Moveworks Policy Validators: Agentic AI, Built for Compliance

As enterprise technology continues to evolve, companies are turning to AI agents to streamline business processes, boost productivity, and enhance user experience. Yet, the promise of AI agents operating reliably within the bounds of business rules remains a challenge.

Simply feeding text instructions into an LLM and expecting it to reliably understand and enforce those instructions is not a viable solution. LLMs are inherently susceptible to misinterpretations and can even be manipulated by malicious users to bypass intended restrictions. 

Moveworks' Agentic Automation Engine, and specifically its Policy Validators feature, are designed to address this challenge by enabling AI agents to be trained to adhere to business rules. Policy Validators is able to provide a robust mechanism to translate deterministic business rules into actionable steps for AI agents and enable agentic AI compliance. 

Counting on natural language to enforce rules can be risky

When you’re building AI agents to live within your business – they have to be able to follow your organization’s policies. Just because it’s “powered by AI,” doesn’t give it an excuse to follow business rules only 70% of the time. You want to make sure that they are always able to be enforced.

Unfortunately, many AI agents today have a primitive approach to policy validation – they feed text instructions into an LLM and expect the LLM to figure out how to enforce it. 

This LLM enforced approach generally fails to be able to validate your specific policies. Just because an LLM can follow instructions, doesn’t mean they should be your system for enforcing instructions. LLMs are subject to misunderstanding your instructions, or worse, a malicious user could jailbreak past the LLMs instructions.

Let’s take a simple example – booking a room. Pretend your organization has a special policy:

In AI agents today, you have to write this as a text instruction: 

You are at the mercy of the LLM to understand “if the facility has an executive business center” and enforce it reliably. You might get lucky, but realistically, you’re probably going to get a chat experience like this:

To get this to work as expected, your developers will have to become prompt-tuning experts . Even in ideal conditions, this will probably only work ~70% of the time. That’s not okay when it comes to building critical plugins like:

• Generating sales quotes – but rejecting requests below a minimum user count
• Submitting expenses – but requiring an itemized receipt or list of attendees over a limit
• Resolving a ticket – but requiring a summary of how the ticket was resolved

Introducing Moveworks Policy Validators

I believe we should reject the notion that an LLM following instructions is the right way to enforce your business policies. I believe that Policy Validators offer a far superior approach.  Policy Validators are simple to use, whether you’re an Excel whiz or a systems engineer, you’ve written an “IF” statement before. Good news – all a developer has to do is specify a rule in our friendly condition syntax, and the conversation is able to naturally follow the requirements. See the example below:

First, a Slot Resolver that converts “Guardian conference room” into an “OfficeSpaceFacility” object. If you missed our last blog on Slot Resolvers, check it out here.

Then, our Agentic Automation Engine validates that all requirements are met. In this case, it noticed that there was a Policy Validator on the “account” slot. It noticed the OfficeSpaceFacility had an Executive Business Center listed in the amenities.

Finally, our Reasoning Engine used this deterministic check to request the name of the customer from the user. Amir was able to reliably & accurately book the room.

When deploying AI agents, you don’t have to choose between doing things fast and doing things right – you can do both. Moveworks Policy Validators make that possible.

Trust, but verify – without burdening developers

This robust approach to policy validation allows developers to build AI agents that reliably operate within the bounds of business rules. Instead of relying on potentially fallible natural language instructions for LLMs, developers can use Policy Validators to clearly define and enforce rules with a high degree of confidence. This shift eliminates the need for extensive prompt engineering and reduces the risk of AI agents misinterpreting or circumventing policies.

Policy Validators simplify the process of translating complex business logic into actionable steps for AI agents. Developers can express policies in a clear, concise format that the Agentic Automation Engine can readily interpret and enforce. This streamlined approach can potentially lead to faster development cycles, help reduce tech debt, and can increase confidence in the reliability of deployed AI agents.

Agentic AI built to enable compliance

As businesses continue to adopt and integrate a multitude of software applications, the need for AI agents capable of navigating this complex landscape becomes increasingly critical. Policy Validators are important for enabling these AI agents to operate across diverse systems while adhering to organizational policies. By providing a reliable mechanism for overseeing business rules, Policy Validators pave the way for the development of truly versatile and trustworthy AI agents.

With Policy Validators as a core component of the Agentic Automation Engine, organizations are able to build AI agents that can act as a unified interface to their entire business ecosystem. These AI agents can intelligently interact with various systems, automate tasks, and provide valuable insights, all while operating within the defined boundaries of organizational policies. This capability is able to empower businesses to better streamline operations, enhance productivity, and unlock the full potential of AI across their organizations.

Want to dive deeper into the Agentic Automation Engine? Why not read our white paper to learn more.