Firms evaluating agentic AI at ILTACON this year kept circling back to the same questions, whether they were buying a platform or building their own proprietary tools. What will the agent touch, what can it see, and can anyone reconstruct why it did what it did.
It means that the questions being asked at the executive level have changed too, from which tool to buy toward whether the underlying data can support what that tool is being asked to do.
Here are the ILTACON takeaways that matter most for technology leaders evaluating an AI investment, a CRM modernization project, or both at once: the return on an agentic AI investment in legal work depends less on model sophistication than on the data underneath it. Clean, governed, defensible data is what determines whether that investment produces measurable impact. Model selection tends to get most of the attention, but law firm data readiness and the foundation underneath any given model is what determines the return on an agentic AI investment.
Table of contents
Agentic AI changes the stakes for information governance
For the last two years, AI at most firms meant a lawyer asking a question and getting an answer back. That’s starting to change. Agentic AI in legal settings, capable of executing multi-step tasks with limited human oversight, reviewing documents, monitoring compliance obligations, cross-referencing matters, running research workflows end to end, are already running in production at a share of firms. A 2026 survey of 200 legal AI leaders found 38% report at least one agentic workflow already in production, with another 29% in active pilot. Wherever a firm sits on that curve, it changes what governance has to cover: an agent that acts rather than just answers touches privileged, client-specific data with a degree of autonomy legal tech hasn’t required firms to govern before.
Agents are well suited to some of that work and poorly suited to the rest of it, so governance has to draw a clear line between the two. Operations, pitches, and client relationship management sit on one side of it: agents can run there with appropriate oversight. So do repeatable, high-volume tasks within legal work itself, document review, compliance monitoring, cross-referencing. Final legal judgment sits on the other side. No agent should sign off on work or exercise legal discretion on your firm’s behalf.
Three principles should anchor that framework:
- Access control: An agent only touches the data and systems required for its specific task, nothing broader.
- Traceability: Every agent decision carries an auditable reason, something you can produce if a client, a regulator, or opposing counsel asks how a piece of work got done.
- Reversibility: Never let an agent fully automate an action that can’t be undone.
A version of that last principle came up at ILTACON, framing every delegated task as either a reversible “two-way door” or an irreversible “one-way door,” the latter requiring far stricter oversight.
The risks compound as autonomy increases. Hallucinations don’t disappear just because an agent is executing rather than answering. Token usage climbs without anyone noticing until the bill arrives. Data from different matters starts comingling across agent workflows that were never designed with ethical walls in mind. And frameworks like the EU AI Act are beginning to formalize obligations around exactly this kind of autonomous decision-making, which means legal tech AI governance is quickly becoming a compliance requirement rather than a best practice. Build the governance infrastructure before the agents go live, not after. That was the clearest instruction to come out of the ILTACON sessions on this topic. It’s far harder to retrofit oversight onto an agent already touching client data than to build that oversight in before the agent goes live.
Underneath all of that sits a harder problem: none of it holds up on disorganized data. Access control depends on knowing which data belongs to which matter, and traceability depends on records that aren’t duplicated, unattributed, or scattered across systems nobody owns. Without that, an agent operating in a legal environment removes the one thing that used to catch bad data before it caused a problem: a person checking it manually.
Legal information governance and the shadow AI threat
Information architecture came up at ILTACON too, and it’s a problem most firms will recognize regardless of which platform they run. Firms invest budget and good intentions into tools like Microsoft 365, and the gap that follows usually isn’t the technology, it’s the governance layer around it: unclear ownership, inconsistent policy, data that was never classified. A recurring observation at the conference: most SharePoint environments end up with a majority of sites that are functionally abandoned, technology deployed, never really adopted.
Three causes most often explain that gap:
- No clear content policy: People put data wherever is convenient because there’s no defined answer for where it belongs.
- Inconsistent governance: One practice group’s discipline does nothing for the next one over when standards aren’t applied firm-wide.
- Missing data classification: Client versus internal, sensitive versus general, without it, nobody downstream can tell what needs protecting and what doesn’t.
This is where legal information governance stops being an IT checklist item and starts being a prerequisite for responsible AI deployment. Without content classification, an AI tool can’t distinguish a client-privileged document from an internal memo, and it will treat both the same way. Without consistent permissions, an agent connected to your document management system inherits whatever access sprawl already exists. Shadow AI in law firms is the practical result: confidential client data flowing into AI tools in ways nobody planned, nobody approved, and in most cases nobody can detect until something goes wrong.
