Vertical AI - What is it and what are the benefits?
Vertical AI specializes in specific industries and their unique use cases, as opposed to horizontal AI, which is designed for a variety of applications across different industries (such as the ChatGPT tool), vertical AI focuses on solving specific problems and optimizing processes within a specific vertical sector.
The main features of Vertical AI are:
- Specialized knowledge: Vertical AI solutions are based on deep, industry-specific expertise.
- Highly customizable: they are designed specifically for the requirements and workflows of a particular industry.
- Accuracy: by using relevant, industry-specific data sets, Vertical AI achieves much higher accuracy than general AI tools.
- Seamless integration into workflows: Vertical AI applications can be easily integrated into real-world processes, improving productivity and efficiency.
How does epilot go about developing Vertical AI?
Vertical AI will fundamentally change the way energy suppliers and grid operators handle their core processes. The energy sector faces unique challenges that cannot be solved with off-the-shelf AI solutions. That's why we develop specialized AI tools based on a deep understanding of industry-specific workflows, regulatory requirements and the precise needs of energy suppliers and grid operators.
Our AI implementation strategy for the epilot platform is very results-oriented.
We do not develop AI to follow the hype, but carefully evaluate all processes and functions in epilot to determine where Vertical AI can bring the most value to our customers. We then analyze whether a generic AI model can deliver the required accuracy or whether we need to improve performance through specific customizations and contextual data.
Our approach at epilot differs from others in that we link several AI tasks together to create intelligent “agents” that can handle even complex processes and support our customers as a personal AI assistant.
epilot AI Mail Agent, for example, supports all tasks from initial email classification to data extraction and pre-formulation of the response when processing service requests.
We are thus working on Vertical AI solutions that gradually eliminate manual tasks while ensuring the high standards of accuracy and reliability required in the energy industry.
Which AI models does epilot use?
epilot does not train its AI models itself. Instead, we use pre-trained large language models (LLMs) that do not require additional training with customer data and specialize them for the specific needs of the energy industry.
epilot primarily uses the Claude 3.5 Sonnet family of models developed by Anthropic. We access these models via Amazon Bedrock, an AWS-managed service that provides secure access to state-of-the-art LLMs.
We continuously evaluate and adopt the latest LLM technologies to ensure our customers benefit from the most advanced AI capabilities available.
The specialization of the models takes place through:
RAG (Retrieval Augmented Generation): We combine these models with our energy-specific knowledge databases and data sources. This allows the AI to draw on accurate, industry-specific information when generating answers.
Prompt Engineering: We develop specific prompts that are tailored to energy industry use cases and guide the AI to produce more relevant and accurate results.
This approach allows us to “verticalize” the general purpose AI models for the energy sector without having to train our own models on customer data, which is a major advantage in terms of data privacy.
Data protection and security
Amazon Bedrock follows a zero storage policy by default. This means that no call histories or prompts are stored or logged when the LLMs are used via this service. Amazon Bedrock does not use the prompts and completions to train AWS models and does not share them with third parties.
The specific data we can include in prompts are:
- Entity data (opportunities, contacts, messages)
- User data (preferred language, e-mail addresses)
For quality improvement purposes, we store some input and output requests from LLMs, always following strict data protection standards and protocols.
Our security concept includes, among other things:
- Strict separation of clients
- End-to-end encryption
- Regular security checks
- Strict access controls
- Compliance with industry standards
For the AI-generated email responses, we use the open source vector database Weaviate to leverage a technology called Retrieval Augmented Generation (RAG).
This RAG technology allows us to build comprehensive knowledge bases for our clients, enabling much more accurate and customized email responses for each client company.
Before using Weaviate, we remove all personally identifiable information (PII) from the email conversation and anonymize the content. This ensures complete protection of personal data while maintaining the contextual value of the communication.
Example: The epilot AI Mail Agent
Communication with customers is a key part of the workload for energy suppliers and grid operators in sales and service, and the AI-based epilot Mail Agent is there to help them process emails as a personal assistant so that they can complete these tasks as efficiently as possible.
This saves employees time, which they can use for other tasks, and customer enquiries are answered more quickly.
The epilot Mail Agent analyzes and categorizes incoming emails, understands their content and formulates a suitable response tailored to the company on this basis. Thanks to the company-specific knowledge database, these responses are tailored precisely to this customer.
The agent also learns from the employee's previous responses. If another end customer asks a similar question or requests the same information, this is stored in the epilot mail agent's memory and is taken into account in the newly suggested email response.
This cross-conversation memory runs via the secure Weaviate database, in which all email communication with end customers is anonymized and stored with strict separation of the clients. The data of epilot customers is therefore never mixed together.
This approach ensures:
- Complete data isolation between different customers
- No data loss between organizations
- Fully compliant with data protection regulations
- Secure, encrypted storage of all vector data
The epilot Mail Agent not only supports the reading and writing of emails, but also helps our customers with their work in the epilot platform itself. Based on the data in the incoming end customer email, it suggests updating or creating new business objects in epilot, such as opportunities, contacts, counters and more. As a result, manual effort can be reduced to a minimum.
epilot's vision for Vertical AI and how we got there
At epilot, we take a systematic approach to implementing Vertical AI based on a step-by-step procedure:
- First, we digitize our customers' business-critical processes, which forms the basis for the structured collection of process data.
- This data is then intelligently classified in order to gain a deep understanding of the process flows. Based on this, we identify the most important and most frequent tasks that are suitable for AI-supported automation.
- These tasks are then converted into specialized AI tools and integrated into our epilot Agent Framework.
- At the heart of our vision is the intelligent linking of these AI agents, which work together to execute complex processes autonomously.
- In doing so, we create a unique Vertical AI for the energy industry that revolutionizes the process flows of energy suppliers and grid operators.
