The Rise of Agentic Systems: From Bots To Agents

When generative AI first emerged, people mostly used it for simple things like creating recipes or planning trips. But as they became more familiar with what it could do, they’ve also grown more curious about its potential for handling common work-related tasks like drafting emails, summarizing meeting notes, and composing documents.

Now, every company wants to be an AI company. Indeed, executive urgency to put the technology to use has increased seven times over the past six months and is now a top concern, above inflation or the broader economy. What’s more, 77% of business leaders worry they’ll miss out on the AI revolution if they don’t deploy it soon.

"Now, every company wants to be an AI company." – Jayesh Govindarajan, ​​EVP of Salesforce AI Platform

They’re absolutely right. Companies that haven’t already implemented AI risk losing significant ground to competitors, and this could happen more quickly than they think, as we move from chatbots to copilots to autonomous AI agents or “agentic systems.”

Long before emerging large language models (LLMs) and generative AI spurred serious business and even consumer interest in artificial intelligence, many of us were already interacting with rudimentary AI chatbots without even knowing it. These bots were all around us, handling simple, predefined tasks like answering common FAQs or recommending products based on a shopper’s history. Companies have actively used them to deliver better customer experiences more efficiently and cost-effectively. Salesforce Einstein Bots, for example, are used by more than 3,000 customers and handle around 65 million sessions per month.

Heathrow Airport has used Einstein chatbots to offer round-the-clock support, answering 4,000 questions a month and helping cut call service volumes by 27%. Heathrow has seen a 450% increase in live chat usage since it launched in May 2023, freeing up agents’ time and improving efficiency. And today, Heathrow is seeing around 40- to 60-second quicker per-contact interactions in the call centers with Einstein chatbots.

Bots, however, have typically been limited to specific scripts and can sometimes seem robotic because they lack natural language and reasoning capabilities. What’s more, they sometimes lack nuance, context, and personalization when they aren’t grounded in corporate data and metadata involving their customers.

Copilots began to change that by adding generative AI, natural language processing (NLP), and, for business use cases, CRM to the mix to simplify routine tasks and provide more dynamic and less hand-crafted suggestions in the flow of work. Salesforce’s Agentforce Assistant, for example, enables enterprises to use their own unique data and metadata through Data Cloud, built on Agentforce, to produce powerful customer insights and recommendations while using the Trust Layer to help maintain privacy and data governance. Unlike other AI assistants or copilots that lack adequate company data to generate useful responses, Agentforce Assistant is an enterprise-class copilot that enables customers to generate responses using their own private and trusted data.

Bombardier, a leading jet manufacturer that designs, builds, and maintains high-performance aircraft for individuals, businesses, and governments worldwide, is using Agentforce Assistant to consolidate need-to-know information on prospects for sales reps in advance of meetings, and provide recommendations on how to best engage. Agentforce Assistant saves the sales team time that can be spent meeting new prospects and supporting existing clients by transcribing voice notes from customer meetings and interactions.

"Agentforce Assistant saves the sales team time that can be spent meeting new prospects and supporting existing clients by transcribing voice notes from customer meetings and interactions." – Jayesh Govindarajan, ​​EVP of Salesforce AI Platform

Still, copilots aren’t fully autonomous. For business use cases, they’re extremely helpful for assisting in activities like scheduling meetings, updating CRM records, drafting emails, and conducting preliminary research. Broadly, they can orchestrate complex actions on behalf of the user, but they require those skills to be configured and need a degree of hand-holding to perform at their best. They’re almost like interns or new hires who are super smart and good at straightforward work, but need guidance and oversight to do much more.

For those things, you need agentic systems, which can be thought of as trusted digital colleagues as opposed to digital assistants.

An advanced form of AI, they can perform higher-order planning, reasoning, and orchestration without needing much in the way of human handholding. Unlike traditional software programs that follow predefined rules, autonomous AI agents not only improve productivity but augment employees with new skills and abilities, building deeper customer relationships that span every interaction and deliver higher margins by fully automating routine tasks. They also interact with human colleagues and customers in human-like ways.

For example, Salesforce recently announced two new fully autonomous sales agents to help scale and train sales teams. Built on Agentforce, Agentforce SDR Agent autonomously engages with inbound leads in natural language to answer questions, handle objections, and book meetings for human sellers. Agentforce Sales Coach Agent, meanwhile, autonomously engages in role-plays with sellers, simulating a buyer during discovery, pitch, or negotiation calls.

This announcement follows the July launch of Salesforce’s first fully autonomous AI agent, Agentforce for Service which delivers trusted customer support 24/7 on a broad range of service issues without preprogrammed scenarios. That same month, Salesforce also announced a partnership with Workday to deliver a new AI-powered assistant for employee services, such as onboarding, health benefits, and career development. In the coming months, expect Salesforce to release other AI agents to automate work functions for specific professions. Some of these agents will be ready to use right out of the box. Others will eventually be customizable to meet a company’s specific needs.

Early on, many of these agents will work independently, meaning they won’t interact with other agents focused on different tasks. But that will change because, just as a sales rep must interact with service agents and marketers, or an HR lead must regularly consult in-house attorneys or hiring managers, autonomous AI agents will eventually need to team up with other agents.

Of course, this can all get pretty complicated, which is why a unified platform of agents like Agentforce will be needed for building, training, and supervising custom autonomous AI agents that work independently or together. Like any company in the physical world, the agentic world will need systems such as this for supervising and monitoring agents, quickly deploying them when and where they’re needed, and auditing their performance while holding them accountable for meeting their goals.

There’s still plenty to explore and understand with agentic ai. This journey is akin to the evolution of autonomous driving. The technology started in earnest with cars providing specific features that drivers have the option of enabling, like lane departure warnings, automatic parking, and emergency braking. But as ‌technology advanced, we started seeing driverless taxis carrying passengers down the busiest of city streets.

The point: With autonomous anything, you’ll have a spectrum of options operating independently or together. It won’t just be chatbots or copilots or agents serving business needs‌ — it will be all of them operating as one to shape the future of enterprise IT.

"It won’t just be chatbots or copilots or agents serving business needs‌ — it will be all of them operating as one to shape the future of enterprise IT." – Jayesh Govindarajan, ​​EVP of Salesforce AI Platform

Many business leaders are cautious about fully autonomous AI agents, as trust in the technology is still building. They are aware of the possibility of errors with outdated data, which can affect its efficiency.

But, by continually using a company’s own data rather than training models on publicly available information every few years, as some LLMs do, the accuracy and relevancy issues can be overcome. Also, by running AI inquiries through systems like the Trust Layer, which perform functions like masking personally identifiable information (PII) and defining clear parameters and guardrails for AI agents to follow, trust problems can be similarly addressed.

It’s essential for AI agents to understand and execute their expansive capabilities, but it’s equally, if not more, important for them to recognize their limitations and understand when human intervention is necessary. Our agents are trained to fall back, to know when they don’t know something and then, with proper guardrails, engage in agent-to-human handoff.

Throughout their AI implementation process, we work with our customers and offer tailored solutions, fostering a sense of ease and confidence as they embrace the spectrum of autonomous capabilities of AI on the market today. It is now time to shift our focus from what is achievable to preparing for what is inevitable.

• Discover Agentforce
• Learn more about Einstein Sales Agents and Einstein Service Agent
• Unlock the Four Keys to Enterprise AI Success
• Find out more about the Trust Layer and the Einstein 1 Platform
• Say hello to Salesforce Data Cloud