ERP and AI

March 3, 2025

Enterprise resource planning software is a $300 billion market. SAP alone accounts for roughly a quarter of it. Oracle, Microsoft Dynamics, and Workday split most of the rest. These systems run payroll, manage supply chains, close the books, and enforce compliance across thousands of the world's largest organizations. They are deeply embedded, operationally critical, and, for most companies, the single most expensive software investment on the balance sheet. They are also increasingly misaligned with how enterprises will operate in an agent-driven world.

The conventional wisdom is that ERP incumbents are untouchable. The switching costs are astronomical. The data gravity is immense. The regulatory lock-in is real. All of this is true, and none of it is permanent. Three forces are converging that will fundamentally reshape the ERP landscape: the rise of vertical-specific systems, a structural shift from buy to build, and the emergence of AI agents that change what it means to "use" enterprise software in the first place.

The Monolith Tax

Before examining where ERP is headed, it is worth being precise about why the current model is breaking. SAP and its peers were designed in an era when the primary constraint was data integration. If you could get finance, procurement, manufacturing, HR, and sales onto one system with one data model, you could eliminate reconciliation overhead, enforce process consistency, and generate consolidated reporting. This was a genuine engineering insight, and it created enormous value for decades.

The cost of that integration was complexity. SAP S/4HANA contains over 25,000 database tables. A typical enterprise implementation involves configuring thousands of custom transaction codes, authorization roles, and posting rules. The implementation itself takes 18 to 36 months for midsize companies, and three to five years for large enterprises. Total cost of ownership, including license, implementation, customization, and ongoing maintenance, routinely exceeds $100 million. For the largest deployments, it can reach $500 million to $1 billion.

This complexity is not incidental. It is structural. ERP systems must model the full operational surface area of a business. Every industry has different requirements for inventory valuation, revenue recognition, regulatory reporting, and process workflows. A monolithic ERP handles this by parameterizing everything: configuration tables, condition records, enhancement points, custom exits. The system is general enough to model any business, but the generality means that modeling your specific business requires enormous implementation effort.

The result is a paradox. The system that was supposed to simplify enterprise operations has itself become the most complex system in the enterprise. Companies employ armies of consultants, not to build new capabilities, but to maintain the configuration of existing ones. The big four consulting firms generate tens of billions annually from ERP implementation and maintenance work. That is the monolith tax: the ongoing cost of operating a system whose complexity has outgrown its original value proposition.

Headwind One: Vertical-Specific ERP Is Real

The first crack in the monolith is vertical specialization. For decades, the ERP value proposition was "one system for everything." But the operational reality is that a pharmaceutical manufacturer, a construction firm, and a retail chain have fundamentally different process architectures. SAP handles this through industry solutions, preconfigured templates layered on top of the general platform. But industry solutions are a compromise. They constrain the general system to approximate industry-specific workflows rather than modeling those workflows natively.

A new generation of vertical-specific ERP systems is taking a different approach. Instead of starting with a general-purpose platform and narrowing it, they start with the domain model and build outward. Procore owns construction project management. Veeva owns life sciences CRM and regulatory. Toast owns restaurant operations. These are not lightweight SaaS tools. They are full operational platforms with financial controls, compliance workflows, and reporting, built for a specific industry from the ground up.

The structural advantage of vertical ERP is that the domain model is the system, not a configuration layer on top of a generic data model. A construction ERP that natively understands change orders, retainage, AIA billing formats, and lien waivers does not need thousands of configuration tables to approximate these concepts. They are first-class entities. This means faster implementation, lower TCO, and workflows that match how the industry actually operates rather than how SAP thinks it should.

The same dynamic is playing out geographically. Regional ERP providers in Southeast Asia, Latin America, the Middle East, and Africa are building systems that natively handle local tax regimes, regulatory formats, and business practices. SAP's localization modules exist, but they are afterthoughts, maintained by small teams and perpetually trailing behind regulatory changes. A local provider that builds for Brazilian nota fiscal or Indian GST compliance from day one will always have a structural edge in those markets.

Headwind Two: The Build vs. Buy Calculus Is Shifting

The second headwind is subtler but more consequential. Enterprises are increasingly building core operational software in-house rather than buying it from ERP vendors.

This is not a new impulse. Large enterprises have always built custom systems for their most differentiated processes. What has changed is the economics. Cloud infrastructure commoditized compute and storage. Open-source frameworks (React, Next.js, FastAPI, Kafka, Postgres) commoditized the application stack. And now, AI-assisted coding is compressing development timelines by 30-50% for well-scoped systems. The cost of building a purpose-built operational system has dropped by an order of magnitude over the past decade.

