Why information barriers are under more pressure than ever in the agentic AI era

For decades, information barriers have been one of the quiet workhorses of legal risk management and ethical compliance. They protect privileged and confidential information, manage conflicts, satisfy probity plans, safeguard price-sensitive deals, and shield sensitive matters or content from unauthorised eyes. Lawyers rely on them. Clients demand them. Compliance and risk teams administer them. IT enforces them.

Legal work carries structural features that make this especially acute: imputed conflicts that spread firm-wide from a single lawyer's exposure, privilege that can be waived through inadvertent disclosure, probity and regulatory regimes with hard compliance consequences, and clients who are themselves increasingly sophisticated about AI risk and asking pointed questions before instructing a firm. An information barrier that worked perfectly well when it only had to stop a person from clicking into the wrong folder now has to hold up against systems that can search, summarise, and act across every connected repository in seconds, and, in the agentic case, without anyone deciding in the moment that the search should happen at all.

Agentic AI is changing the landscape, making it critical for law firms to manage information barriers at scale. As firms shift from predominantly open document management systems to more closed environments to protect confidentiality, data protection, and data sovereignty, they are under growing pressure to lock down and ring-fence more matters. But the nuance and granularity that determine which matter teams, offices, or regions can access particular content are now being tested and stretched by the pressure of agentic AI. The question is no longer only which lawyers can access what, but also which agents have access. A familiar, largely static control is now being challenged by systems that can read, reason across, and act on firm content far faster than any human. This enterprise-wide need for scale is making this need more acute. Legal professionals and internal risk teams are right to be more anxious about this than they were even twelve months ago. Here’s why.

The barriers themselves haven't changed — but the stakes have

The prevailing duty of confidentiality hasn't changed. The professional and ethical rules that mandate how legal professionals manage conflicts and confidentiality are the same. Nor have the classic use cases changed: price-sensitive M&A and MAR-related work, matters involving children, multidisciplinary practice divisions, jurisdictional and reserved-area restrictions, client-requested walls, psychologically sensitive matters, the many flavours of conflicts (confidential information conflicts, former-client conflicts, perceived conflicts, current-client "Aussie exemption" scenarios), new-starter and secondee restrictions, probity plans, AML tipping-off, whistleblower files, and the list goes on and on.

What's changed is that what sits behind those walls now has an entirely new class of user trying to get at it: AI.

The adoption curve tells the real story

The iManage AI adoption curve, explained in a recent Security Policy Manager (SPM) webinar with Madeleine Porter, Legal Industry Expert, APAC, and Luke Creswick, Head of Pre-Sales, APAC, clarifies why risk is escalating in this current AI-dominated landscape.

• Assisted: AI lives in a browser tab. A person copies content in and out of a tool like ChatGPT. The human is the barrier, deciding what's shared and what stays protected.

• Augmented: AI moves inside the workflow itself (Copilot in Word, Ask iManage in Work). The human becomes a checkpoint; they review and approve, rather than the gatekeeper deciding what the AI can even see.

• Orchestrated: AI tools connect directly into multiple systems at once via APIs and MCP connectors, along with the DMS, Teams, and practice management systems, and can search and combine information across all of them in seconds.

• Delegated: Knowledge workers start building their own AI agents for repetitive tasks, such as pulling documents together, checking matter status, and generating first drafts. The agent is then bound by the human's identity and permissions, but it isn't a human. It moves faster, more frequently, and without judgment. If there is a mistake, it will not detect this, which compounds the error and potentially replicates this at scale.

• Integrated: AI agents become embedded, event-triggered participants in day-to-day work, and essentially become closer to digital employees than to tools. Nobody has to trigger the agent.

As organisations move left to right along that curve, human judgement, historically the last line of defence and the internal gatekeeper, steps progressively further out of the loop. That is the crux of why legal professionals are more concerned now: The control that has quietly underpinned information barrier compliance for years is being stretched and potentially compromised, exactly as the volume and reach of automated access increase.

The numbers back up the unease

According to results from the iManage Knowledge Work Benchmark Report 2026 a large majority of organisations are already piloting AI in some form, and more strikingly, more than a third have already experienced an AI-related data leak or compliance issue. Critically, these incidents were not the result of sophisticated attacks. They're happening through the everyday AI tools people are already using at work.

Three new pressure points

Beyond the general erosion of human gatekeeping, three specific challenges are surfacing repeatedly in conversations with legal organisations:

1. "AI must never touch this content." Clients are increasingly stipulating an outright prohibition on the use of AI in any form on their content. This augmented compliance measure goes beyond the traditional dictate that access should be restricted to certain people within a firm, but that no AI tool should interact with the content, full stop. Highly sensitive matters, such as those involving children, raise the same requirement. Furthermore, firms must not only comply with these enhanced compliance requirements but also prove that they are compliant; this means defensibility, auditability, and transparency, thereby substantiating adherence to client demands.

2. AI "vaults" don't inherently respect existing information barriers. As AI tools let users spin up dedicated project spaces or vaults for a deal or a discovery exercise with different AI platforms, content gets pulled in via API or MCP connectors; legitimately, at first, because the user creating the vault has access. However, the barrier breaks the moment that the vault is shared with a colleague who was never authorised to see the underlying content in the first place. This essentially renders the information barrier ineffective and exposes the firm to potential ethical, regulatory, or financial penalties.

3. Agents are being built without governance. Because building an agent no longer requires IT, as any knowledge worker who can describe a task in a few steps can create one, firms are accumulating agents faster than they can answer basic governance questions. These questions include: Whose identity is the agent acting under? What is it actually allowed to do, such as read-only, or can it edit, delete, or even send documents? Can it drift into the wrong client or matter? Can someone see what it did and shut it down if something goes wrong? Without clarity around governance, firms are at risk of agents going rogue and a catastrophic compliance blunder occurring.

What "good" looks like going forward

The response isn't to slow AI adoption, but to ensure that the governance layer scales with it. That means:

• Centralising policy enforcement so information barriers defined once cascade automatically across the DMS, practice management, time and billing, Teams, and file shares, rather than being manually recreated and propagated system by system.

• Extending barriers to AI explicitly, so that tools like Ask iManage and MCP-connected assistants respect the same access rules as everything else, and users simply never see what they aren't entitled to see. Bringing AI to your documents, rather than taking your documents to an external AI platform, ensures that your firm’s internal security, governance, and permissions are respected and preserved at all times.

• Building at the collection level, not matter by matter, such as grouping content by office, practice area, or client type so that need-to-know security can be applied at scale without creating unworkable admin burden or friction for legitimate users. This alleviates the resource burden within firms as some firms still require the manual oversight of ensuring information barriers have been accurately implemented, with horror stories of people working late into the evening waiting for the technology or walls to update.

• Governing agents as a distinct category with clear answers on identity, permission scope, monitoring, and audit trail, before agent-building becomes as ubiquitous as spreadsheet macros once were.

• Closing the content lifecycle loop, from conflicts and intake through retention and disposal to threat detection and extending protection (e.g., via sensitivity labelling) even after content leaves the platform.

Information barriers were never just a compliance checkbox; they're the mechanism that lets firms take on adverse, sensitive, and price-critical work with confidence and in accordance with their ethical and professional obligations. Agentic AI doesn't create a new duty. It just removes the human safety net that duty has quietly relied on for years, and it does so at exactly the moment client and regulator expectations for defensible and auditable control are rising.