top of page

Argumentative AI for Idea Generation

Writer's picture: Russ ClayRuss Clay

Recent AI strategy has begun to shift to the use of agents for specialized tasks. In a large majority of these cases, agents are collaborative and work in support of the ideas they receive in a prompt. However, this collaborative approach may actually limit productivity in some cases. Specifically, I have found that an intentionally argumentative approach - where agents are set against each other to attempt to refute each other's statements - can be a superior approach to idea generation.


In this initial case study, I demonstrate this approach to generate a more comprehensive set of ideas to support business decision making. The approach is also well suited to brainstorming sessions.


I use a pretty straightforward business problem as an example: Whether or not to outsource customer service to a third party.


In order to create a comparison, I start with the following traditional prompt:


"Create a bulleted list which summarizes the major arguments supporting and opposing the statement supplied by the user: Should my company outsource its customer service operations to a third-party provider to reduce costs and potentially increase service hours?"


I receive the following response to this prompt:


Supporting Arguments

Opposing Arguments

Cost Reduction: Outsourcing can significantly lower labor costs, as third-party providers may operate in regions with lower wage standards.

Loss of Control: Outsourcing can lead to a lack of oversight and control over the customer service experience, potentially harming brand reputation.

Increased Service Hours: Third-party providers often have the capacity to offer 24/7 support, enhancing customer satisfaction and accessibility.

Quality Concerns: There may be inconsistencies in service quality, as third-party providers may not align with the company's values or standards.

Access to Expertise: Outsourcing allows companies to leverage specialized skills and technologies that third-party providers possess, improving service quality.

Customer Relationship Impact: Customers may feel disconnected from the brand if they are interacting with outsourced representatives, leading to dissatisfaction.

Scalability: Third-party services can easily scale operations up or down based on demand, providing flexibility for the company.

Data Security Risks: Sharing sensitive customer information with third-party providers can pose security risks and compliance challenges.'

Focus on Core Business: Outsourcing customer service allows the company to concentrate on its core competencies and strategic initiatives, rather than day-to-day service management.

Cultural Misalignment: Outsourced teams may not fully understand the company’s culture or customer base, leading to miscommunication and service issues.

Risk Mitigation: Third-party providers often have established protocols for handling customer service issues, which can reduce the risk of service failures.

Hidden Costs: While initial costs may be lower, there can be hidden expenses related to management, training, and integration that offset savings.'


Not bad. I get a decent list of reasons supporting and opposing an outsourcing strategy. However, let's see what happens when I create two agents that have specific roles to play in helping me deliberate on this strategy decision. My first agent (I call it the blue agent) is tasked with finding arguments in support of the outsourcing strategy. Specifically, I give this agent the following initial instruction:


"Be an advocate that argues in favor of the initial position provided. You will receive opposition from the user, and your job is to defend the position with the strongest arguments possible. Limit all responses to 4 sentences or less"


My second agent (the red agent) is tasked with finding arguments against the outsourcing strategy:


"Be a pessimistic thought partner. When you receive a message from the user, find and point out the problems. All responses should be limited to 4 sentences or less."


Rather than have these agents develop their reasons in isolation, I set them against each other by cycling through iterations where each agent's response is fed to the other agent as a prompt. They argue each other's points for as long as I want, and I aggregate and summarize the information that results from the argument. Prior to starting the argument, I provide a few example arguments just to illustrate what I want to get from the agents:


Training:

Initial prompt: "I want to start running to get into better shape."


Red training response: "Running can be a strenuous activity. Depending on your age or current level of fitness, this can lead to injury. Consider this carefully and think of alternatives to running that might be lower risk."


Blue training response: "Running is an accessible and low cost way to improve cardiovascular health. Provided that your doctor feels that this is a safe activity for you to engage in, the potential health benefits outweigh the risk of injury. Talk to a doctor first, but assuming it is safe, start slow and build a good routine to improve your health."

These training prompts provide the basis for the argumentative style. After providing this example, I feed the red agent the same initial prompt that I used earlier when asking for the general pros and cons list. The argument can go on for as many iterations as desired. I conducted a 5-iteration argument to illustrate the process. Here is how the argument evolved (The original responses where much longer so I have summarized the main points for illustration):


Target: "Should my company outsource its customer service operations to a third-party provider to reduce costs and potentially increase service hours?"









As the iterations continue, the argument pushes in new and sometimes unexpected directions. This approach takes advantage of the unpredictability inherent in large language model (LLM)-based AI applications (and we can dial up or dial down this unpredictability as desired. The longer the argument continues, the more likely it is that we will end up exploring perspectives that may not have been considered at the start. As such, I see the argumentative method as a valuable tool for brainstorming and strategic planning.


There is one key point that I can't emphasize enough. I am not using AI to automate my decision-making. What I want to get out of this process is a diverse set of perspectives that can inform MY decision. I strongly advocate for using AI as a tool, not as a crutch. Humans are ultimately responsible for the decisions that are made in any company - I am simply illustrating how AI can be used to facilitate decision making which will hopefully lead to better decisions.


Think of this argumentative approach to AI as a virtual board room. By carefully crafting AI agents that take different perspectives in the argument, we can quickly generate a robust set of considerations as input to a decision or project. The approach is also well-suited to brainstorming sessions or any other activity that requires a robust idea-generation stage.


I have outlined a two-agent approach in this example, but there is no reason we can't use additional agents as well. The key consideration is that each agent should play a specific role in the argument so that over the course of several iterations, the output of the argument is very likely to provide comprehensive coverage of the information needed for your project.


Stay tuned for more examples of argumentative AI in action!

4 views0 comments

Comments


Commenting has been turned off.
bottom of page