Taking AI Beyond Chat and Into Something More Useful - Agents and Assistants
One Year Ago Today We Were Introduced to ChatGPT. Things Have Advanced Significantly Since Then. If You Aren't Using AI Agents, You Aren't Tapping Into The AI Power Available.
The State of AI One Year After the Introduction of ChatGPT
The world of Generative AI is rapidly advancing. Each technological advancement helps build the next advancement. This cumulative effect is driving the AI evolution at a breathtaking pace. Exactly one year ago today we were introduced to ChatGPT. The current state of the AI industry has advanced significantly in that one year.
The irony of ChatGPT in its original form is the intense power of the LLM1, which is severely constrained by the limited use cases of a chat interface. In other words, the current technology offers far more in terms of power and capabilities than we have been accessing. Recent announcements by several of the large players in this space such as AWS have been seeking to extend the reach and capability of the LLM to novel use cases.
So what does “advancement that would allow novel use cases” mean to you and your use of AI? This article looks at what this advancement means for the overall AI landscape and how consumers of AI can benefit from agents.
What is an Agent?
The first issue is what is an agent? If you’re a lawyer like me, your mind quickly jumps to things like respondeat superior, vicarious liability, and other legal terms. That’s not too far off the mark. In the context of this discussion (and more broadly in the AI space) an agent is an application or component of an application that is intended to autonomously perform an action when requested. In other words, the agent is acting on your--the user’s--behalf. Agents orchestrate interactions between foundation models, data sources, software applications, and user conversations, and automatically call APIs2 to take actions, and leverage knowledge bases to supplement information for these actions. Really, it’s a lot cooler than it sounds.
You may have heard of RAG3 or may hear about it from a legal AI provider who extolls its benefits. LLMs are trained on data, but that data doesn’t continually get updated. RAG is an AI generator model which is more narrowly focused on text generation fueled by your specific, existing writings. Using agents expands on the capabilities of RAG. The RAG model focuses narrowly on text generation by tapping into the information of existing writings and documents it was trained on, along with a specific “database” of relevant information. The agent model then enables actions based on data and reasoning.
RAG vs. AI Agent — What Is the Difference?
To understand the difference between RAG and an agent, imagine that you are a partner at a law firm. You have a heavy caseload and a big trial coming up, so you need more junior attorneys, paralegals and legal assistants helping you out with tasks like synthesizing information from depositions, researching statutes and caselaw, proof-reading motions, and even perhaps making sure a motion is timely filed.
A RAG model is like a junior attorney or paralegal that has access to all the documents and information in the case. The RAG model can read through all the documents to determine what legal issues need to be included in a motion or brief. It can begin drafting the brief. It can review discovery documents and suggest questions for a deposition. It might even come up with unique analogies or creative arguments. You can also direct the RAG assistant to only use a limited set of information, legal cases, or material to perform its tasks.
In contrast, an AI agent is more like an apprentice attorney learning from the master--you. Through ongoing conversations, AI Agents incorporate human feedback and preferences to improve their abilities. Over time, this allows the AI Agent to develop more complex behaviors not just derived from pre-written legal books or texts, but also from understanding nuanced human norms and goals.
Agents are a more robust and directed solution. Here are some examples of how AI models have been used or can be used in business:
An agent that reviews customer feedback forms and sends a standard email to the response team if an overly negative review is received.
An agent that receives an electronic intake form from a potential client, categorizes it by type of law involved, stores it in a conflicts database, and then sends an email with the intake form attached to the responsible attorney.
An agent that retrieves data from a CRM database when the LLM is asked to create a welcome message to a new customer, or a termination of representation letter.
An agent searches files in a directory and organizes them based on content.
All these examples have been possible for years using typical software programs. The key differentiation and advantage of using an agent is the ability for automated advanced reasoning. Words matter a great deal to lawyers, and we are so picky about the meaning of every word. So it feels a bit awkward to say “automated advanced reasoning” (how can an inanimate object reason?). But that’s what I mean.
Using an agent allows the database to be integrated with the LLM, by way of an API, and then given certain tasks to perform. Remember the API is simply the software that links the LLM to other applications. This allows for more advanced and specific skills and abilities.
Here are some typical capabilities that are possible by tapping into the technological genius of AI Agents:
Extend the LLM by organizing a requested task into discreet steps.
Gather additional user input, and correct actions based on feedback.
Make API calls to your systems to complete tasks.
Augment available data by querying data sources.
Take actions to complete a task.
Annotate sources for attribution.
Current examples of Agents include the OpenAI Assistants that were featured earlier this month by OpenAI and Bedrock Agents that were recently announced by AWS. So far I’ve only worked with the Bedrock Agents, but wow, they are phenomenal.
Why This Is Valuable To You
Hopefully you’ve gotten an idea of how useful AI Agents can be. In case you haven’t, I’ll spell it out a bit further. AI models can make your job easier, and faster. They can come up with the deposition summary, point out contradictions in testimony, or give you a rough draft of a complaint.
AI Agents, on the other hand, can revolutionize your firm’s or business’s workflow when used correctly. You give the AI Agents a set of “lambda functions” and then they can take the information from an intake form, draft the complaint, draft the summons, email a draft to the client to review and get sign-off approval, figure out which court to file the complaint in, then get the complaint and summons filed4 and paid for, and make arrangements to have copies of the summons and complaint served, with a copy to the client. It can even look up the addresses of the party opponents, and determine if a waiver of service, electronic service, service in person, or service by certified mail is more appropriate. You can look over the complaint before it’s filed, if you’d like, and make sure all the correct claims and legal defenses are there. The AI Agent can then calendar future deadlines and add to or adjust those dates based on a later-issued scheduling order. Yep. All that and more.
Regardless of your approach to generative AI or the focus of your interest, agents and assistants are the latest evolution of LLM-driven capabilities. Understanding how these tools work is fundamental to knowing how you can apply AI to solve specific problems, whether simple or sophisticated. Agents create a multi-prong approach to tasks, combining the latest LLM capabilities with traditional software features.
For More . . .
If you want more information check these out:
https://platform.openai.com/docs/assistants/how-it-works
https://docs.aws.amazon.com/bedrock/latest/userguide/agents.html
An LLM or Large Language Model is a more specialized type of AI model.
An API or Application Programming Interface is the computer code that allows two different applications to function together. Think of it as a contract between two parties. For each application/party, there must be a separate API to link the two programs/clients.
Retrieval Augmented Generation is a program whereby you can insert or inject data into an LLM. The data you inject can be anything from the newest caselaw in the relevant jurisdictions, to your firm’s discovery materials in a case.
Ok, this task is only possible if the court has an API, but I believe we’re not far from that.