AWS Partner Network (APN) Blog
Designing enterprise AI agents with DevRev guardrails powered by Amazon Nova
By: Len Gomes, Partner Solutions Architect – AWS
By: Ruhisar Tikoo, Technical Account Manager – AWS
By: Writom Guha Roy, Sr. Startup SA – AWS
By: Prateek Chaudhury, AI Engineer – DevRev
 |
| DevRev |
 |
Enterprises are embedding AI agents across workflows from customer support to IT service management but face the critical challenge of facilitating responsible behavior at scale. As enterprise AI agents gain autonomy and access to sensitive customer and product data, the demand for robust AI guardrails has increased significantly. With robust AI guardrails in place, organizations gain the confidence to scale AI agents across sensitive workflows knowing that personally identifiable information (PII) is filtered before it surfaces in responses, content stays on policy, business insights remain grounded in verified data, and safety controls hold firm against adversarial inputs. Responsible AI is an operational requirement for enterprises deploying AI agents powered by Amazon Nova or other foundation models (FMs) in production.
The business value is clear. Effective guardrails strengthen regulatory posture, reinforce customer trust, and reduce remediation overhead. For enterprises in regulated industries like financial services, healthcare, or telecommunications, these advantages compound quickly. Organizations that implement custom, domain-specific policies reflecting internal compliance rules gain a measurable operational edge: consistent enforcement around roadmap details, confidential integrations, and regulated industry language. Modern enterprise AI systems that enforce AI safety at the infrastructure level unlock greater agent autonomy with confidence.
DevRev guardrails: Built-in AI safety for every agent
DevRev has addressed this challenge directly by building real-time AI guardrails into its enterprise agent system, powered by Amazon Nova 2 Lite through Amazon Bedrock, a fully managed service from Amazon Web Services (AWS) for building and scaling generative AI applications with foundation models. Agents deployed on DevRev’s system come with a unified guardrail layer that evaluates user inputs before a response is generated, supporting reliable, adherent behavior across enterprise workflows right from the start. Teams can also author custom policies in natural language, making policy enforcement accessible without requiring code.
Amazon Nova is the AWS portfolio of AI offerings delivering frontier intelligence with competitive price performance. Nova is built on AI technologies originally developed for internal Amazon applications, including Alexa+, Amazon Ads, and Amazon Stores. Amazon Nova 2 Lite is a fast, cost-effective reasoning model featuring extended thinking with step-by-step reasoning (at three configurable levels: low, medium, or high), multimodal support (text, image, video, and documents as input), a one million-token context window for expanded reasoning, and built-in tools including web grounding with citations and code interpreter.
Early adopters report significant results: Siemens achieved a 300% improvement in search speed, Trellix saw a 39% accuracy boost in threat classification, and AWS Transform improved tool-calling efficiency by 60%.
Measured impact in production deployments
Production deployments of DevRev’s AI agent system demonstrate measurable impact across enterprise deployments where Computer agents handle support, operations, and knowledge management at scale, as shown in the following table.
| Operational metric | Production result |
| Support ticket automation | 85% (compared to 40% industry average) |
| Support cost reduction | 50% (compared to 20% industry average) |
| Employee time saved per week | 10-plus hours per employee |
| Guardrail evaluation latency | 247 ms average (<300ms threshold) |
The DevRev guardrails solution
DevRev built AI guardrails as a built-in layer within its agent system not as a bolted-on filter, but as an integral component of the execution pipeline that evaluates every user message before a response reaches the end user. This built-in protection covers the most critical enterprise AI issues:
- Prompt injection is designed to alter agent behavior through crafted inputs.
- Sensitive content is evaluated, filtering PII before it surfaces in agent responses.
- Toxic and off-policy responses are blocked, filtering harmful or nonadherent content before delivery.
- Custom domain policies restrict topics such as internal roadmaps, confidential integrations, or regulated terminology.
Administrators configure custom policies using plain natural language rules like “Avoid discussing internal roadmap details” or “Do not reference competitor products” without writing code. After these policies are attached to an agent, they enforce consistent, auditable boundaries regardless of where the agent is accessed across the organization.
How it works: The parallel evaluation architecture
The key design principle behind DevRev’s guardrail system is concurrency. Rather than inserting a sequential safety check that adds latency to every interaction, DevRev runs guardrail evaluation in parallel with the primary agent reasoning loop, achieving near real-time responses while maintaining strict policy enforcement.
Each interaction flows through the system in five steps:
- Input received – Each user message simultaneously triggers the guardrail model and the primary agent large language model (LLM).
- Parallel evaluation – The lightweight guardrail model classifies the input against the active policy set default and custom while the primary model begins generating a response.
- Gated output – The response stream is held until the guardrail model returns its decision.
- Approved path – If the input passes evaluation, precomputed portions of the response release instantly to the user, and the remainder streams subsequently.
- Violation detected – If a violation is flagged, the system aborts the pending response and streams a safe, policy-defined fallback instead. The violation is passed back to the agent reasoning loop for handling.
This concurrent evaluation architecture is what enables sub-300 ms guardrail performance without degrading the conversational experience. The flow is shown in the following graphic.
