a16z: Is the software industry losing its direction under AI intelligent agents?

Author: Seema Amble, a16z Partner; Source: a16z; Translation: Shaw, Golden Finance

Is the software industry losing its way?

Last month, Salesforce announced the opening of its application programming interface (API) and launched headless products. This is essentially a bet: in the era of intelligent agents, the core value lies in the data layer, not the user interface (UI). It’s a strategic repositioning. However, it’s worth noting that on the technical level, there isn’t much change: the APIs that Salesforce now promotes as “headless products” have actually existed for many years. In other words, this is just a typical marketing release by Salesforce. The core logic of this new product is: AI agents can directly retrieve data from record systems without needing to adapt to user interfaces designed for humans to follow workflows.

This release also raises a deeper question: If we strip away the user interface and directly open up the underlying database, what is the true core value left? How does this compare to a PostgreSQL database with well-designed data tables and a set of APIs? Do the classic barriers that have supported the longevity of traditional record systems still hold? Or has the industry formed a new set of evaluation standards? In the SaaS era, record systems have a competitive moat because users heavily rely on interfaces to perform their work. In the era of intelligent agents, this advantage is weakening. The value layers of enterprise moats are beginning to sink into data models, permission systems, workflow logic, and compliance capabilities, while extending upward into ecosystems, proprietary data generation, and real-world business execution.

As software moves toward a headless era, where will the true competitive barriers shift to?

User interfaces are themselves products

Record systems are authoritative sources of real data in specific business domains. Customer relationships, employee records, and financial transactions are stored here, serving as the official versions, and are also the core carriers for other tools to read and write data.

Customer Relationship Management (CRM) systems are the record systems for revenue-generating activities; Human Resources Information Systems (HRIS) are the record systems for personnel management; Enterprise Resource Planning (ERP) systems are the record systems for financial and capital data.

The strength of these systems lies not only in data storage but also in becoming the unified factual basis relied upon for the entire enterprise’s coordination and operation.

Over the past twenty years, Salesforce has essentially sold a way of managing sales teams. Dashboards, sales funnel views, performance forecasting tools, dynamic information flows—these are what users truly pay for. Its business model is built on selling user seat licenses, allowing users to access these features. The underlying database is important, but from a business value perspective, it has become an auxiliary configuration.

This means: user interfaces determine user stickiness. They standardize data entry, establish a unified business terminology system: sales leads, opportunities, customer accounts. They push thousands of salespeople to proactively record data that might otherwise be untracked. The user interface has always been the core mechanism for maintaining data consistency and order.

This product has extremely high user stickiness; many sales executives continue to use Salesforce after switching to new companies, not because of superior interface experience, but due to muscle memory and ingrained habits.

Today, AI agents are disrupting this traditional model. Agents can read and write directly to the underlying data without going through front-end interfaces. This has spawned a large number of new tools and alternatives that bypass the original interfaces (Salesforce is not an exception: we recently analyzed how SAP’s ecosystem is rapidly evolving to support AI-compatible systems).

In the long run, computer operation agents will gradually erode traditional human-centered factors: personal preferences, job training, tacit business background knowledge, etc. In other words, the long-term viability of record systems depends on evolving core conditions.

Traditional Moats of Data Retention

Before exploring the changes brought by the era of intelligent agents, it’s necessary to clarify: What originally built the high user stickiness of record systems? The primary factors are centered on human-software interaction methods and individual usage habits. High stickiness is largely created by user interfaces, usage habits, manual workflows, and embedded business processes.

Frequency of use

CRM and similar systems are used daily by sales teams and related personnel. High-frequency use makes them critical infrastructure for enterprises; meanwhile, the cultural and management habits—fixed workflows, muscle memory, long-established management rhythms—are often the hardest to migrate, and enterprises may not even realize that these systems themselves need to be migrated or reconstructed.

