Where are mid-size law firms actually deploying agentic AI?
Mostly not in legal work, and mostly not yet. About 60% are still evaluating, with no agents live. The quarter that has moved made the same call: they started in the back office, in accounting, HR, and finance reporting, where the processes repeat, the data is their own, and a mistake costs a redo instead of a bar complaint. The firms building real fluency now are doing it on low-stakes work, so they are ready when the high-stakes requests arrive.
| What mid-size firms report | Share |
|---|---|
| Still evaluating agentic AI, with no agents deployed | 60% |
| Have deployed or are actively building agents | 25% |
| Have disabled agentic features or stayed skeptical | 15% |
How we know this
Sidebar puts one question a week to legal management professionals at firms of 10 to 200 attorneys. Members are verified by title, employer, and firm size before they are admitted, and every reply is private. This page draws on every Sidebar cycle that has touched this question, and it is updated as new replies come in. We publish patterns across the group, never individual firms, and only once at least five members have replied to that question. Full methodology at gosidebar.ai/methodology.
Operations before legal, every time
The firms that have deployed or are building agents almost all started in the same place: accounting, HR, and finance reporting. Legal work comes later, if at all for now. The sequencing is not timidity. It reflects where the processes are repeatable and rules-based, where the firm controls its own data, and where the liability if something goes wrong is an internal error rather than a malpractice exposure. Back-office work is the natural first surface for a tool that acts on its own.
The measurable wins are showing up in accounting first
Among firms that have actually shipped something, accounting is the first real win. Reporting tasks that used to take half a day are coming back in minutes. Others are lining up automated financial reports, conflict checks, and client notifications. The pattern is consistent: pick a task where the output is easy to verify and the cost of a bad run is low, and the value shows up fast. That is a very different proposition from pointing an autonomous tool at a client matter and hoping.
Data quality is the unglamorous blocker
The constraint that comes up again and again is not the agents. It is the data underneath them. Inconsistent records and siloed systems mean an autonomous tool has nothing reliable to act on, and several firms named exactly that as the reason they are watching rather than building. The firms moving forward are doing the boring work first: unifying data, cleaning up records, and mapping the process before they add the agentic layer on top. The agent is the easy part. The plumbing is the project. There is a useful consequence buried in that. A firm that decides it is not ready for agents still has a year of unambiguously worthwhile work in front of it, because unified data and mapped processes pay off whether or not an agent ever runs on top of them. That is the rare case where the prerequisite is worth doing on its own terms, which makes it a low-regret place to spend the time while the tooling settles.
The attorneys mostly do not know this is happening
Almost all of this is running inside administration, and the lawyers are largely unaware of it. Firms report that attorneys do not know what agentic AI is, let alone what their own firm has already automated. Right now that is harmless, because nothing agentic is touching client work. It stops being harmless the moment it does. A firm that has quietly built real operational fluency and never told its attorneys will find itself introducing autonomous tools to a group with no frame of reference for them, at exactly the point where the stakes are highest. The people doing this work are accumulating a working sense of where the technology holds up and where it does not, and that understanding currently lives nowhere near the attorneys who will have to sign off on using it. The gap is worth closing before it has to be closed under pressure.
One tool on paper, several in practice
A newer wrinkle in the same back-office story: even where a firm has picked one supported platform, several AI tools tend to be running underneath it with nobody coordinating between them. The gap is not really between sanctioned and unsanctioned use. It is between the tool a firm licensed and the tool an individual already trusts, and those are turning out to be different products more often than firms expect. People who prefer a different assistant than the one the firm deployed keep using it anyway, quietly, which means the firm's real agentic footprint is wider and messier than what shows up on the license list. That also changes what moving from back office to legal work will require. The harder step is not proving an agent can do useful accounting work. It is the move from one person's working prototype to a system the rest of a department can rely on, and that is where several firms report getting stuck now that the earlier split between evaluating and deployed has started to close.
