Get Audit Answers in About Half a Minute: Introducing Kiteworks AI (KAI)
Every compliance team knows the pattern. An auditor, a privacy officer, or an executive asks who accessed a file, which policy fired, or who shared a document outside the organization. The answer exists somewhere in the logs, but getting it means a manual report, an analyst's time, and a wait that often stretches into days. Meanwhile, the business keeps adding AI to the stack, and the people who own risk are asked to approve tools they cannot audit.
This post introduces Kiteworks AI (KAI), the AI agent built into Kiteworks for the people who govern sensitive data. You will learn what KAI does for admin and compliance teams on day one, why a read-only design that runs on your own model is easier for a security team to approve, how KAI fits into the Kiteworks Agents workflow control, and what you need to get started.
Executive Summary
Compliance and security teams spend hours on questions that should take minutes, while AI agents spread through the business without a governed path to sensitive data. KAI gives admin and compliance users a faster, more verifiable way to work. They ask a question in plain language, get an answer drawn from their own activity and audit logs, users, files, folders, and risk policies, and see every lookup KAI made along the way. They can also ask KAI to draft a risk policy, review the draft, and apply it themselves. KAI is read-only, runs on the customer's own Claude model through the customer's own AWS Bedrock account, and keeps conversation history on the customer's appliance. Kiteworks customers can see it in a demo.
Key Takeaways
- Audit questions should not require a reporting project. KAI answers plain-language questions about activity and audit logs, users, files, folders, and risk policies, and a first answer takes about half a minute.
- Every answer shows its work. Show work lists each lookup KAI made, so reviewers can verify an answer instead of taking it on trust.
- KAI drafts and people decide. KAI can draft a risk policy, such as one that restricts external sharing of CUI, and an administrator reviews and applies it.
- The design is built for security approval. KAI is read-only, works within the access controls of the signed-in admin through Kiteworks APIs, is limited to System Admin and Compliance roles, and has its settings changes recorded in the audit log.
- You own the model and the data path. KAI runs on your Claude model in your AWS Bedrock account, and conversation history stays on your appliance.
Why Audit Questions Take Days When the Answers Already Exist
Most regulated organizations already log what happens to sensitive data. The slow part is turning those logs into an answer someone can defend. A compliance officer asks a question in business terms: which users downloaded the most files last week, or which risk policies fired for a specific person. An analyst then translates that question into filters, event names, and date ranges, runs the report, checks the output, and writes it up. Each answer becomes a small project. Frameworks such as NIST SP 800-171, the HIPAA Security Rule, and SOC 2 all expect organizations to review and report on audit activity, not merely collect it.
That friction compounds. Audit season brings dozens of requests at once. Incident response reviews arrive without warning. Privacy officers field questions from legal, HR, and executives who want a straight answer, not a spreadsheet export. Every hour spent translating a question into a query is an hour not spent on the work that actually reduces risk.
At the same time, pressure to adopt AI keeps growing. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. Security and compliance teams are the ones asked whether those agents are safe, and they cannot answer if the agent is a black box.
Ask. Verify. Act. What KAI Does on Day One
KAI lives in the “Ask KAI anything” box at the top of the Governance Center in the Compliance Console. There is no new console to learn and no new tool to deploy beside Kiteworks. The workflow follows three steps.
Ask. Type a question the way you would say it to a colleague. “Who downloaded the most files last week?” “Which risk policies fired for this user?” “What happened to this file?” KAI turns the question into controlled queries against activity and audit logs, users, files, folders, and risk policies, and it maps everyday wording such as “logged out” to the correct event names. A first answer takes about half a minute. When a request is ambiguous, KAI asks a clarifying question instead of guessing.
Verify. Every answer comes with Show work, a list of the lookups KAI made. Reviewers can follow the reasoning, confirm the scope, and reproduce the result. That matters for compliance work, where an answer is only as useful as your ability to explain how you got it.
Act. Ask KAI to draft a risk policy, for example one that restricts external sharing of CUI-tagged files. KAI checks the policy, resolves the users, tags, folders, and countries it refers to, and presents readable conditions and actions for review. Nothing is saved until a person applies it.
Why Security Teams Can Approve a Read-Only AI Agent
Gartner's guidance on AI agent governance recommends proportional trust across four autonomy levels: observe, advise, act with approval, and act autonomously. KAI sits at the advise level. That is the level regulated buyers find easiest to approve, and it is a deliberate choice, not a limitation.
- Read-only. KAI explains and drafts. A person applies any policy.
- No back door. KAI reads through Kiteworks Core's own APIs as the signed-in admin, so there is no separate permission model to review.
- Role-limited. Chat access is limited to System Admin and Compliance roles, and settings access to System Admin and Application Admin roles.
- Audited. AI settings changes are recorded in the Kiteworks audit log, credentials are stored encrypted, and KAI runs a FIPS self-test at startup.