Read-only or read-write

A highly sticky record system must be a two-way read-write system. Take CRM as an example: it’s not just an archive for storing data; it’s continuously and frequently read. Every call record, opportunity stage update, task creation—these are actively entered by relevant personnel. This bidirectional data flow means any replacement system must handle real-time operational data, not just historical archives. Enterprises cannot find a seamless switch-over window, so once deployed, they tend to be long-term locked into the original vendor. In contrast, applicant tracking systems (ATS) are mostly write-only or read-seldom systems: after hiring is completed, there’s little need to revisit the data, so user stickiness is naturally weaker.

How many undocumented Standard Operating Procedures (SOPs)?

Implicit knowledge critical to business is never documented in knowledge bases but is embedded in the long-standing workflow rules built by administrators and system integrators. For example, in sales scenarios, unwritten rules include: enterprise orders over $100k require VP approval; orders in EMEA must undergo privacy compliance review; strategic key account discounts can only be approved at quarter-end without financial approval. These implicit rules often directly determine whether business can be executed on time, whether key compliance standards are met, or whether the deal can be closed. To migrate such systems, one must either reverse-engineer every automation logic or risk losing organizational experience accumulated over years.

Are there many internal and external dependencies?

The core criterion is: how many internal systems, team workflows, and external parties depend on this record system? Internal connectivity refers to downstream software and workflows; external connectivity involves auditors, accountants, regulators—e.g., ERP systems often need to grant data access directly to such external entities. The higher the internal and external connectivity, the more complex the related links that need to be untangled during migration.

From a compliance perspective, how critical is the data?

The key question is: does this system qualify as a compliance-critical system? Payroll, ERP, HR data—these are systems that must have authoritative, legally compliant data sources, strict admin controls, and require direct involvement of regulators during migration and audits. This greatly increases user stickiness and migration difficulty. Conversely, sales data, Zendesk, and other customer service tools are less critical: enterprises value business continuity and context but data migration or permission changes do not pose regulatory risks.

Not all record systems have the same migration cost. Comparing CRM and ATS under the same standards, the gap is clear. ATS only serves a closed, fixed recruitment process: after a candidate joins or is eliminated, the related records are mostly a one-time entry with minimal subsequent changes; the scope of system integration is narrow, and the user base is small and focused.

ERP, on the other hand, is at the other extreme: accounting ledgers are audit trail evidence. System migration involves auditors, regulators, and external stakeholders—making replacement much more complex. Replacing ATS is troublesome but manageable; replacing CRM is akin to open-heart surgery; replacing ERP is like running a marathon while undergoing open-chest surgery.

Historically, core record systems have not leveraged proprietary data or network effects to build moats; they rely mainly on business workflows to establish barriers. In consumer-facing businesses, tools and ecosystems are often integrated, but record systems have traditionally not achieved this.

Proprietary Data

Although most record systems collect customer data, they rarely deeply activate this data (and often contractual restrictions limit their use). Therefore, even if CRM contains rich data and could aggregate cross-customer insights, it has yet to produce practically valuable applications (despite some attempts, such as Salesforce’s Einstein AI products).

Network Effects

Network effects are the ultimate barrier many industries dream of. If a system can generate network effects, the value of CRM increases as the ecosystem grows, enabling software vendors to precisely match buyer resources on the platform. But similar to the current state of data value, record systems’ network effects have always been weak.

What remains when user interfaces fade and AI agents take center stage?

AI agents do not need browsers; they only need APIs, business context, execution commands, and action capabilities. Two major technological conditions make large-scale deployment possible: first, large models now possess sufficient logical reasoning ability. Today’s agents can autonomously understand business context, formulate execution plans, call tools, perform operations, and review results—most tasks no longer require human intervention. Second, the MCP protocol standardizes tool invocation, providing a unified interface for external capabilities. Agents integrated with MCP can perform all human operations within milliseconds, operate at scale, and do so entirely without browsers. As long as they have complete business context, even computer operation agents can directly adapt to and control traditional software interfaces without relying on official APIs.

In simple terms, today’s software buyers have three options:

1. Use existing systems + add AI agents. Relying on the command line and API capabilities of existing systems, they can use vendor-native AI products (like Salesforce’s Agentforce, SAP’s Joule) or develop their own intelligent agents based on current systems. (This is an idealized assumption: APIs are fully available, headless transformation is operationally simple, which is often not the case in reality.)