Figure 1: Parallel evaluation flow for real-time guardrails
Why Amazon Nova 2 Lite?
DevRev’s guardrail model runs on every single user message, in parallel with the primary agent LLM. This means the model must be fast, accurate, and cost-efficient at scale simultaneously. After evaluating multiple models available through Amazon Bedrock, DevRev selected Amazon Nova 2 Lite as the best fit for production-grade enterprise safety enforcement based on three deciding criteria:
- Speed – Amazon Nova 2 Lite achieves a 247 ms average evaluation time, which is fast enough for real-time guardrails without disrupting user experience. This meets the critical threshold for maintaining conversational fluidity at enterprise scale.
- Classification performance – Amazon Nova 2 Lite delivers strong accuracy in detecting prompt injection, flagging sensitive content, and enforcing custom natural-language policies, the exact workload DevRev’s guardrail layer requires.
- Cost efficiency – Per-interaction costs remain low across thousands of concurrent agent sessions. Built-in availability on Amazon Bedrock also simplifies scaling and monitoring within DevRev’s AWS based infrastructure.
System context: Computer by DevRev
DevRev is a modern enterprise AI system that unifies fragmented enterprise data through Computer, an AI assistant designed to go beyond chatbots that merely read and summarize. Computer can create, update, and delete records across systems such as Salesforce, Jira, and Zendesk. Understanding the architecture explains why guardrails are not optional: they’re essential infrastructure for agents operating at this level of autonomy.
Computer is powered by two core technologies:
- Computer Memory – DevRev’s permission-aware knowledge graph transforms structured and unstructured data from across the enterprise into a dynamic network of interconnected information, maintaining data permissions, context, and compliance requirements throughout.
- Computer AirSync – A real-time synchronization engine that continuously synchronizes data bidirectionally across systems, alleviating silos and keeping connected systems aligned.
The following graphic shows DevRev at the center of an interconnected architecture that unifies enterprise data across multiple systems such as Jira and Teams.
Figure 2: Computer by DevRev unifies enterprise data across multiple systems
Computer operates as a true AI teammate across the enterprise, finding answers and reasoning across business context. It acts through AI agents that create tickets, update records, route work, and automate tasks using Computer Agent Studio, delivering enterprise-grade governance through built-in security, permission awareness, and compliance controls. DevRev’s Computer Agent Studio provides an intuitive, low-code environment for teams to design, test, and deploy AI agents. Each agent moves through a defined lifecycle (from design to simulation, validation, and finally publishing) and inherits identity and permissions like a human user, limiting access to only authorized data and operations. This is the execution environment in which guardrails operate: a deterministic, observable, permission-bounded runtime where safety enforcement is built in at the system level.
The following graphic shows Computer executing a multistep ticket reassignment in which it updates eight records, applies escalation tags, and logs changes across multiple systems, including Salesforce and Slack.
Figure 3: Computer executing a multistep ticket reassignment
The following screenshot of the Build dashboard shows the agent configuration interface for a Developer Evangelist agent, including modular sections for knowledge sources, skills, and guardrails alongside a live chat preview panel.
Figure 4: The Build dashboard showing the interface for a Developer Evangelist agent
Enabling safe agent autonomy at enterprise scale
When enterprises can trust that every interaction is evaluated against policy before reaching the user, they can confidently expand the scope of what agents are allowed to do, taking on more complex workflows, accessing more sensitive data, and operating with less human oversight. AI safety, properly implemented, makes greater autonomy possible rather than constraining it.
The production results reinforce this value. When agents resolve 85% of support tickets automatically and save employees more than 10 hours per week, efficiency gains only hold if the underlying AI operates within defined boundaries. By building real-time evaluation directly into the agent execution pipeline powered by the 247 ms average evaluation time of Amazon Nova 2 Lite, DevRev demonstrates that speed and safety can coexist without sacrificing either. For organizations exploring enterprise AI adoption, deploying AI agents is a must, and their challenge is to choose a system that can best enforce the policies, permissions, and AI safety boundaries required by production deployments.
Conclusion
DevRev’s guardrails capability powered by Amazon Nova 2 Lite through Amazon Bedrock demonstrates that enterprise AI safety and operational performance coexist comfortably. With real-time policy evaluation running at 247 ms average latency, natural-language policy authoring, and parallel execution that preserves conversational fluidity, DevRev gives enterprises the foundation to deploy AI agents confidently and at scale. The combination of Computer’s permission-aware architecture with the classification performance of Nova 2 Lite demonstrates how responsible AI and operational efficiency can scale together.
To get started, request a demo and see DevRev’s agent system and built-in guardrails in action. Visit DevRev’s homepage to learn about Computer by DevRev and check out the DevRev blogs for technical articles and updates. Watch Building safer agents at DevRev, a live session with DevRev CTO Ahmed Bashir, for a comprehensive walkthrough of guardrail implementation and real-world use cases.
.
DevRev – AWS Partner Spotlight
DevRev is an AWS Partner that provides an enterprise AI system unifying customer support, product development, and business operations through Computer, an AI agent system that executes work across enterprise systems while enforcing safety policies with built-in guardrails powered by Amazon Nova 2 Lite on Amazon Bedrock.