion.

It is within these processes and pain points that AI offers an opportunity to reshape how software is implemented and used. A simple way to understand this opportunity is to look at the enterprise software lifecycle: first implementation or migration, then daily use, and finally, continuous adaptation amid business changes. At each stage, the core task is to convert chaotic human intent into correct, auditable actions recorded in the system.

Next, let’s explore how AI can improve traditional software use at each stage.

Implementation Stage

Starting with implementation, this is the riskiest phase, most sensitive to budget, but also where the returns are clearest. Specifically, it involves transforming scattered research data—meetings, documents, tickets—into structured requirements, and automatically generating the workflows needed for deployment, including process and field mappings, configurations and code, test scripts, cutover plans, migration manuals, and data cleansing and validation before go-live. This process is highly complex and error-prone. For example, Lidl, a German retail giant, spent $500 million on an SAP transformation project but ultimately abandoned it.

Several companies are building tools to assist migration and implementation, such as co-pilot systems and project management tools. Examples include:

· Axiamatic offers an AI-driven ERP assurance layer, building project knowledge graphs that flag potential issues in requirements and change management within Slack or Teams, reducing risk and accelerating S/4HANA projects, integrated with SAP Build and consulting firms like KPMG, EY, and IBM.
· Conduct is a co-pilot tool for code and process mapping, generating semantic and technical documentation during ECC to S/4 migration, supporting Q&A on custom tables and APIs to speed internal takeover.
· Auctor provides agent-based implementation delivery for system integrators and professional services, converting research into structured requirements and managing SOWs, design docs, user stories, configurations, and test plans.
· Supersonik focuses on product enablement, using visual and voice agents within real interfaces to teach and reduce the need for manual solution engineering, supporting channel and customer-driven expansion.
· Tessera builds AI-native system integration capabilities, directly connecting to existing ERP systems, assessing implementation status, and automatically identifying and fixing issues during migration for end-to-end transformation management.

These companies aim to make transformation faster, cheaper, and more controllable—by identifying issues early, reducing delays, converting scattered project data into structured knowledge, and automating mapping, documentation, testing, and training—thus reducing reliance on large system integrators.

We believe this space still has room for more startups, especially those collaborating with existing partners rather than competing. Directions include:

· Implementation agents tied to project outcomes and risks, such as requirements tracking, configuration comparison, switch simulation, code generation, and deviation detection.
· Semantic documentation tools that keep knowledge current and accessible.
· Empowerment agents that turn training and channel promotion into reusable products.

Because startups can effectively alleviate enterprise burdens, they can price their solutions based on the cost savings from delays, directly target CIO and CFO budgets, and replace bloated system integration projects.

Use and Maintenance

After a system is implemented, the real challenge begins. Daily use involves navigating complex, chaotic interfaces. Work often spans dozens of screens, with personnel constantly resetting experience, and many edge processes lacking good product support. Users spend time searching for fields, manually syncing data across systems, or repeatedly requesting reports from operations. This leads to slower workflows, errors, and high training costs.

Here, AI’s opportunity is to build a more friendly, powerful action layer on top of these traditional systems.

These companies aim to help teams extract more value from existing systems. They typically develop co-pilots within Slack or browser sidebars that answer questions like “Where is this data?” or “How do I complete this task?” and can perform safe API actions—creating tickets, entering journal entries, updating vendor terms—when APIs are available. They can also connect multiple systems into cross-application workflows, such as pulling last quarter’s purchase orders from SAP, verifying contracts in Coupa, drafting discrepancy reports in ServiceNow, with manual approvals, audit logs, and fine-grained permissions. Good products track usage, time savings, and error reduction.

However, many critical tasks in enterprises are not exposed via standardized APIs but exist within various interfaces—traditional clients, virtual desktops, or poorly documented management portals. Therefore, modern computer operation agents supplement API-driven co-pilots by automating the last 30-40% of workflows that cannot be called via interfaces.

Their core capability is not just clicking buttons but executing reliably in chaotic environments. These agents must understand interface structures, locate stable elements, recover from layout changes, and record progress at key points for safe interruption and resumption. When combined with validation mechanisms (like difference checks, reconciliation, sandbox testing) and enterprise controls (single sign-on, key management, least privilege, auditing), they can transform manual tasks—such as ticket sorting, end-of-period closing, customer updates, or price adjustments—into manageable, repeatable automation, even in parts of SAP, ServiceNow, or Salesforce not originally designed for automation.