2. Build a completely custom record system. From scratch, develop proprietary data models, business logic, permission systems, audit trails, and system integrations, along with custom AI agents (often leveraging third-party tools for AI and databases).

3. Purchase native AI replacement software. Choose next-generation software designed from the ground up for the AI agent era: built with machine-readable design, with agent orchestration as a core native feature, often in a headless architecture.

So, what other standards for system stickiness still hold? Factors relying on human behavior and habits—such as usage frequency and read-write attributes—are gradually diminishing. AI agents may erode muscle-memory-based moats, but cannot replace the operational logic and contextual barriers formed by business processes. On the contrary, the importance of these logical structures is increasing—because agents need clear rules, permissions, and process definitions to operate safely and compliantly.

In the short term, undocumented implicit SOPs remain critical. The native business logic embedded in workflows is essential for AI agents to execute business correctly and is the most difficult part to reverse-engineer. Currently, this implicit knowledge cannot be fully and cleanly exported, especially when human involvement remains in the process. However, digitization of business context is becoming easier; as agents gradually replace more manual steps, the importance of implicit experiential knowledge will decline.

Disentangling system dependencies remains challenging and has broader implications. The core logic of connectivity has shifted: no longer just about adapting to human operations, but about integrating previously siloed business functions and software systems. CRM agents need to connect sales, billing, customer success data, and workflows. If your platform becomes a transaction hub for multiple external entities (buyers, sellers, partners), dependency chains deepen. Whether adding AI to traditional vendors or building a proprietary database and AI ecosystem, coordinating underlying capabilities across diverse software remains highly difficult.

The importance of compliance-sensitive data remains unchanged. Data subject to regulatory requirements and legal risks must have a single, trusted source. As long as enterprises trust existing products, their willingness to replace diminishes. For example, payroll and financial data—if AI agents need to access them—are rarely self-developed and maintained by companies due to compliance concerns. In the era of full AI integration, a major unresolved issue is: which AI agents are authorized to act on behalf of whom, perform which operations, and how to ensure full auditability? If a record system can serve as the identity and permission hub for AI interactions, it will establish an almost irreplaceable structural position—its moat lies not in stored data but in the trust and permission governance it creates.

Looking ahead, the core factors that determine the competitive moat of native AI startups are becoming clearer.

How difficult is it to rebuild a core record system?

The importance of data manifests on multiple levels. In the short term, it depends on how easily existing record system data can be extracted and replicated. AI tools have significantly lowered the barriers to data migration and replication. In the near term, traditional vendors may deliberately raise migration barriers: designing APIs that are cumbersome, restrictive, feature-limited, or priced unattractively, or not opening APIs at all. But as data extraction tools and especially computer operation AI capabilities continue to improve, the barriers to replication will further decrease. Meanwhile, emerging vendors are capturing richer native data dimensions from emails, call recordings, internal documents, etc. AI reduces the cost of replicating the core 80% of a record system’s basic functions; the remaining 20%, involving special exception workflows, approval rules, compliance requirements, and edge-case scenarios, is the key to distinguishing “partial substitute modules” from “truly complete replacements.”

Do you possess truly exclusive, proprietary data?

Second, the strategic value of data is increasingly prominent. The real moat isn’t external data imported into the system but the unique data generated by the product’s own business processes. The so-called data moat refers to proprietary, regulated, and continuously evolving data assets. Software providers investing in authoritative, comprehensive datasets have a natural advantage over generalist vendors or competitors. Another dimension of data moat is internal business behavior data: top companies no longer just store external input data but deepen their participation in business processes to generate new data assets—such as user behavior observations, response rates, timing patterns, workflow outcomes, industry benchmarks, anomaly detection, and agent operation traces. The core logic has shifted: data itself is now business context.

Do you control the business execution layer?

In traditional models, storing business records was enough to create value; in the new era, where agents lead decision-making, the moat shifts toward products capable of forming closed business loops: initiating actions, capturing results, and using feedback to optimize subsequent decisions—forming a self-sustaining cycle. For example, ERP systems that automate expense approvals, trigger payroll, reconcile invoices, and send business notifications create a deeper moat. Products that embed business processes and generate unique data, becoming more intelligent with use, are more likely to cause business process disruptions if replaced. The richer the business context they cover and the more edge cases they handle, the higher their value.

