Vertical AI / Horizontal AI & Other Specialized AI Models in 2024

Foundation models like ChatGPT with many capabilities (e.g. translation, text generation) trained on public data have launched the generative AI wave. However, businesses need to work with specialized enterprise generative AI systems trained on private data for increased effectiveness.

AIMultiple’s specialized generative artificial intelligence framework splits specialized AI systems in 3 categories to facilitate comparing similar systems:

Horizontal AI

Horizontal AI systems are focus on one business function or process like customer service, accounting.

Vertical AI

Vertical AI systems are specialized in a specific industry like banking or pharma. Currently, vendors are building specialized models for various industries.

Real-life example:

Healthcare is a specialized domain with significant private (e.g. patient records) and public data (e.g. scientific papers). Vendors are building specialized healthcare models that can outperform other large language models.¹

Common AI

These systems leverage machine learning to provide specific capabilities like search, integration or automation across the enterprise.

Real-life example:

Numerous automation companies have rolled out generative AI offerings that allow users to develop automation solutions with prompts. This is a further improvement to low code automation and allows users to use natural language to build automation solutions. These features are typically called copilots or assistants.²

Why is specialized AI relevant now?

Higher performance

From an inference (i.e. running a machine learning model to produce predictions) perspective, we have hit the limits of increasing the scale of dense transformer models. A state of the art 8xGPU cluster can not serve a multi trillion parameter dense transformer model at a fast enough speed to keep a human reader engaged.

This is why OpenAI relied on a Mixture of Experts architecture in GPT-4.³. Such architectures enable building smaller expert machine learning models that act together to solve a diverse set of problems.

Since our capabilities in building larger deep learning models may grow slowly, we need to increase the specialization of these models to increase performance. There has been numerous examples of specialized data improving model performance.⁴

Improved UX

Specialization allows user interface and functionality improvements as well. For example:

• A transcription software running on a mobile app which has the capability to understand voice commands can unlock more use cases than the same specialized AI model running on a desktop.
• An accounting model with tax rules embedded into the solution can provide relevant subject matter expertise to its users. Such solutions can combine machine learning models, rules-based programming and human intelligence to solve complex problems.

Lower computing costs

Foundation models are pre-trained on large amounts of text data, such as websites, books, and articles, to learn the underlying structure and patterns of human language. As a result, frontier models require trillions of parameters and require GPUs for inference.

Specialized models can focus on a more specific training dataset, require fewer parameters and therefore require less computing power for inference.

Data security

Specialized models have lower resource requirements making it easier for enterprises to run these models on their own cloud infrastructure. Therefore, enterprises can provide their proprietary data with ease as training data for these models without transferring data to cloud infrastructure managed by 3rd parties.

How can you find vertical AI / horizontal AI solutions?

Vertical AI can be found like vertical SaaS (vSaaS). Users can search industry analysts like AIMultiple to find new solutions. For example, AIMultiple’s AP AI list is contains horizontal AI vendors in the accounts payable (AP) domain.

What is the future of specialized AI technologies?

Both established SaaS companies and AI startups are launching new specialized AI solutions. As these solutions mature, AIMultiple expects vendors to consolidate their offerings to help enterprises:

• minimize application switching for employees
• vendor sprawl

However, given that specialized AI is an emerging field, we don’t expect this consolidation in the next few years.

Notes & external links

All models mentioned in this article are narrow AI models and not generalized artificial intelligence models. Generalized AI is not likely to happen in this decade.

External Links

• 1. “Benchmarks“, Hippocratic AI, Retrieved July 29, 2023
• 2. “Introducing Boomi AI, a Generative AI Game-Changer for IT and Business Success“, Boomi, Retrieved July 29, 2023
• 3. “GPT-4 Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE”, SemiAnalysis, Retrieved July 29,2023
• 4. “Introducing BloombergGPT, Bloomberg’s 50-billion parameter large language model, purpose-built from scratch for finance”. Bloomberg. March 30, 2023. Accessed May 24, 2023