How to Steer Moveworks Copilot Responses

Overview

The Moveworks Copilot leverages advanced LLMs and machine learning models to offer conversational flexibility and enhanced reasoning capabilities. While the reasoning engine is globally optimized for broad, cross-environment performance, we understand that each business has unique requirements that a one-size-fits-all approach cannot fully address. Variations in preferences—such as plugin selection, response in specific situations, resource prioritization, and even tone and wording—are common among our customers. To meet these diverse needs, we developed a suite of steer-ability tools that empower customers to provide organization-specific, situational instructions to the Moveworks Copilot's reasoning engine, allowing it to be finely tuned to each organization’s specific needs.

How do Steerability tools work?

When a user submits a query to the Moveworks Copilot, it undergoes a series of steps designed to process and respond effectively. These steps include reasoning, planning, plugin selection, plugin execution, response generation, and ultimately delivering an answer back to the user. Throughout this lifecycle, the Moveworks Copilot makes key decisions at each step, guided by data and instructions. The data encompasses the user’s request, relevant content, and user attributes, while instructions are prompts that shape how the LLM approaches the request, based on current conditions (such as available plugins). The Moveworks Copilot uses this combination of data and instructions to plan its responses.

So where do the Moveworks Copilot steer-ability tools come in? By default, the instructions provided to the Moveworks Copilot across customer environments are standardized, ensuring optimal collective performance through extensive tuning for enterprise use cases. However, the Moveworks Copilot steer-ability tools enable us to customize these instructions at an organizational level, allowing the Moveworks Copilot to make decisions tailored to each customer’s specific business needs.

Currently, two types of instructions can be provided to Moveworks Copilot:

  1. Global Instructions: These instructions apply to all user queries across every step of the query lifecycle. They act as standing guidelines that Moveworks Copilot continuously follows. For example, you could instruct Moveworks Copilot to always respond with a tone similar to SpongeBob's.
  2. Situational Instructions: These instructions apply only to specific use cases or stages within the query lifecycle, activating only when particular conditions are met. For instance, you could instruct Moveworks Copilot to “Always call the Forms plugin when a user requests a laptop replacement.”

What types of response can be steered?

Below are a few examples that the Moveworks Copilot Steer-ability tools can help with.

Global

  • Tone of responses: Adjust Moveworks Copilot’s response style to be more succinct, lighthearted, or professional, based on organization’s preferences.
  • Terminology customization: Use preferred terminology for specific terms, such as referring to "Human Resources" as "People and Culture."

Situational

  • Steer the plugin selection: When a user request could be resolved by multiple plugins, specify which plugin Moveworks Copilot should prioritize. For example, if a software access request could be addressed either by the software plugin (for automatic provisioning) or by the forms plugin (for catalog items), you can direct Moveworks Copilot to always use the software plugin.
  • Prioritize specific resources: In some cases, you may want Moveworks Copilot to prioritize a particular resource over others for certain types of questions. For example, if an employee asks about medical benefits, you can specify the employee handbook as the primary resource Moveworks Copilot should reference.
  • Control information included in responses: Once a plugin is executed, the Moveworks Copilot summarizes information from the plugin response metadata to construct an answer that is helpful to the user’s question. If the default summarization doesn’t fully capture what’s needed, you can specify key information for Moveworks Copilot to emphasize in the response.
  • Enable advanced reasoning capabilities: In specific scenarios, you may want to enhance Moveworks Copilot’s reasoning and planning, allowing it to execute advanced tasks step-by-step. This is experimental and should be applied selectively on a case-by-case basis.

What are the limitations?

  • These steer-ability tools are still experimental and require extensive case-by-case testing for effective implementation. Performance can vary significantly depending on how each instruction is configured, with even slight wording changes impacting outcomes.
  • Providing instructions to the LLMs does not resolve all potential Moveworks Copilot issues and cannot guarantee a complete fix. LLMs are inherently dynamic systems that may still make errors.

How to use it?

Currently, steer-ability tools are available only to our internal customer success and support teams. We’re keeping these tools internal due to their experimental nature and high dependency on debugging and testing resources. Our teams are well-trained in using these tools, so if you have any use cases where they may be beneficial, please reach out to your customer success or support team.