Meanwhile, the cost of configuring a general-purpose ERP to match your exact requirements has not dropped. If anything, it has increased as systems have grown more complex. The crossover point, where building a tailored system costs less than configuring a generic one, has moved sharply in favor of building.

Consider a mid-market logistics company. Their core operational loop involves order intake, route optimization, driver dispatch, proof of delivery, invoicing, and settlement. They could implement SAP Transportation Management, a 12-month project requiring specialized consultants and significant customization. Or they could build a purpose-built system on modern infrastructure that models their exact workflow, integrates with their specific carrier APIs, and handles their particular billing structures. Five years ago, the second option was viable only for companies with large engineering teams. Today, a team of ten engineers with AI tooling can build and ship it in four months.

This does not mean enterprises will build everything. General ledger, accounts payable, payroll, and tax compliance are commodity functions with low differentiation potential. No company gains competitive advantage from a novel approach to double-entry bookkeeping. But the operational processes that define how a company competes, its supply chain logic, its pricing algorithms, its fulfillment workflows, are increasingly too important and too specific to delegate to a generic ERP module.

Headwind Three: AI Agents Decouple the Interface from the System

The third headwind is the most disruptive, and the one the ERP industry is least prepared for. AI agents are about to fundamentally change the relationship between users and enterprise software.

Today, ERP usage is a human-interface problem. A procurement manager logs into SAP, navigates to transaction code ME21N, fills in a purchase order header, adds line items, checks pricing conditions, and releases the order. This interaction model, human navigates complex UI to execute a structured transaction, has been unchanged for 30 years. SAP Fiori modernized the visual layer but did not change the fundamental interaction pattern. The user still needs to know which screens to visit, which fields to fill, and which sequence to follow.

AI agents collapse this entire interaction model. An agent that can interpret a natural language instruction ("Create a purchase order for 500 units of part X from supplier Y at the contracted price, routed to our Dallas warehouse"), resolve it against the ERP's data model, execute the necessary transactions, and handle exceptions does not need a human-navigable UI at all. The agent needs an API. Or, in the absence of an API, it needs the ability to operate the existing UI programmatically, the same way a human would, but faster, more accurately, and without training.

This is not hypothetical. The building blocks are production-ready. Large language models with function calling can decompose complex business instructions into structured API calls. Tool-use frameworks enable models to interact with external systems in reasoning loops. Computer-use capabilities allow agents to operate legacy UIs through screen interpretation and programmatic interaction. The orchestration layer, planning a multi-step business process, executing each step, handling errors, and escalating ambiguous cases to humans, is a solved problem at the framework level. What remains is the domain-specific integration work: teaching agents the semantics of specific ERP transactions and business rules.

The implication for ERP vendors is profound. If the primary user of ERP software becomes an AI agent rather than a human, then the competitive dimensions change completely. UI quality becomes irrelevant. The factors that matter are API coverage, data model quality, transaction semantics, and integration surface area. SAP's enormous investment in Fiori, its modern UX layer, becomes a stranded asset in a world where the primary consumer of the system is not a person but an agent.

ERP Does Not Disappear. It Becomes Infrastructure.

None of this means ERP goes away. Enterprises will still need systems of record for financial transactions, regulatory compliance, and operational data. The general ledger is not going to be replaced by a chatbot. Audit trails, posting rules, and tax determination logic are not amenable to AI improvisation. These are rule-based, deterministic processes that belong in structured, governed systems.

What changes is the role ERP plays in the enterprise architecture stack. Today, ERP is the primary interface for operational work. Employees spend hours daily inside SAP, Oracle, or Dynamics, executing transactions, running reports, and managing exceptions. In an agent-driven architecture, ERP becomes a transaction substrate. It still stores the data. It still enforces the business rules. It still produces the regulatory outputs. But humans rarely interact with it directly.

The analogy is databases. PostgreSQL is critical infrastructure. Every production application depends on it. But application developers do not spend their day writing raw SQL against production tables. They interact through application layers, ORMs, APIs, and now AI-generated queries. The database moved from a primary interface to background infrastructure without losing any of its importance. ERP is on the same trajectory.

In this architecture, the agent layer becomes the primary interface for operational work. Users describe intentions in natural language. Agents decompose those intentions into transactions, execute them against the underlying ERP, handle exceptions, and report results. The ERP is still doing the work. The agent is doing the translation. The human is doing the thinking.

