The latest judgment from a partner at Andreessen Horowitz (A16Z): There are only two paths for AI startups, either like an oil well or a pipeline.

Not long ago, Joe Schmidt and Angela Strange, partners at A16Z, published an in - depth reflection on the path of AI entrepreneurship.

In this article, they used a classic metaphor from the energy industry to summarize the two choices facing founders today:

First, drill an "oil well". Deeply explore a specific scenario, master core data, and ultimately form a complete record - keeping system.

Second, build a "pipeline". Connect dispersed systems and processes, and automate tasks that originally relied on manual judgment and cross - departmental collaboration.

Although the two paths seem different, both may give rise to large - scale and well - fortified companies.

A16Z believes that oil wells and pipelines are not in opposition but two complementary wealth - creation logics in the AI era. The key is not which one is better, but whether founders clearly know which game they are playing and can firmly see it through to the end.

01

The Oil Well Path: Replacement and Reconstruction

In the early days of the energy industry, there were two distinct ways to accumulate wealth. An oil well means "finding a spot". Once a well with rich oil reserves is drilled, it can generate a continuous cash flow, allowing one to rely on a single reserve for decades.

A pipeline, on the other hand, means "connecting the dots". It doesn't involve owning the resources themselves but becoming the essential route for resource flow, creating stable returns through large - scale transportation. Simply put, oil wells rely on scarcity, while pipelines rely on connectivity. Both strategies have given rise to industry giants.

Today's AI founders face a similar choice. They can either choose to "drill a well", deeply explore a specific workflow, thoroughly understand the processes, data, and customers, and ultimately form a complete record - keeping system; or they can choose to "build a pipeline", connect different systems and processes, and automate tasks that require manual judgment and cross - departmental transfer.

Although the two paths seem different, both can give rise to large - scale companies, but they differ in terms of construction methods, sales logics, and sources of competitive moats.

Looking back at the history of enterprise software, the most profitable and resilient B2B companies are almost without exception "record - keeping systems". Systems like ERP, CRM, and HRM essentially lock customers into their ecosystems by controlling the underlying data of enterprises, making work processes dependent on them, and thus building long - lasting competitive moats.

The advent of artificial intelligence has greatly accelerated this trend. Unlike the bulky and functionally fixed systems of 30 years ago, today's AI startups can provide an order - of - magnitude improvement in efficiency.

Old systems appear slow and vulnerable in the face of AI. Boards of directors and management are already discussing "buying AI", which means that the sales cycle is shortening, and replacement opportunities are emerging at an unprecedented pace.

The "oil well" strategy is most suitable for scenarios where data is unstructured and scattered across different systems. Once someone can integrate this messy data into a clear model, the improvement in customer experience will be revolutionary. There are two types of opportunities with this approach:

Type one, Replacement and Reconstruction.

When old systems are too outdated and full of problems to support AI, startups can replace them with brand - new, AI - native systems. As long as the new solution brings about a significant enough improvement to make customers feel that replacing the old system is worthwhile, an opportunity arises.

For example, Valon built a mortgage service system from scratch, integrating all the processes that were previously scattered across 25 different old systems into Valon OS. This system can automatically generate auditable ledgers, set programmable workflows, and has an AI assistant to assist with compliance checks and customer service. As a result, a business that could barely break even now has a profit margin of over 60%.

Another example is Vesta, which developed a new mortgage approval system. The old data architecture was so outdated that only one person could handle a loan process at a time, which is why loan approvals often took more than 30 days. Vesta's system allows different stages to be processed in parallel, increasing both the approval speed and accuracy several times over.

Type two, Starting from Scratch.

When there is no mature software system in the market and many processes have to be handled manually, startups have the opportunity to enter the market, acquire customers first, and then grow with them. Usually, they start with small and medium - sized businesses (SMBs) and then enter the large - enterprise market as their functions become more and more complete.

For instance, Rillet developed an AI - driven ERP tool that can automate financial tasks such as month - end closing and real - time reporting. Many of its early customers used to keep accounts with pen and paper, Excel, or Quickbooks. Rillet became their first official system and has grown with them. Now, Rillet is in a position to challenge established systems like NetSuite.

