Published by Surojit Chatterjee in Agentic AI
Copy Link
Every week, I engage with numerous customers and encounter many CTOs, CIOs, and Chief AI Officers struggling to determine which GenAI systems would be the most beneficial for their enterprises. The focus is often on automation — how can we achieve more with less? What role will co-pilots play? Are RPAs still necessary? What about assistants? Navigating this landscape can be complex, with countless startups and established companies introducing new products. In this edition of our blog series, we delve deeper into agentic applications and their potential to transform enterprises.
What are Agents or Agentic Systems?
Large Language Models (LLMs) and Retrieval Augmented Generation (RAG) systems, which query large documents with precision, have already gained popularity across enterprises. Today, every application claims to have a co-pilot, essentially a conversational interface for existing functionalities. Agentic systems, however, take this a few steps further. These systems encompass planning, loops, and other structures, enabling them to accomplish tasks end-to-end. Agents can use tools, making life much easier. Think of these tools as the ultimate sidekicks, helping our model interact with external data, whether by pulling in information or sending it out.
Imagine a claim adjuster in the insurance sector as a collection of agents. This adjuster has a defined role and set of tasks, utilizing tools to review claim details, assess damage reports, or send approval emails for claim settlements. Agents benefit from LLMs’ reasoning abilities. As foundational models become more powerful, creating agents will become increasingly straightforward.
While basic LLMs and co-pilots are excellent for specific tasks, agents excel at automating entire complex workflows. But how do agents differ from RPAs? The difference is clear to anyone familiar with Gen AI. RPAs are essentially ‘if-then-else’ statements meticulously woven together to perform complex tasks. They are extremely fragile; any change in context requires rewriting RPA scripts. RPAs lack reasoning abilities, making them incapable of handling tasks that require understanding complex data (text, video, audio) and dynamically figuring out how to act.
What are Ema's AI Employees?
We started Ema with a single mission: to create AI employees capable of assuming any role within an enterprise and automating monotonous, repetitive tasks. AI Employees are an agentic system where multiple agents collaborate to emulate a human role. Each agent may access different models and tools. Ema’s AI Employee Builder allows customers to specify the goals, resources, and constraints for Ema in the role they want to hire her for. A single conversation is enough for Ema to create and deploy a new AI Employee for a new role. Ema has already been trained on hundreds of enterprise applications, for both read and write actions. Ema can also call into RPAs and has API interfaces, enabling seamless integration with internal and external applications. An AI Employee typically spawns multiple agents, orchestrating them to complete complex tasks. Each agent can dynamically figure out what actions to take based on a high-level description of the task it receives from another agent (or humans).
What Can Ema’s AI Employees Do?
Ema’s multi-agent architecture makes it the most flexible agentic system in any enterprise. In her role as a customer support agent, Ema not only acts as a chatbot to answer many customer questions but also resolves complex customer issues. For instance, in customer refund cases, Ema’s AI Employee for support assistance examines policy documents, accesses databases, decides whether a refund is warranted, processes the refund, and emails the requesting customer maintaining the tone and brand voice your company uses. Unlike the previous generation of chatbots, customers do not need to manually configure and maintain a refund workflow. Ema automatically creates this workflow based on a simple conversation with users and stays up to date with any changes in the refund policy documents, database schemas etc.
Today, Ema’s AI Employees are working on a variety of use cases, from aiding in the prior authorization of medical procedures and managing various healthcare workflows to adjusting insurance claims and writing business proposals.
Building Adaptive Software: How Ema Aligns with Human Feedback
Working with customers since day one has taught us that building products in Gen AI is fundamentally different. Traditional software changes behavior only when a manufacturer pushes a new update. Ema’s behavior and performance can be modified through direct customer feedback, much like a human employee. Proper onboarding and continuous performance feedback enhance Ema’s performance over time. Ema’s AI Employees are designed to easily receive and act on human feedback, which can be given in plain natural language. With Ema, we’re ushering in a new type of software — adaptive software that morphs based on human feedback.
How Can Enterprises Build Their Own Ema AI Employees?
We created a platform that enables not just engineers but also business users to build AI employees. With Ema, you can start building your next multi-agent AI Employee through a simple conversation. You can always switch to a visual builder or configuration window for further customization. In a follow-up blog post, we’ll go into the details of Ema’s AI Employee builder feature.
Ensuring Trust and Safety
A recurring concern for hundreds of CIOs, CTOs, CAIOs (Chief AI Officers) is a potential for AI agents to go rogue or leak private data. Ema is built with trust at its core, compliant with leading international standards such as SOC 2, ISO 27001, HIPAA, and GDPR. Robust security measures, including automatic redaction and safe de-identification of sensitive data, audit logs, real-time monitoring, robust encryption of all data at rest and in transit, and explainability across all output results, ensure your enterprise data remains private, secure, and compliant. For document generation use cases, Ema also checks for any copyright violations to ensure customers don’t incur IP liabilities. Additionally, Ema never trains models on one customer’s data to benefit other customers.
Embrace Agents in the Enterprise
Agentic systems offer unprecedented opportunities to accelerate your enterprise. Agents learn from your best employees, providing consistent and high-quality automation services that evolve with your enterprise context. We built Ema as an operating system for the enterprise (EmaOS), enabling you to conversationally build any agentic application (Ema’s AI Employees). Working with customers has highlighted the complexity and enormity of this task, driving us to create real innovation and multiple patent-pending technologies.