The SAP Problem

SAP's strategic position in this transition is weaker than the market appreciates. The company has three structural vulnerabilities that AI-driven disruption will expose.

The complexity moat cuts both ways.SAP's stickiness has always been a function of its complexity. Migrating off SAP is so painful that most enterprises choose to stay. But that same complexity makes SAP the hardest system for AI agents to operate. The 25,000 tables, thousands of transaction codes, and deeply nested configuration hierarchies create a massive integration surface that agents must navigate. A simpler, more API-native ERP system is structurally easier for agents to use. As agent-driven operation becomes the norm, SAP's complexity stops being a moat and starts being a liability.

The consulting ecosystem is a tax, not a moat. The system integration industry around SAP, Accenture, Deloitte, IBM, Infosys, exists because SAP is difficult to implement and maintain. This ecosystem is often cited as a competitive advantage: the partner network as a distribution and support channel. But it is also a massive value drain. For every dollar an enterprise spends on SAP licenses, it spends three to five dollars on consulting. When AI agents can configure, operate, and maintain ERP systems with dramatically less human intervention, the consulting tax shrinks. And with it, one of the primary mechanisms through which SAP maintains its installed base.

SAP's AI strategy is defensive, not transformative. SAP's AI investments, Joule (its copilot), embedded AI features in S/4HANA, and partnerships with hyperscalers, are oriented around making SAP easier to use. This is the right short-term move, but it fundamentally misreads the long-term dynamic. The question is not "how do we make SAP easier to use?" The question is "what happens when humans rarely use SAP directly at all?" SAP is optimizing for a human-interface paradigm that agents are about to render obsolete.

The New Enterprise Stack

The enterprise software stack is being redrawn. At the bottom, systems of record, ERP, HRIS, CRM, handle transactions, compliance, and persistence. These become infrastructure: critical, governed, and largely invisible to end users. In the middle, an integration and orchestration layer normalizes data and actions across systems. This is where semantic models, API gateways, and event buses live. At the top, AI agents serve as the primary interface for operational work, interpreting intent, planning actions, executing workflows, and handling exceptions.

This architecture has several important properties:

• ERP becomes swappable. If the agent layer abstracts the system of record, then the specific ERP underneath matters less. Enterprises can migrate incrementally, replacing SAP modules with vertical-specific systems or in-house alternatives, without changing the agent-facing interface. The lock-in transfers from the ERP to the agent-and-integration layer.
• The value shifts to the orchestration layer. The most strategic software in the enterprise is no longer the system that stores transactions but the system that understands operational intent and executes it across multiple backends. Whoever owns this layer captures the strategic relationship with the enterprise.
• Operational knowledge becomes a first-class asset. When agents execute business processes, they generate structured logs of decisions, exceptions, and outcomes. This operational telemetry is far richer than what ERP systems capture today. It creates a feedback loop: agents learn which actions succeed, which fail, and which require human judgment, continuously improving their operational model.
• Implementation cost drops dramatically. The most expensive part of ERP deployment is configuring the system to match business processes and training users to operate it. If agents handle both configuration (translating business rules into system parameters) and operation (executing transactions on behalf of users), the human cost of ERP falls by an order of magnitude.

Where This Goes

The ERP market will not collapse overnight. SAP will continue to print billions in maintenance revenue from its installed base. Oracle will continue to win cloud ERP deals in finance-heavy enterprises. The transition will be gradual, driven by attrition rather than replacement.

But the trajectory is clear. New deployments will increasingly favor vertical-specific systems over general-purpose monoliths. Enterprises with strong engineering teams will build their differentiating operational software in-house. AI agents will progressively take over the operational interface, reducing the importance of ERP usability and increasing the importance of ERP programmability. And the strategic center of gravity in enterprise software will shift from the system of record to the orchestration layer that sits above it.

For enterprise technology leaders, the implications are actionable now. Invest in API-first ERP architectures, whether that means pushing your current vendor for better APIs or selecting new systems with agent integration as a first-class requirement. Build the semantic layer that normalizes your operational data across systems. Start piloting AI agents on well-defined, high-volume operational workflows, not as chatbots stapled to the ERP sidebar, but as autonomous operators that can execute end-to-end business processes with human oversight at decision points.

ERP is not dying. It is being demoted. From the center of enterprise operations to the plumbing underneath. From the application you train people to use to the infrastructure your agents call. The companies that recognize this shift early will build the orchestration and agent layers that become the new strategic platform. The ones that keep investing in making their ERP easier for humans to click through will find, in a few years, that the humans are no longer clicking.