Anthropic investors share their latest thoughts on vertical AI implementation

With the widespread application of AI technology in various industries, the implementation of vertical AI has become the key to the digital transformation of enterprises. However, the success of vertical AI does not rely solely on technology, but also requires a deep understanding and precise grasp of industry needs. This article summarizes the ten judgments of Anthropic investors on the implementation of vertical AI for your reference.

Software is considered one of the most important scenarios for AI implementation. Sequoia Capital once mentioned that AI has the potential to replace services with software, creating tens of trillions of dollars in market opportunities.

Despite the huge opportunities, there is still no clear path for how AI software can be truly implemented. Bessemer recently made a valuable point about this issue:

Vertical AI software will be the future.

People familiar with the SaaS industry may not be unfamiliar with Bessemer, which is one of the most professional investment institutions in the SaaS field in the United States and has invested in more than 200 SaaS companies in the past 10 years.

Although vertical AI is still in its infancy, we can still see that after the rise of generative AI, a number of excellent companies in the field of vertical AI have emerged, such as AI legal unicorn EvenUp (founded in 2019), AI medical company Subtle Medical (founded in 2017), AI medical company Abridge (founded in 2018) and automatic collaboration software platform Fieldguide (founded in 2020).

Based on the business cases of these vertical AI companies, Bessemer has developed 10 roadmaps for the implementation of vertical AI, covering aspects such as vertical AI functional value, economic value, competitive position and defensibility.

01 The implementation of vertical AI should start from the actual needs of customers

The core workflows of different industries have different needs for automation. However, whether the workflow has the foundation for automation is not the only factor that vertical AI companies need to consider when building their business.

Customers' interest in automation and different requirements for automation will also have a great impact on the implementation of vertical AI.

Sometimes these preferences or requirements can be addressed in product design. For example, a dental office might want to set up automatic review for medical supply purchases if the order is below a certain cost, but still require manual review for larger orders.

In other words, the AI ​​procurement solution needs to have a certain degree of flexibility, not only to achieve automated procurement of some orders, but also to enable human participation in other orders.

As another example, a law firm may be willing to fully automate the process of servicing client payments, but when it comes to core workflows like writing legal briefs, they need human feedback to create the final output (e.g., creating a first draft) because they want to control what is produced.

The implementation of vertical AI requires thorough research into the market and user needs of vertical scenarios.

For example, in healthcare, AI solutions for managing workflows provided by AI companies such as Abridge are seeing widespread adoption as clinicians look to automate administrative tasks such as record-keeping.

While there is also interest in the application of multimodal AI in diagnostics, penetration remains low because healthcare payment models lag behind innovation in healthcare technology.

Therefore, the implementation of AI in vertical scenarios not only needs to consider whether it has the conditions for automation, but also needs to pay attention to the actual needs of customers and their expectations for artificial intelligence.

02 Only by seamlessly integrating into existing scenarios can we build a product moat

Vertical AI solutions need to not only perform tasks excellently, but also build a real moat.

AI solutions that are easily copied will face enormous competitive pressure.

For example, in the financial services sector, there are more and more application cases for accounts receivable and accounts payable (AR/AP) automation solutions, where AI capabilities for data matching and invoice verification may provide some value, but these subtle functions can easily be integrated into a workflow tool and replaced by industry-specific workflow vertical AI solutions.

In order to reduce the risk of large-model commercialization, the best vertical AI applications not only need to fully cover the entire business process, but also need to achieve seamless integration with existing systems through APIs/plug-ins.

Many B2B AI startups achieve the latter by partnering with established platforms, especially large incumbent platforms, to create value through seamless integration.

For example, AI insurance company Sixfold is embedded in the existing policy management system (PAS) in the form of API or plug-in, so insurance companies do not need to completely transform the old system or rebuild the workbench. This "plug and play" integration method allows underwriters to effortlessly introduce Sixfold's AI capabilities directly into their daily workflows.

03 Look for opportunities to implement productivity-limited

AI is reshaping the division of labor in the workplace: it not only replaces repetitive work and frees up manpower, but also gives companies breakthrough operational capabilities. Vertical AI products that truly have transformative value often have two core advantages: full-process automation and massive data processing capabilities, which are areas that humans cannot reach.

For example, Rilla, an AI company in the home economics field, records and analyzes face-to-face interactions between sales representatives and customers, and provides customized feedback and suggestions to help sales staff improve their performance. Without Rilla, sales managers must personally accompany sales representatives on site visits, but ultimately they are still limited by their personal energy.

On the other hand, Rilla can also review large amounts of conversation data from sales reps across the company, which means the guidance it provides to sales reps is based on a much larger amount of data than any sales manager has at their disposal.

This is why certain industries, such as sales and marketing, services, and law, are particularly suitable for AI implementation:

The success in these fields is based on the knowledge generated by a large amount of written text and practical records. In the past, this was a time-consuming task, but now AI can do it better or even take over completely.

04 Efficiency improvement, the key point of vertical AI products

By using data to intuitively show customers the efficiency improvements brought by AI solutions, you can greatly speed up the sales cycle and improve customer retention.

This efficiency improvement generally comes from two aspects: controlling costs and generating more revenue.

For example, Abridge can automatically record conversations between doctors and patients, reducing doctors' workload, improving doctors' job satisfaction, and thereby increasing doctors' retention rates.