In essence: APIs make standard paths more efficient; computer operation capabilities extend automation to long-tail workflows.

Companies like Factor Labs and Sola have already deployed such agents in production, replacing traditional outsourcing and enabling large organizations to scale task automation.

Extension Layer

Even if you make SAP, ServiceNow, and Salesforce easier to use, enterprises are constantly evolving. New products, policies, mergers, regulations, and long-tail processes that are not worth standalone development push software to adapt to real business conditions. Historically, teams had two options: deep customization—bearing fragility costs—or building scattered standalone apps, which complicate integration, governance, and maintenance.

AI offers a third path: build small, manageable, governable application experiences on top of core systems without disrupting them.

This involves creating a unified data and action plane: reading data from systems via APIs and events (supplemented by secure interface scraping if needed), standardizing it into semantic models—orders, vendors, tickets—and providing permissioned, auditable operation interfaces.

On this foundation, teams can rapidly develop focused applications for specific scenarios—more modern, more aligned with actual needs. For example, instead of multiple steps in SAP for vendor onboarding, provide a single lightweight app that collects data, checks for duplicates, routes for approval, and writes data back to SAP. Instead of toggling between Salesforce interfaces to update renewal terms, offer a spreadsheet-like editor for bulk edits, compliance checks, impact previews, and full audit logs. Or, instead of building new portals repeatedly, provide a unified entry point for frontline teams to perform high-frequency tasks—creating returns, extending credit, initiating secondary tickets, or recording expenses—across systems without switching pages.

These extension layers can also enable cross-system workflows and automation—something no single vendor can fully cover. For example, event-driven automation: when an invoice posts with over 3% discrepancy, generate a note and submit for approval; or when a ticket is reopened twice, automatically create a problem record, assign responsible staff, and update customer status, with manual review at key points.

Over time, the most valuable practices will become reusable intent modules—such as quoting, payment collection, vendor onboarding, or period-end closing—that define not only what to do but how to do it securely and compliantly in specific enterprise contexts.

Products like General Magic’s Cell make building such custom workflows concrete: upload OpenAPI specs to turn each API into an executable operation; embed simple scripts into native command bars to call APIs directly; supported by analytics, multi-tenant architecture, security, and permissions.

What will the end state look like?

We believe most traditional systems will continue to exist, but they will no longer be the primary interfaces for work. ERP, CRM, and ITSM systems are deeply embedded and cannot be replaced at the pace of consumer software; they will evolve slowly and remain as system records. The real change will be in the user-facing action systems built on top: AI will become the default entry point for understanding system operations, executing workflows across systems, and creating lightweight modern apps that bypass traditional interfaces. In other words, the bridge layer will become the main highway.

In this paradigm, long-term winners will not be chatbots but operating systems: a unified data and action plane built on semantic models of business objects, equipped with robust security and governance, enabling AI to reliably operate in production. For end users, there will be no need to learn which interface, field, or transaction code to use; no need to relearn after interface or process changes. Just describe the desired outcome, and the system will handle the rest—asking clarifying questions, showing previews, and executing under approval and audit controls.

For example, you could instruct: “Create a return and notify the customer,” “Open a secondary ticket and fetch the last three related events,” or “Complete vendor onboarding, including data collection, approval, and payment terms.” Today, these require switching between SAP, Salesforce, ServiceNow, and spreadsheets; in the new paradigm, they will be integrated into a seamless execution flow.

This shift results in fewer errors and rollbacks, lower reliance on experience, faster processing cycles, and significantly reduced training costs—since interactions are driven by intent, perceived by roles, and support self-service by default.

The moat will also accumulate through real use: each successful workflow execution becomes a reusable intent; each exception becomes a new security constraint; each migration artifact becomes part of a continuously evolving system narrative; each integration deepens understanding of actual enterprise operations. Over time, this AI layer will become the core interface for understanding change impacts, preventing drift, measuring ROI, and building new workflows—even if the underlying systems remain unchanged.