Accountability is the reason it stops at the office door
The hesitation about attorney-facing agents is not squeamishness about technology. It is a specific and well-founded worry about autonomous tools acting on behalf of a licensed professional. Firms are direct about this: the question is not whether an agent could draft the thing, it is who answers for it when the agent is wrong and nobody caught it. That maps onto real malpractice exposure, and it is why the same firms comfortable letting an agent assemble a financial report will not let one near a filing. The distinction they are drawing is not about difficulty. It is about who carries the consequence, and it explains the sequencing better than any assessment of what the tools can technically do.
The wider market is at the same stage
The caution here is not a mid-size trait. Thomson Reuters Institute reports that fewer than one in five organizations describe themselves as engaged in widespread agentic AI adoption, while roughly half are planning to use it or still deciding whether to. Their reporting singles out the same sticking point our members name: autonomy is the one area they flag as most likely to limit how far agentic AI spreads in legal work. One lawyer in that research described agentic AI as removing oversight a step too far, and questioned giving a machine that much latitude in the doing of a thing without concrete review. That is the same instinct behind starting in accounting. The difference at mid-size firms is that there is no innovation team to run a controlled pilot, so the back office is not a staging ground chosen on principle. It is the only place with enough repetition and enough tolerance for a redo.
The advantage goes to whoever practiced first
There is a quiet edge here for the operations side of the firm. You do not have to start where the pressure is loudest. Start on work you already control, hand one task to an agent, and let it be imperfect while you learn what the technology is genuinely good at. Attorneys, by and large, do not yet know this is happening, which buys the back office time to build fluency before agents ever touch a client workflow. The firms that use agentic AI well will not be the ones that bought first. They will be the ones whose legal management professionals learned the tools early, on work where a mistake costs a redo and nothing more.
What to do with this
Map one repeatable administrative workflow this week and count how many of its steps require no legal judgment at all. The firms that moved fastest did not start from an AI strategy. They started from a single rules-based process in accounting or HR, documented every step, and marked the points where a human genuinely has to look. Whatever is left after that subtraction is the part an agent can hold, and it is usually more of the process than anyone expects. Treat the first attempt as practice rather than deployment. Let it be imperfect on work where being wrong costs a redo. That is how you find the edges of what the technology can hold before a request arrives that you cannot afford to get wrong. The firms that will use agentic AI well are not the ones that bought first. They are the ones whose operations teams built real fluency early, on work where a mistake cost nothing but time. Fluency is not something a firm can buy in the quarter it turns out to need it, which is the whole argument for starting somewhere unglamorous now.
Maybe the safest way to get someone to believe that they could be a builder is to build a couple of things for them and let them see how they work. Kind of like the cooking show. You built it, you show it to them, they use it, and then you're like, let me show you how I did that.
Members get more. The breakdown by firm size, practice area, and tech stack. A new question every Tuesday, the outlier answers that cut against the pattern, and most weeks an expert take with one clear next step.
Frequently asked questions
- What is agentic AI in a law firm context?
- An AI agent that acts on a task with some autonomy, rather than a chatbot that only responds to prompts. In a law firm, that means agents that can complete a repeatable process on their own, like drafting a report or flagging an exception, instead of waiting for an attorney to ask a question each time.
- Should a law firm start agentic AI in legal work or the back office?
- Peer data points to the back office first. Accounting, HR, and finance reporting are repeatable, use data the firm already controls, and a mistake costs a redo rather than a malpractice exposure. Legal work is where firms are moving slower, if at all yet.
- Why are so many mid-size firms still just evaluating AI agents instead of deploying them?
- The most common blocker is not the agents themselves. It is data quality. Inconsistent records and siloed systems give an autonomous tool nothing reliable to act on, so firms are doing the unglamorous work of cleaning up data before adding an agentic layer on top.
- What is a realistic first agentic AI use case for a law firm?
- A back-office task with an easy-to-verify output and low cost if something goes wrong, such as an automated financial report or a routine conflict check. Firms that have shipped something started narrow rather than pointing an agent at client-facing work.