AI Agents vs Chatbots: A Paradigm Shift in Business Software
Since starting our journey at Frontline earlier this year with my co-founder Esteban Paredez, the key insight that we developed is that deploying AI in a business environment, particularly AI Agents, represents a paradigm shift from deploying traditional software. What has been interesting so far though, is the market’s perspective on this. Our experience so far based on meeting with hundreds of customers, investors and partners is that most people don’t fully understand the fundamental difference between traditional chatbots/IVR systems and AI agents. The most common reaction is: “Chatbots! Yes, they’ve been around for a decade and there are thousands of companies doing this. How is Frontline any different? I wanted to share my thoughts on this and why I think AI Agents are a totally different software category, similar to how a car is different from a train in terms of transportation.
To provide some context, my background is in building communication software for businesses. Previously I built PhoneIQ, a contact center platform for companies running on Salesforce, which we scaled to hundreds of customers across 15+ countries. But my entrepreneurial journey started in 2013 when I built a mobile app called “Tellmi” which allowed consumers to text businesses, when WhatsApp Business didn’t exist. Although we had some decent traction and reached #1 in the App Store in Uruguay, we failed, mainly because businesses wouldn’t respond to users in time. We tried to fix this in several ways, but one cool thing I recall is our amazing investor Nicolas Jodal suggesting we use AI to automate responses. This seemed impossible at the time, but fast forward 10 years, and it has now become inevitable.
What’s interesting is that the software that we built at my two previous companies is foundationally different from what we are now building at Frontline. I’ll get back to that in a minute. But first, customers really don’t care about this. They care about results and getting as much value from their money as possible. Our bet at Frontline is simple: although AI Agents might seem similar to traditional Chatbots or IVR systems today in capabilities, they will become better at an exponential rate, making these older technologies obsolete and driving unparalleled business value. Businesses will need to quickly shift to this new technology to stay competitive, but shifting to AI Agents will require entirely new platforms purposely built from the ground up to support this new type of software.
WHY DO WE NEED NEW PLATFORMS FOR AI AGENTS?
Traditional software is deterministic and rule-based, meaning a specific input will reliably produce the same output every time. For example, in the case of chatbots or IVRs, selecting an option within a decision tree will always yield the same result. From a business process perspective, a company can easily:
- Diagram Expected Behavior: Outline the exact flow of interactions and responses they expect from the chatbot.
- Program Responses: Setup the chatbot to respond to each input in a pre-defined way.
- Test Using Traditional Methods: Employ traditional quality assurance methods, such as automated or manual testing, to ensure the chatbot behaves as expected.
- Confident Deployment: Move the chatbot into production with confidence, knowing it will perform reliably.
- Repeatable Cycle: Easily repeat this cycle whenever changes or updates are needed.
However, AI, and AI Agents in particular, operate on fundamentally different principles. Unlike traditional software, AI systems, especially those powered by Large Language Models (LLMs), are non-deterministic. This means they can produce different results even with small changes in input or system configuration (e.g., base model, temperature settings, or prompt variations). Given that AI Agents typically communicate through natural language, whether text or voice, the range of potential inputs and interactions is virtually infinite.
The traditional software development cycle, which includes straightforward diagramming, programming, testing, and deployment, breaks down when applied to AI systems. The inherent variability and complexity of AI require a completely new approach and thus new platforms. At Frontline we’ve identified the following steps to reliably build and deploy AI Agents in production:
1. Defining the Agent’s Persona, Goals, and Guardrails
To create a successful AI Agent, teams must start by meticulously defining the agent’s persona, goals, and guardrails. The persona outlines the agent’s character, including its tone, style, and behavior, ensuring it aligns with the brand and meets user expectations. Goals clarify what the agent is designed to achieve, such as improving customer service or increasing sales. Guardrails set the boundaries for the agent’s actions, ensuring it operates within company policies, ethical and legal frameworks. This foundational step is crucial for developing an agent that is both effective and trustworthy.
2. Adding Agent Tooling and Workflows
Once the agent’s persona, goals, and guardrails are defined, the next step involves integrating the necessary tooling and workflows. These tools and workflows enable the agent to execute specific tasks, such as fetching an order status or updating customer information in business systems like a CRM. By equipping the agent with the right tools and processes, businesses can ensure it performs its functions efficiently and accurately, thereby enhancing overall customer experience.
3. Testing and Quality Assurance
Developing an AI Agent doesn’t end with its creation. It’s imperative to test the agent to ensure it responds and behaves as expected. Traditional quality assurance software and processes fall short in this context due to the infinite number of possible test cases. Innovative companies like Maihem are revolutionizing Quality Assurance for the AI era by offering platforms that automate the testing of LLMs and AI Agents. These platforms provide simulation runs at scale, complete with performance and security benchmarking, ensuring the agent is robust and reliable before deployment.
4. Proactive Monitoring and Continuous Improvement
Deployment is not the end of the journey for an AI Agent. Proactive monitoring is essential for continuous improvement. Regularly auditing the agent’s responses and expanding its knowledge base can significantly enhance its performance. Addressing unaccounted intents and queries, introducing new workflows, and updating existing ones are key strategies for maintaining and improving agent efficacy. This ongoing process ensures the agent evolves with user needs and industry trends, providing consistent value.
5. Thorough Retesting After Updates
AI is evolving at an unprecedented pace, so updates and improvements are inevitable. However, even minor changes, such as upgrading from GPT-4 Turbo to GPT-4o, necessitate a thorough retesting of the entire system. This is crucial as small updates can introduce significant changes and potential risks. Businesses must repeat the development cycle, including retesting prompts, tooling, and conducting extensive quality assurance, to ensure the updated agent maintains or improves its performance and security benchmarks. This diligence helps prevent commercial and security risks, ensuring the AI Agent continues to operate effectively.
What lies ahead in AI ✨
We are in the very early stages of AI and AI Agents. Reflecting on historical transformations, consider the analogy between cars and trains I made before: in 1916, trains accounted for 98% of passenger travel, but by 1956, that number had plummeted to 4%. I believe we are witnessing a similar transition with AI and AI Agents today. Chatbots, IVRs, and traditional software still represent the vast majority of enterprise software by a country mile, but this will change.
As AI continues to advance at an exponential rate, businesses will need to embrace this technology to remain competitive. The future will see a significant reallocation of dollar spend from traditional software and manual labor to AI systems that operate 24/7/365, can scale to unlimited capacity, and can be upgraded seamlessly with just a single line of code — from one language model to another. This is unprecedented in software!
The challenge for startups lies in surviving the current market conditions while waiting for the broader adoption of AI. Success will come to those who can position themselves to ride the wave of innovation and capture the early majority as they transition from traditional software to AI systems.
These are my insights and reflections on building in this space. If you are interested in learning more about Frontline, feel free to book a meeting with me or get started now.
Thanks for reading ✌️