Exploring AI Agent Models

Show Notes

Jacob Bank, founder and CEO of Relay.App, talks about the rise of AI agents, a type of chatbot that can work on your behalf in the background. He explains that AI agents can perform tasks similar to junior-level employees or interns. 

How AI Agents Work

An example of an AI agent working on your blog post is Chat GPT, which can automatically draft a blog post about a new recipe. The agent may ask for feedback and then publish it for you. This makes AI agents less of a thought prompt partner and more like an intern who takes on a mission on your behalf. There are two ways AI agents can take action on your behalf: making direct computer calls called API calls, or controlling your computer. API calls allow agents to make direct connections with tools like Salesforce, Calendly, Microsoft Teams, Google Calendar, and HubSpot. The second approach involves the AI controlling your computer, i.e., constantly looking at the screen and clicking buttons on it. Relay.App focuses on business productivity applications and automated calls, so it cannot log into your bank and perform actions on your behalf. However, a tool using the computer use capability would need two factor authentication and captcha.

AI Agent Interaction: Solutions and Problems

AI agents can interact with any website or tool that has an API, such as email browsers, CRM systems, and business productivity tools. There are three categories of AI agents: APIs, which perform tasks on a computer screen, AI in-built capabilities, and capabilities in reasoning. One problem AI agents need to solve is how they interact with their tools, such as reading and writing data from Salesforce, and how they can do this either via an API or by controlling the browser. Additionally, AI agents have the ability to extract information from PDFs, translate language information, turn text to speech, create videos automatically, and browse the web and do research.

Three Models of AI Agents

There are three models of AI agents to keep in mind: one class is a tool like relay.app, where the tool comes on your cloud or services, and interacts with things, while another class is a tool owned by the customer or freelancing agents. Models to consider when building AI agents: pre-built AI agents, which are commonly used in customer support tools, custom AI agents, and freelancing agents. These models are designed for specific vertical use cases and can be hosted on a platform or servers. AI agents can interact with various tools and platforms, including email marketing tools, CRM systems, and cloud-based versions of Microsoft tools. They can also perform tasks such as transcription, summary notes, and internet research.

Customer Service Agents

Customer service agents can be trained on a company's knowledge base. These agents can take various actions, such as replying to emails, triggering password reset emails, or issuing refunds. There are three main types of customer support agents: pre-built agents for specific use cases, custom built agents on easy-to-use platforms like lyndee.ai, relevance.ai, and Zapier, and engineers building their own agents using developer-focused frameworks. There are two options for building agents: one that interacts with APIs, and another that almost takes over your desktop. The fully browser-based approach is less reliable and predictable, but API-based approaches provide clearer guardrails for the agents.

Common Use Cases for Relay.App 

The most common use cases for Relay.app include email handling, calendar management, customer interaction and relationship management, and marketing content creation. Email calendar management involves extracting information from emails, summarizing PDFs, forwarding them to others, drafting or applying to emails, labeling them, and archived emails. Personal productivity use cases involve managing emails, scheduling meetings, and reminding people to RSVP. Customer interaction and relationship management involves researching prospects, sending personalized emails, creating contracts, and reviewing support tickets. Marketing content creation involves creating life cycle marketing campaigns, blog posts, LinkedIn posts, and Twitter posts. These are the big three use cases for AI agents. However, there are also a wide range of businesses with different bespoke use cases, so they can build AI agents to do custom tasks.

AI Agent Meeting Prep

Jacob shares the step-by-step process to using the AI agent. The agent will help prepare for meetings by providing information about the person they are about to meet. The next step is to add a trigger, which is what happens in the outside world that causes the agent to wake up and start working on our behalf. The trigger can be based on an event in an application or on a scheduled basis. In this case, the trigger will be when an event is upcoming in Google Calendar. The agent will then be able to check if any changes in the person's calendar are made and reschedule the meeting. The AI agent will then be able to look at all the meetings that the person has tomorrow, and the time picker will show that the trigger will happen daily at 5pm.

Using AI Agents to Find Events

Jacob introduces a new step in Google Calendar that allows users to find events and filter them based on criteria. The step is set up to find all events that match the specified criteria, such as start time coming after today or before today plus two. If no events are found, the system can either notify the user or continue with the day without meetings. The next step is to use an iterator to iterate over the list of events found in the previous step. The output of one step can be used as input for the next step, as it often references previous information. In this case, the list of events is the list of events found in the previous step.

Using AI Judgement

To use AI judgement, Jacob adds a step to an AI step, selecting Custom Prompt. This prompts the AI to provide detailed instructions and insert relevant data for context. For example, if everyone has responded to a meeting, Jacob sends an email stating "looking forward to seeing you tomorrow" and if not, an email asking if the meeting still works for everyone. A path is created, which allows the user to decide whether to proceed with the AI's task. The first branch of the path is everyone replied, where the user can choose which rules determine whether to go down that path. In this case, the user selects "everybody replied," which will send an email to the list of guests, stating "looking forward to seeing you tomorrow." The email can be written manually or sent by the user.

AI Agents' Primary Target Audience

The primary target audience is non-technical individuals who have never written code or used code tools. The goal is to make the product easy for everyone to use. The company offers a partner program with automation and AI experts and agencies to help businesses set up workflows or advise on AI usage. Jacob also provides YouTube tutorials that can be helpful for creating workflows. Once people watch tutorials and follow them, they can understand the process at each step. The hope is that everyone can create their own workflows without needing a partner. However, complex business needs may require assistance.The goal is to make the product accessible to everyone, regardless of their technical skills. The company also offers a partner program with automation and AI experts to help businesses set up workflows and advise on AI usage.

Relay.App Pricing

The pricing of the tool is a standard self-serve freemium model, with subscriptions based on how much users use the product. The free tier offers 200 steps and 500 AI credits, while the first paid tier is $9 a month, which includes more steps and AI credits. The team tier starts at $59 a month for companies using it with multiple people. As users use the tool more, it needs more steps and AI reasoning. Bundles of credits are available on top of the subscriptions.The typical cost for a typical business is between $100 and $200 a month. However, if users are getting value from the product, they will pay $100 to $200 a month. 

Finding Talent with Relay.App

On the Relay.App website, which features a gallery of common use cases, such as competitive research, pre-meeting dossiers, and email extracting. The app also includes a human-in-the-loop step where users can ask for help in identifying the right profile based on their information. This human-led approach can help AI agents make more informed decisions and ensure the effectiveness of their work. The app also includes a human-in-the-loop step, allowing users to identify the most suitable profile based on their information. 

Timestamps:

03:12: Capabilities and Limitations of AI Agents 

10:01: Interaction Models and Use Cases 

14:13: Building an AI Agent: Step-by-Step 

29:56: Advanced Features and Customization 

31:25: Pricing and Availability 

33:33: LinkedIn Profile Finder Use Case 

38:02: Conclusion and Resources

Link:

Company website: https://www.relay.app/

Unleashed is produced by Umbrex, which has a mission of connecting independent management consultants with one another, creating opportunities for members to meet, build relationships, and share lessons learned. Learn more at www.umbrex.com.