Companies with core data models can not only develop functions that others can't replicate but also gradually make customers dependent on their workflows, resulting in high conversion costs. Just like an oil well, the drilling process is long, but once successful, it can provide a deep and long - lasting competitive moat.

02

Two Scenarios Most Suitable for the "Pipeline" Model

According to traditional views, some might say that building around a record - keeping system is just a function and not enough to support a company. Indeed, in some scenarios, established enterprises can incorporate new orchestration tools during their development.

However, the reality is that many traditional "oil wells" are deeply entrenched, with extremely high migration costs or strict compliance restrictions, making it difficult for them to transform quickly. Meanwhile, the market's demand for efficiency is more urgent than ever. This creates a new opportunity for AI: intelligent agents can now capture market opportunities that were previously too small and fragmented to be addressed.

The "pipeline" doesn't aim to replace the core systems but rather hands over the "glue work" that people used to do between systems to AI. For example, handling messy unstructured information, making judgments based on context, and running between different processes and departments to complete tasks. These tasks used to be done manually, but now AI can take them over, presenting software with a great opportunity to solve long - standing problems.

Specifically, there are two main scenarios suitable for the "pipeline" model:

Scenario one: Dispersed Old Systems.

Many large companies have been using old systems for decades, and these systems are completely incompatible with each other. As a result, information is scattered everywhere, and communication between departments is inefficient. However, tearing down and rebuilding these core systems would be too costly and time - consuming, so they value "immediate results".

This is where the value of the "pipeline" comes in: it can unify the data and processes between different systems, like adding a "master control console" between old machines.

For example, Further built an AI workspace for the insurance industry that can automate paper - based processes such as submissions, loss records, and compliance. Using a small number of common industry documents (policies, ACORD, SOV, etc.), it can "string together" the previously dispersed systems to form a smooth workflow.

Scenario two: The Human "Middle Layer".

In many industries, although software exists, a large amount of manual work is still needed to "patch" the operations. People are forced to act as the "middle layer" between systems: moving files, entering data, and conducting checks.

Now, large language models (LLMs) can take over these tasks, digitizing the processes that were previously only possible with human labor, thus achieving scale. These opportunities, which couldn't be addressed in the past, have now become a new market that can "nurture unicorns".

For example, Concourse developed an AI assistant for corporate finance teams. Without replacing the underlying systems, it can connect to all the company's financial software, automatically complete queries, analysis, and reporting, replacing hours of manual work.

Sola developed an AI - native back - office automation tool. Users only need to record an operation process once using a plugin on their computer, and Sola can generate a real - time, self - adapting AI agent to perform tasks such as invoice reconciliation, claims processing, and data entry that were previously all done manually.

Compared with the "oil well", the charm of the "pipeline" lies in that it doesn't require customers to start from scratch.

A large amount of manual labor in enterprises is essentially the "glue" between systems - manually moving files, communicating across departments, handling unstructured information, and making decisions based on context. These are exactly the strengths of AI.

The pipeline can quickly reduce human input and connect previously isolated systems. Over time, with each new workflow added, the value of the platform will grow exponentially, creating stronger and stronger stickiness.

Importantly, customers don't have to choose one over the other. In complex enterprises, these two needs often coexist: some business segments require a completely new record - keeping system, while some business processes only need lightweight automation. It's the entrepreneurs themselves who really need to make a choice.

Therefore, the key question is not "which is better, the oil well or the pipeline", but to clearly know which game you're playing: if the opportunity lies in mastering key data and unlocking new workflows, then go for the oil well; if the market is too fragmented and labor - intensive to replace old systems all at once, then build a pipeline to unlock value through automation.

The bottom line is that oil wells and pipelines are not in opposition but two complementary paths.

The value of an oil well lies in mastering the basic facts; the value of a pipeline lies in efficient orchestration around those facts. Both can give rise to far - reaching companies. For entrepreneurs, the most important thing is not to try to do both but to consciously make a choice and firmly see it through to the end.

This article is from the WeChat official account "Crow Intelligence Talk". Author: Intelligent Crow. Published by 36Kr with permission.