- Switchable. An administrator can turn KAI off at any time without removing it.
The stakes of getting access control right are real. IBM's 2025 Cost of a Data Breach research found that 97% of organizations that reported an AI-related breach lacked proper AI access controls. A tool that inherits your existing permissions and records its own configuration changes starts from a stronger position.
Your Model, Your AWS Account, Your Appliance
KAI does not ask you to trust a vendor's model. It uses Anthropic Claude through your own AWS Bedrock account. You hold the credentials, and model usage sits under your own AWS agreement, not a vendor's. Your AWS terms govern Bedrock processing, and your existing cloud controls apply.
Conversation history is stored in a database on your own appliance, and the KAI service log does not hold prompts, answers, file content, names, or email addresses. That supports data minimization reviews, including those run under GDPR-style requirements.
Customers in regulated industries should also know about two deployment details. KAI needs direct outbound HTTPS access on port 443 to the Bedrock and STS endpoints, and it cannot route through a customer web proxy today. KAI uses Bedrock global cross-Region inference profiles, so inference can run outside your home region. If strict data residency applies to you, review your requirements with Kiteworks before enabling KAI.
Where KAI Fits in the Agents Workflow Control
KAI is the first agent in the Kiteworks Agents workflow control. The second half is the Kiteworks MCP Server, which lets customers connect their own AI clients, such as Claude, to Kiteworks data under the same permissions and the same audit trail. The umbrella idea is simple: your agents and ours, one governed data layer.
Most teams start with KAI for governance and insight, then add the Kiteworks MCP Server when business teams want their own AI tools working with Kiteworks data. Either way, the agent passes through the same permissions and lands in the same audit trail.
What You Need to Get Started
KAI is designed to be light on operations. It runs as one lightweight pod with automatic upgrades, and a single-node RKE2 deployment is enough. To run KAI, you need:
- A Kiteworks Enterprise deployment on version 10.0.0 or later
- The Solr search role and RKE2
- An AWS account with Bedrock access to Anthropic Claude
- Direct outbound HTTPS on port 443 to the Bedrock and STS endpoints
KAI runs as a single pod in this release. If its node fails, KAI is unavailable until the pod is rescheduled, and the rest of Kiteworks is unaffected. Customers without Bedrock, with proxy-only egress, or with strict in-region processing requirements can talk with Kiteworks about alternatives, including starting with the Kiteworks MCP Server.
Kiteworks Puts an AI Agent You Can Audit Inside Your Governance Center
Compliance is a discipline of evidence, and evidence is central to data governance. KAI helps admin and compliance teams produce that evidence faster, without adding a black box to the control plane for secure data exchange. You ask in plain language, see every lookup, and apply policy on your own terms, using your own model and your own account.
To see KAI answer real audit questions in your environment, schedule a custom demo. Current Kiteworks customers can contact their customer success manager or use the early access sign-up form to learn more.
Frequently Asked Questions
KAI is the AI agent built into Kiteworks, part of the control plane for secure data exchange. Admin and compliance users ask questions about activity and audit logs in plain language in the Governance Center, see the lookups KAI made through Show work, and can ask KAI to draft risk policies that a person reviews and applies. It is the first agent in the Kiteworks Agents workflow control.
KAI is read-only in this release. It reads through Kiteworks Core’s own APIs within the access controls of the signed-in admin and drafts risk policies for review. A person applies any policy, AI settings changes are recorded in the audit log, and an administrator can switch KAI off at any time without removing it.
KAI uses Anthropic Claude through your own AWS Bedrock account, and you hold the credentials. Conversation history stays in a database on your appliance, and the KAI service log does not hold prompts, answers, file content, names, or email addresses. Your AWS terms govern processing in Bedrock. This supports data minimization reviews.
You need a Kiteworks Enterprise deployment on version 10.0.0 or later with the Solr search role and RKE2, an AWS account with Bedrock access to Anthropic Claude, and direct outbound HTTPS on port 443 to Bedrock and STS endpoints. A single-node RKE2 deployment is enough. If strict data residency applies to you, review your requirements with Kiteworks before enabling KAI.
Chat access is limited to the System Admin and Compliance roles, and settings access is limited to System Admin and Application Admin roles. KAI works with the permissions of the signed-in admin, which keeps it inside your existing role-based access control model and adds no separate permission model.
Additional Resources
- Blog Post
Zero‑Trust Strategies for Affordable AI Privacy Protection - Blog Post
How 77% of Organizations Are Failing at AI Data Security - eBook
AI Governance Gap: Why 91% of Small Companies Are Playing Russian Roulette with Data Security in 2025 - Blog Post
There’s No “–dangerously-skip-permissions” for Your Data - Blog Post
Regulators Are Done Asking Whether You Have an AI Policy. They Want Proof It Works.