Can they execute in the real world?

Business models that connect to offline operations—such as logistics, field services, or physical asset management—can form unique moats. For example, companies like DoorDash, which have extensive offline networks, are not traditional record systems but are highly valuable references. More broadly, any software that links service delivery, fulfillment, logistics, field operations, and payments has a differentiation that pure SaaS cannot match. These companies do more than store records and give advice—they dispatch personnel, allocate resources, and complete physical service delivery.

For entrepreneurs, this presents huge opportunities: in sectors where software can make autonomous decisions and coordinate AI-driven scheduling, but the last mile still requires offline execution, there is vast potential. For example, vertical SaaS companies specializing in field services.

Can they generate network effects?

Most traditional record systems have weak network effects because they mainly serve internal enterprise processes. But in the AI agent era, if systems are deeply embedded in multi-party workflows, network effects will become much more significant. If a platform becomes a hub for multi-party business interactions—between buyers and sellers, employers and employees, enterprises and auditors, vendors and customers, payers and service providers—each new participant increases the overall value for others.

Network effects mainly manifest in three ways:

• Shared workflows and collaboration: becoming a unified platform for business supply-demand transactions, synchronized context, and exception handling;

• Industry benchmarks and intelligent insights: leveraging behavior patterns across the network to produce industry standards, anomaly alerts, and optimization suggestions, complementing proprietary data moats;

• Trust and standardization infrastructure: once all parties rely on the same platform for approvals, handoffs, compliance checks, and payments, the product evolves into an industry-wide collaborative infrastructure, creating a high barrier to entry.

How strong is the buyer’s own technical capability?

In theory, any enterprise can develop its own AI agents, but practical implementation varies greatly. Especially in vertical industries and traditional departments, most lack dedicated engineering teams capable of building, maintaining, and iterating proprietary databases, workflows, agent architectures, and governance systems over the long term. Cost is also a key factor: self-development may seem to save licensing fees but shifts costs to implementation, maintenance, and internal complexity. This reveals a clear market opportunity: industries with complex workflows but insufficient technical service providers—such as manufacturing, construction, industrial maintenance, field services, and finance & auditing.

There are also essential foundational conditions that serve as industry entry barriers: first, the core data architecture must be redesigned. Many “self-built databases” underestimate the business value of object models. Traditional software revolves around dashboards, reports, and manual workflows, with core objects like opportunities, work orders, and candidates. For AI agents, data architecture must support logical reasoning, action execution, state tracking, exception handling, task delegation, and cross-system coordination, shifting focus to tasks, user intents, session flows, governance rules, and business outcomes.

Second, permission systems must be upgraded to manage AI agents, not just personnel. Clear rules are needed: who has authority to operate, which agent performs which actions, what governance standards are followed, what approvals are required, how audit trails are maintained, and how to handle exceptions and rollbacks.

Of course, all these depend on cost constraints: investments in building and maintaining AI agents and databases, API call expenses, etc., ultimately circle back to data replication difficulty and system dependency complexity.

Where will all this ultimately lead?

When traditional software vendors shift toward headless architectures, they are implicitly betting that: the core value will still reside in the data layer. In some industries, especially in highly regulated fields like financial services, this judgment may hold for a longer period, and the full transition to headless may lag.

For software entrepreneurs, the current shift of traditional vendors toward headless systems rewrites the opportunity logic for building long-term moat software products.

Next-generation core record systems are already evolving into a new form: they are no longer just containers for storing data to record manual work, but possess native intelligent agent capabilities, capable of automatically capturing business context, initiating workflows, and accumulating full-process derivative data.

Furthermore, the most promising companies will extend into real-world business execution—coordinating field personnel, logistics providers, service teams, and physical assets, or acting as intermediaries in multi-party transactions. This new model will blend traditional business paradigms; meanwhile, the core that traditional record systems rely on—data—will recede into the background, serving as the foundational infrastructure.