Tencent just shipped something called Team Memory, a feature that lets a group of AI agents share what they’ve learned across a whole team instead of starting fresh with every conversation. VentureBeat covered it as a genuine step forward for multi-agent collaboration. Buried in the same piece was a detail that should stop any managing partner mid-scroll: there’s no governance layer yet for what happens when that shared memory is wrong. One bad context gets picked up by an agent, propagates across the team, and nobody catches it until it’s already in front of a client.

That’s not a Tencent problem. That’s an AI-in-professional-services problem, and law firms are exposed to it more than almost any other vertical.

Why context errors matter more in a law firm than anywhere else

Most industries can absorb an AI mistake with an apology and a refund. A law firm can’t. If an AI agent drafts a client update that misstates a filing deadline, misreads a clause in a contract, or pulls the wrong precedent because it lost track of which matter it was working on, that’s not a customer service issue. That’s a malpractice exposure with your firm’s name on it.

The VentureBeat piece cites a figure worth sitting with: 57% of enterprises using AI agents have traced a wrong answer back to missing or corrupted context, not a model that was “bad” in some general sense, but an agent that lost track of what it was supposed to know at the moment it needed to know it. For a law firm, that’s the difference between an agent that correctly recalls a client is under an NDA on a related matter, and one that doesn’t, and drafts something it shouldn’t have.

The uncomfortable truth is that most firms adopting AI tools right now have no equivalent of this governance question answered internally. They’ve got an intake bot, a drafting assistant, maybe a research tool, and no defined checkpoint where a human confirms the output before it reaches a client’s inbox. That gap is exactly where exposure lives.

The manual work this is actually about

Before we talk about fixing it, it’s worth being precise about where AI touches client-facing work in a typical $1M-25M firm, because the risk isn’t evenly distributed.

Intake and first response. A prospective client calls after hours or fills out a form at 9pm. Someone, or something, has to capture the facts, run a conflict check, and get the matter in front of the right person. Firms we talk to typically lose 30-40% of after-hours intake because nobody responds fast enough and the caller moves to the next firm on their list. If an AI agent is doing that first response, and it misreads the conflict check or mischaracterizes the matter, that error is baked into the file from day one.

Document review and discovery. Junior associates spend days doing first-pass review on contracts and discovery batches, work that runs $200-400 an hour of billed or written-off time depending on how it shakes out. AI can compress that timeline dramatically. But if the agent loses context between batches, say it flags a clause as standard in batch three that it correctly flagged as a problem in batch one, that inconsistency doesn’t just cost time to fix. It can walk straight into a filing.

Matter admin and billable-hour leakage. Most attorneys we talk to lose 4-6 hours a week to admin that never gets billed: status updates, file organization, chasing signatures. AI can take a lot of this off their plate. The risk isn’t usually catastrophic here, but it’s where sloppy habits form. If your team gets used to letting AI output go out the door unchecked on the low-stakes stuff, that habit doesn’t stay contained to the low-stakes stuff.

None of this means AI doesn’t belong in a law firm. It means the firms getting real value out of it have built a specific structure around where a human has to look before a client sees anything.

What “no governance” actually looks like in practice

Here’s a scenario that maps directly onto the Tencent story, translated into a legal practice. Say your firm runs a shared AI assistant across three associates working related matters for the same client. One associate asks it to summarize a deposition. The agent, pulling from shared context, blends in a detail from a different matter involving the same client, one that’s actually under a separate engagement letter with different terms. The summary goes to the partner. The partner, trusting the tool, forwards it. Now you’ve got a document in a client file that references information it had no business referencing, and there’s no record of anyone catching it before it moved.

That’s the failure mode VentureBeat is describing at enterprise scale. In a law firm, it’s not an efficiency problem. It’s a bar complaint waiting to happen.

The fix isn’t “don’t use AI.” The fix is treating every AI-generated output the same way you’d treat a first draft from a first-year associate: useful, often very good, and never sent to a client without a second set of eyes that knows what they’re looking at and why.

What a properly governed AI agent setup looks like

This is where the design of the agent matters as much as the model underneath it. At Enterprise DNA, when we build AI systems for law firms through Omni, we build the human checkpoint into the workflow rather than treating it as an afterthought.

Take the Intake Voice Agent, part of Omni voice. It answers every call, after-hours, during lunch, on weekends, runs a conflict check against your existing matter database, and captures the essential facts of the inquiry. It books a consultation directly into the firm’s calendar. But it doesn’t draft an engagement letter or give legal guidance on the call. It captures and routes. The line between “gather information” and “give advice” is drawn into the agent’s design, not left to hope.

