-optimized.jpg){kind=avatar}
As organisations race to adopt gen AI tools to boost efficiency and deliver client value, they are also creating new risks around highly sensitive data, making strategies like Data Loss Prevention (DLP) and Data Segregation mission-critical, says Aaron Rangel, Director of Product Management, iManage.
That tension leads to a central challenge: How can organisations embrace AI responsibly without compromising confidentiality or compliance โ and what strategies should they use to reduce risk and prevent accidental data leaks while still enabling their practitioners to work smarter?
AI raises the stakes
The โPre-AI Security Modelโ focused on a few core areas: Identity and access control, to ensure only authorised users can access systems and data; Data Encryption, to protect data at rest and in transit at every layer; Activity Auditing, to log every action against sensitive data for accountability; and Threat Monitoring to detect and respond to anomalous behaviour in real time. These are all still important. But AI changes the game โ and raises the stakes. With AI, the question is no longer just โwho accesses whatโ but โwhat can the model see, infer, and surfaceโ?
A crucial factor here is credential inheritance. Most AI agents operate under the end userโs credentials โ i.e., models see what the user sees. Thatโs fine when the agent is restricted to browsing a centralised repository like a document management system (DMS) that has ethical walls in place.
But DMS content that has been downloaded to AI-native platforms โ in other words, taken out of the DMS โ may not be protected by ethical walls. That copy no longer carries the DMSโs ethical walls or matter-level permissions.
The AI agent can now surface that content in a different context, combine it with other matters, or leak it through inference. And it can do this all while technically โrespectingโ the userโs permissions โ because the user had access when they exported it. The ethical wall was in the DMS, not in the AI platform.
The above scenario also brings up the issue of agent oversight. What monitoring exists when an agent behaves unexpectedly? What visibility do you have into what it accessed? Keep in mind, that in an agentic world, you have agents talking to agents via MCP (Model Context Protocol), so they have enormous potential reach into different systems. When an upstream agent has made an important conclusion or suggestion, can you see the underlying data that was used to make that decision? Can you shut down a rogue AI agent?
Finally, adding another layer of complexity, thereโs the issue of outside counsel guidelines and large language model (LLM) compliance. How can organisations comply with mandates that content cannot be sent to a LLM or unapproved LLMs?
Building safeguards without burdening professionals
When it comes to successfully addressing these complex issues, there are a few key points for organisations to keep in mind if they want to reduce risk and prevent accidental data leaks while not creating hurdles that prevent their professionals from working smarter.
Know where your data flows โ especially into AI-native platforms. The moment content leaves a governed system like a DMS and enters an AI-native environment, it often sheds its access controls. Organisations must map these data flows and implement controls that travel with the data, not just guard the perimeter. Where possible, favour AI integrations that operate via API and return insights or summaries rather than pulling full documents into an uncontrolled context. Additionally, several established software vendors integrate AI in ways that content does not leave the DMS. This limits the potential damage if something goes wrong.
Address the human layer, not just the technical one. Policies and training only work if they are practical enough for people to actually follow. Human error remains one of the leading causes of data breaches, and in the AI era, those errors often happen at the point of export โ when a practitioner copies a file into a tool that feels faster or more convenient.
Acceptable use policies need to be communicated clearly and reinforced regularly, not buried in documentation no one reads.
Establish agent oversight before you need it. In the new agentic world, agents are making key decisions โ so, you need to understand what documents the agents analysed to base their decisions.
Auditing and reporting supports this goal. Build logging, monitoring, and intervention capabilities into your AI deployments from the start. Audit trails for AI agent activity should be held to the same standard as those for human users. And if an AI agent behaves unexpectedly โ accessing data it should not, surfacing confidential content in the wrong context, or acting on stale permissions โ you need to detect it and shut it down quickly.
Understand capabilities AI natives offer vs. what established vendors offer: While AI natives have a head start in capabilities, established vendors are catching up fast. Additionally, established vendors have an advantage with data governance as duplicate copies of data in third-party applications are no longer needed.
Start with what you have. Many organisations already have DLP tools, ethical walls, and access controls in place. The priority is to extend those existing controls into AI environments incrementally, rather than waiting for a comprehensive solution that may never arrive.
At the end of the day
Responsible AI isnโt an unobtainable goal for organisations to seek โ it just requires a few key steps and some advance planning if they want to minimise the risk around their data while avoiding creating unnecessary friction for practitioners. Building the discipline now will pay dividends later, setting organisations up for the better business outcomes that AI can deliver.