By improving retention rates, Abridge significantly reduces the costs of recruiting and training physicians—costs that typically run into the millions or even tens of millions of dollars each year.

In addition to controlling costs, Abridge is also increasing revenue by saving each physician one to two hours per day.

This extra time enables doctors to see more patients, directly improving the hospital's operational efficiency and generating more operating revenue. Abridge's detailed records and summaries of every patient visit also prevent revenue leakage by ensuring comprehensive coding and billing.

The EvenUp case also illustrates this point.

EvenUp uses AI technology to generate demand packages for personal injury law firms, whereas in the past paralegals would have spent days collecting data from clients, sorting through hundreds of documents, extracting data from medical and police reports, etc.

Because EvenUp’s legal operations team reviews every letter, law firms can maintain high quality standards while significantly reducing (or eliminating) the time their teams spend on on-demand packages. This extra time enables firms to take on more cases, thereby increasing revenue.

05 AI reshapes service delivery and pricing, bringing new business opportunities

New delivery and pricing methods brought by vertical AI solutions are creating new opportunities.

In the past, many vertical scenarios did not have enough TAM (total potential market) to build traditional software businesses. Now, this part of the market gap is expected to be filled by AI with lower costs and more standardized services.

Historically, service businesses have been difficult to make money because of the high cost of specialized workers. AI will completely change this. By 2024, the average gross margin of service companies in Bessemer's vertical AI portfolio will be about 56%, and the average capital consumption rate will be 1.6 times, that is, only $1.6 of operating capital will be invested for every $1 of profit.

Some AI service products have demonstrated better delivery results with the support of manual QA, and other service products with AI products as their core have also performed well.

06. Build for neglected categories and workflows

In the sales and marketing field, there are already large and resource-rich competitors, such as Salesforce or ADP. In this case, AI vertical companies should look for areas with relatively less competitive pressure.

While gaining first-mover advantage in a broad market is ideal, most vertical categories already have at least one incumbent.

But there are opportunities. When incumbents are stretched thin or slow to integrate AI, startups that move quickly can gain a competitive advantage by building superior, high-ROI AI products and services that optimize valuable but non-obvious workflows with automated AI solutions.

07 Provide services for customers with specific needs

Vertical AI companies differentiate themselves by targeting customers in neglected categories, who often have complex requirements that cannot be easily met by AI solutions.

For example, an AI startup that provides services to banks or government contractors needs to build industry-specific security and compliance tools to sell to customers. This complexity based on specific industry needs creates a moat for AI companies’ products.

To reduce the risk of LLM commoditization, we may start to see foundational model players (such as OpenAI and Anthropic) also start to build corresponding vertical models for customers in these industries.

08 Models are not a reliable moat, but multi-models can be

As model infrastructure costs continue to fall, models will no longer be a moat. Early vertical AI founders need to ask themselves: "Why are the products we build with AI better than those built with public models and data?"

Building new technical architectures to solve specific problems may be one approach, for example, fine-tuning LLM to better reflect the writing style of a client, or using retrieval-augmented generation (RAG) to better perform information retrieval.

Bessemer believes that applying RAG technology to industry-specific data sets is also a way to establish commercial barriers.

New business niches will be found in solutions that can handle more complex (especially multimodal) workflows.

For example, Bessemer portfolio company Jasper is a great example of an AI solution that is ultimately used by marketers to create long-form blog posts based on their text-based GenAI capabilities.

Generally speaking, once a post is generated by AI and edited by a marketer, the next step is to find the right image to go with it. Therefore, Jasper acquired Clickdrop to strengthen its Jasper Art product, using multimodal capabilities (text and images) to meet all the needs of marketers.

09 Focus on modularity and extensibility of the model stack

Traditional SaaS relies on the arrangement and combination of standard technology stacks, while vertical AI companies must build a customized infrastructure system: integrate open source models and commercial solutions through self-developed capabilities, flexibly fine-tune large language models, and achieve the best results for customers.

This approach allows AI companies to seize the initiative in the rapid iteration of large models. At the same time, it reduces the cost of trial and error. When the open source model can achieve 90% of the effect of the commercial model after tuning, there is no need to take the risk of self-development.

More importantly, this approach also allows companies to invest resources in what matters most: providing customers with quality products.

In this regard, Jasper is an excellent example of a product built for flexibility. The platform sits at the core of the marketing technology stack, acting as an “AI brain” that helps users develop, design, and execute plans for all marketing specialties.

The Jasper team designed a modular platform that uses multiple LLMs, and can run marketing inputs through multiple LLMs based on customer needs, model performance, and cost. For example, if Claude 3.5 outperforms GPT-4 in a certain case, Jasper can support interchangeable model infrastructure.

10. Don’t over-pursue data quantity; data quality is more important

It is widely recognized that proprietary datasets can build moats.

But for many early-stage startups, they can’t get the amount of data they want. This is when you can start with the quality of the data, because high-quality data (regardless of the amount) can have a compounding effect, and over time, the company will benefit greatly.

For example, in the early days of EvenUp, the team invested heavily and intentionally in legal operations to have humans review all claim letters; in this case, data size was not as important as data quality, and over time, large amounts of high-quality data feedback would further refine the model to improve the product.

In the early stages of a startup, it is more important to create a product with a high return on investment, meet the pain points of core customers, and sell well quickly. As the scale of use expands, proprietary data will follow, and these high-quality data can also lead to product upgrades.

Text/Lin Bai