What ILTACON’s panelists proposed had nothing to do with buying another tool. Define team sites by department with clear ownership. Apply Microsoft 365 policies consistently rather than department by department. And build a framework around three things: communication, so people actually know where content belongs; convenience, so following the rules is easier than working around them; and classification, so both your permissions system and your AI tools can enforce the access boundaries that matter. That’s what closes the gap that lets shadow AI take hold, not another point solution layered on top of the mess.
These structural fixes, clear ownership, consistent policy, and classification, are some of the more practically useful ILTACON takeaways for anyone responsible for law firm data readiness or information architecture.
If shadow AI becomes a problem at a firm, it probably won’t start with a breach. It’s more likely to start small, for example a lawyer standing up an agent without firm-wide approval, an agent touching data it was never scoped to see, or client matters comingling because no one defined the boundary before the agent went live.
That same governance question, who can stand up an agent and what it can access, sits underneath a decision most technology leaders are already weighing: build or buy. But that question collapses two separate decisions into one. Where relationship context comes from, and what a firm builds on top of it, are not the same choice. A firm can build its own AI agents and workflows while still sourcing the relationship data layer underneath them. Getting precise about which layer is actually in question is where a real law firm CRM modernization conversation should start.
The data layer is where firms underestimate the commitment. A working proof of concept can be running in days. What takes longer, and rarely shows up in the pilot business case, is capturing relationship data at scale, enriching it from outside sources, deriving relationship strength and warm paths from raw activity, maintaining that picture as contacts and roles change, and reaching the security and compliance standard your firm already operates under, SOC 2 Type II, ISO 27001, GDPR, PIPEDA, and whatever client-specific terms sit in your engagement letters. None of that is a one-time build. Every new integration, every new agent touching that data, reopens the question.
The bigger issue isn’t cost. It’s what that data actually is. Email, calendar, and CRM data are inputs, not relationship context. An AI model can tell a partner when they last emailed a contact, but it can’t reliably tell them how strong that relationship is, whether their contacts are engaged or disengaging, or who else at the firm has a better path in. And an AI working from incomplete data doesn’t flag what it’s missing. It answers the question as if the picture is complete, which is a quiet failure mode until the moment it costs a firm a deal or a client’s trust. Law firm CRM modernization efforts that solve for contact discovery and stop there still leave that governance problem sitting underneath it, unaddressed.
Measuring impact in the business of law
One of the more counterintuitive ILTACON takeaways was that attendees who already run sophisticated AI stacks also showed strong interest in relationship mapping. The case for data quality, what it actually buys a firm day to day, hasn’t fully landed yet, even among people who understand the underlying architecture.
Clean, connected, well-governed relationship data shows a firm where the growth already is, inside accounts it holds today, not just the ones it’s chasing. It streamlines the pitch and RFP process by surfacing who at the firm already has a relevant relationship before anyone starts cold outreach. And it delivers that intelligence inside the tools attorneys already use, whether that’s Outlook, Word, whatever sits in their normal workflow, instead of asking a partner to log into another system.
Technology investment only counts if you can measure what it produced. That principle applies as much to your data infrastructure as it does to any AI tool sitting on top of it. Law firm data readiness isn’t compliance box, but the layer that determines whether every subsequent investment, agentic workflows, CRM modernization etc. produce something you can point to.
Conclusion: secure your data foundation first
If there’s one thread running through every one of these ILTACON takeaways, it’s that law firm data readiness gets decided long before anyone opens an AI tool. None of this requires taking a vendor’s word for it, including ours. The pattern holds regardless of who you talk to next: firms that treat agentic AI in legal work as a governance problem first tend to deploy it with fewer surprises.
A workable plan starts in the same place no matter which vendor or build path you eventually choose. Know where your firm’s client and relationship data actually lives, and who can vouch for its accuracy. Close the classification and access gaps before you ask any governance framework to enforce boundaries that do not yet exist. Only then decide whether the fastest, most defensible route to AI value is buying an enterprise platform or building one internally.
Introhive works with professional services firms on exactly that sequence, providing a secure, enterprise-grade relationship intelligence foundation built to the compliance standard legal data requires, so the data feeding your next AI investment has already been verified rather than assumed. Get the sequence right and every agent, every CRM modernization effort, and every AI initiative that follows it inherits data your firm can defend. Get it wrong and you’re left governing agents, and answering for their decisions, on a foundation nobody actually checked. Law firm data readiness is what separates those two outcomes, and it’s decided well before any agent goes live.
If you’re ready to see where you stand on law firm data readiness, book a demo with our team.
BOOK A DEMO
