OpenAI Teases AI Agent in Paris

OpenAI presented an AI agent, based on GPT-4o, during a live session at the GenerationAI Conference in Paris, demonstrating its reasoning and user interaction capabilities for handling complex tasks.

During the demonstration, Katia Gil Guzman, a member of the developer experience team at OpenAI, illustrated how it could assist with resolving customer service issues, such as handling a return for a Galaxy S23 smartphone. 

https://youtu.be/KsBT5G91fzI

Unlike traditional AI systems, the agent generated a dynamic, interactive user interface (UI) based on structured outputs to simplify the resolution process.

“I have a problem with my Galaxy S23 and want to return it,” Guzman said, starting with the problem statement.

The assistant fetched purchase details, leveraged decision-tree reasoning, and analyzed policy details. It also made a tool call to o1, which helped determine the customer’s eligibility for a replacement valued at up to 110% of the original price.

Unlike GPT-4o, which provided an incorrect outcome due to policy complexity, o1 accurately assessed the issue and navigated the policy tree to provide the correct resolution.

The demo also explored o1’s ability to recommend replacement products and assist in purchasing decisions. For instance, when the Galaxy S23 was unavailable, o1 suggested alternatives, compared specifications like RAM, CPU, and camera capabilities, and facilitated decision-making through custom-generated visuals, such as bar charts and tables.

Moreover, in a simulated phone call, o1 coordinated with a store to confirm the availability of a preferred replacement model, showcasing its ability to integrate real-time communication tools. “What we saw is an AI agent capable of reasoning and decision-making in complex customer service scenarios," said Guzman.

On the first day of the '12 Days of OpenAI', the company released the full version of o1 and a new $200 ChatGPT Pro model. The ChatGPT Pro plan includes all the features of the Plus plan and access to the additional o1 Pro mode, which is said to use ‘more compute for the best answers to the hardest questions’. Furthermore, the plan is set to offer unlimited access to o1, o1-mini and GPT-4o along with the advanced voice mode.

[Update] December 11, 2024, 17:09 | The article previously misstated the details of OpenAI's demonstration. The assistant showcased during the GenerationAI Conference in Paris was built using GPT-4o and incorporated a tool call to o1 for analyzing a customer service policy. It is not an "o1 agent" nor a preview of a new product, but an example of what can be built today using OpenAI's existing offerings. Additionally, Katia Gil Guzman is part of the developer experience team at OpenAI, not the head of the department.