Agentic AI is grabbing headlines these days, and justifiably so. As the industry moves to the next stage of deriving true business value from the generative AI advances of the last two years, agentic AI is at the forefront—and Moveworks has been paving the way as an early leader on this agentic front since the launching of our Copilot in February 2024.
We recently shared a great explainer on agentic AI — it’s a hot topic in the AI world and we wanted to build on the subject and share a look behind the scenes as our own Copilot uses it.
Agentic AI is an AI system that can act via autonomous “agents” to achieve complex goals on behalf of users. In this framework, a combination of ML models is used to create an AI that can independently create and execute a reasoning-based plan that can achieve the user’s goal instead of being limited to simply following programmed logic or dialog flows.
An agentic AI Copilot that can reason is much more capable of helping employees throughout the day with a plethora of tasks. Similar to the role a personal assistant would play, this AI, too, has the capacity to understand challenges and problem-solve on your behalf. State the challenge you’re facing and Copilot can do the diligence and come up with the best solution for the situation — that’s the value of reasoning in action.
The Moveworks Reasoning Engine
A reasoning engine powers the AI to meaningfully achieve an employee’s goals. It does so with abilities to:
- understand a user's objective
- develop a plan to achieve that objective
- execute function calls according to its plan
- evaluate the success of this execution
- iterate on the plan until it successfully achieves its original objective
Importantly, a reasoning engine is able to check back with the user wherever necessary to seek additional context about the goal, confirm its direction, and take user input for adapting the plan on the fly.
In some ways, it attempts to mimic how human beings solve problems, which is why having a reasoning Copilot is so powerful.
The Moveworks Copilot’s reasoning engine is an agent-based AI system that showcases many of these human-like problem solving capabilities. Our engine uses reasoning, decision-making and plugin calling capabilities of multiple LLMs to accomplish specific elements of the bullets we touched on above.
And from a universal support perspective, our reasoning engine scales to reliably operate within the specifics of any of our customers’ unique environments (i.e. system integrations, available plugins, business logic, processes, etc.).
Success for the Moveworks Copilot means our reasoning engine is able to create business-specific plans using a distinct set of available tools and processes, and consistently execute whatever the task may be in a goal oriented way—for millions of users across the world, and all in real-time. Let’s talk further about how that happens.
Reasoning in action
When users interact with Copilot, their queries, questions, or conversations are run through a reasoning process that informs how it responds back to them. For easy understanding, we can think of that process unfolding in three distinct steps: understanding, planning, and executing.
In each step, we've fine-tuned the Reasoning Engine’s various LLMs to apply the best reasoning and decision-making techniques. This approach aims for each employee interaction to be met with a thoughtful, high-quality response.
Understanding user needs
The first stage, understanding, begins as soon as a query is received. The first step is to enrich the text of the query with extracted metadata that serve as critical signals for the reasoning process. The reasoning engine uses a collection of machine learning models, including LLMs, to determine properties of the utterance such as topic, intent, sentiment, domain and language.
Next, it establishes the query in the specific business context of the customer by leveraging the Moveworks knowledge graph to map entities mentioned by the user to their standard or canonical forms. This ensures that all the components of the Moveworks Copilot are able to understand the variations of how an entity is referenced, including very unique and organization-specific entities that only exist within the customer’s environment. In this stage, the reasoning engine also checks for whether the query is toxic or offensive and prevents such requests from being served and entering the system.
The next important piece of understanding is conversational context. The reasoning engine incorporates previous interactions between the user and the bot, which provide a rich vein of situational information to help identify the most effective solution. This includes identifying whether the user is starting a new topic or continuing a previous one, deciphering implicit references such as “it,” “that,”, and so on to understand follow-up questions and to check what other resources and options were previously offered to the user.
From the combination of all of these inputs, the reasoning engine rewrites the user’s query and is able to piece together a complete and well-defined problem statement to solve for.
Planning for success
The next step is planning a course of action. This is where the reasoning engine maps its understanding of the problem and the prior context to select the best plugins or tools to serve the request. You can think of plugins as specialized tools custom-built to handle specific use cases and integrate with certain systems.
For example, plugins may reset passwords in Okta, provision users in Google Workspace, onboard staff in BambooHR, manage access in AccessHub, deploy servers, update Salesforce, pull reports from analytics platforms, and sync product data between ERP systems. The list goes on and on. Each organization has a unique set of plugins, based on their chosen configuration and the custom plugins they have developed using the Creator Studio extensibility platform.
The reasoning engine reviews the list of available plugins and quickly filters the plugin catalog down to the most relevant options with the previously extracted information. It then enters what is called a plan-evaluate loop. In this loop, the reasoning engine uses its planning module in conjunction with an autonomous evaluator module.
The planner and evaluator rapidly iterate on approaches using Chain-of-Thought to break down the problem using the available plugins, and this independent iteration produces a high quality first pass that eliminates wasted effort and unnecessary burden on the user to guide the reasoning engine. From a performance perspective, we have also optimized this to be very low latency to ensure that the plan can be charted out quickly.
But how do we know that a given plan makes sense in the business context? This is enforced with the use of policies in the reasoning engine, which provide guidance on how to think about the problem in the context of employee support, and to lay out the standard objectives and best principles for how an employee support assistant should work. This imbues the reasoning with a form of grounding that focuses the solution space and allows the planner and evaluator a way to determine the quality and efficacy of a candidate plan.
Executing with purpose
The reasoning engine is now ready to start executing its identified plan. Just like a human problem solver, it is able to attack the problem but pays close attention to the outputs at each stop, so that it can adapt the plan, if required.
It invokes the plugin executor component, which is responsible for making the appropriate function calls to each plugin selected in the plan and for monitoring the responses sent back. As the plugins perform their specific tasks, they return results or error messages, which the plugin executor analyzes and sends to the reasoning engine, which determines what to do next.
At every juncture, Copilot analyzes the latest available information and decides whether to summarize the output as a final response, or to ask the user for more information, or to try another approach by calling other plugins. This iterative approach ensures that the reasoning engine dynamically tries to solve the problem in a variety of ways without any human intervention. This flexibility is built into its core reasoning approach, mimicking how humans would handle similar situations.
As an example, for a complex IT troubleshooting scenario, the plan could involve using search plugins that scour knowledge base articles or files, or serve pertinent support forms, or workflows that can guide the user through a designed resolution process. Therefore, we can see that the reasoning and planning processes work in lockstep to remain focused on the user’s goal and find the fastest path towards providing a meaningful solution.
All these steps are taken by the reasoning engine and plugins in a seamless, autonomous way, in a matter of seconds. The reasoning engine produces user-facing responses with a concise, referenced summary of the most relevant findings. What’s more, at every step the reasoning engine continues to provide the user with reasoning steps for explainability, which also helps the user understand what steps it has tried, and how the response was generated.
If none of the available plugins can help the user, Copilot enables the user to easily seek human support by filing a ticket or speaking with a live agent.
Better reasoning, better support
The Moveworks Copilot puts powerful AI directly at the fingertips of employees to aid them in their day to day. The Reasoning Engine is able to perform the same diligence as a human agent when it comes to comprehending a user's issue, developing a plan to fix it, and executing the appropriate actions to do so with the same degrees of success—only AI can do it faster, for cheaper, and around the clock.
Ready to see how reasoning can reshape how your employees work? Get a demo today and watch AI problem-solve employee support challenges live.
See the Moveworks Reasoning Engine in action. Sign up for a personalized demo here.
Table of contents