The Matter Triage Agent, built on Omni ops, reviews incoming form submissions and emails, classifies the practice area, scores fit against your firm’s typical matter profile, and routes to the right partner with a one-paragraph brief attached. That brief is a starting point for a human decision, not a final answer. The partner still reads the actual submission. The agent just makes sure nothing sits in a queue for six hours while a competitor’s intake line picks up the call.

The Document Review Agent, also part of Omni ops, does first-pass review on contracts, discovery batches, and matter files. It flags clauses, summarizes positions, and produces an associate-grade memo. That memo goes to an actual associate or partner before it goes anywhere near opposing counsel or a client. The agent compresses days of first-pass work into hours. It does not replace the review that catches the agent’s own mistakes.

That last point is the whole argument. An AI agent without a mandatory review checkpoint is a liability generator with good production values. The same agent with a checkpoint built into the workflow is a genuine efficiency gain, because the human review step is fast when the first-pass work is already 80% done and clearly documented.

The checkpoint is the product, not the afterthought

Firms that get this wrong tend to make one of two mistakes. Either they avoid AI entirely because they’re worried about exactly the scenario above, and they keep bleeding the 4-6 hours a week per attorney and the 30-40% of after-hours intake to competitors who moved faster. Or they adopt AI tools quickly, get a taste of the speed, and start letting output go out without anyone checking it, because checking it “defeats the purpose.”

Both are wrong. The firms doing this well build a specific rule: no AI-generated output reaches a client, opposing counsel, or a court without a named human confirming it first. That’s not a suggestion buried in a policy document nobody reads. It’s a workflow step, logged, with a name attached to the approval. When something does go wrong, and eventually something will, you can trace exactly where the check happened and who made the call. That’s the difference between a firm that can defend its process and one that’s guessing.

If you want a starting framework for building that checkpoint into your intake process specifically, we put together a practical worksheet, the AI Client Intake Checklist for Law Firms, that walks through exactly where conflict checks, human review, and matter routing need to sit relative to any AI touchpoint. It’s built for firms doing this for the first time, not for firms that already have a compliance department dedicated to it. You can grab the direct version here.

What this costs you if you get it wrong, or do nothing

Firms in the $1M-25M range typically leave somewhere between $80,000 and $250,000 a year on the table across unbilled attorney hours, lost after-hours intake, and associate time spent on first-pass review that AI could compress. That’s the leakage side of the ledger, and it’s real money whether or not you touch AI at all.

The liability side is harder to put a number on, because it shows up as a single bad incident rather than a steady drip. A malpractice claim tied to an AI-generated error that went out without review isn’t a cost you amortize. It’s a cost that can undo years of a firm’s reputation in one filing. That asymmetry is exactly why the governance question matters more here than in almost any other industry adopting these tools right now.

The firms getting this right aren’t the ones avoiding AI. They’re the ones who’ve decided, deliberately, where the human checkpoint sits, and who’ve built their agents around that decision instead of bolting it on after something went wrong. That’s a design choice, not a policy memo, and it’s the difference between AI that makes your firm faster and AI that makes your firm exposed.

Where to start

You don’t need to overhaul your practice management system or hire an AI compliance officer to get this right. You need a clear picture of where AI touches client-facing work today, where the review checkpoints actually sit, and where they don’t. That’s what an Omni Audit is built to give you.

It’s a 60-minute conversation, no deck, no sales pitch dressed up as a workshop. You walk away with three concrete things: a map of where your firm is currently leaking billable time and intake conversions, a specific read on where AI-generated output might already be reaching clients without a defined review step, and a rough estimate of what closing those gaps is worth in dollar terms for a firm your size. See Omni for law firms for a sense of what we look at before you book anything.

If you’d rather just get on the calendar, book a 60-min Omni Audit and we’ll go through your intake process, your document review workflow, and where the governance gaps actually sit, specific to your matters, not a generic checklist.

For more on how firms are structuring AI adoption without creating new exposure, our insights section has a growing set of write-ups on this exact tension, and the guides library has more detail on how we scope agents like the Matter Triage Agent for firms still deciding what to automate first. If you’re earlier in the process and just want to understand what’s realistic for a firm your size, the AI audit for law firms is the right next stop, and Omni advisory is where we go deeper on governance design once you’re past the initial conversation.

The Tencent story is a signal, not an isolated incident. Shared AI memory across teams is coming to every industry that uses agents for real work, legal practices included. The firms that build the human checkpoint in now, before it’s forced on them by an incident, are the ones who’ll get to keep using these tools with a straight face when a client asks how their file was handled.