79% of legal professionals now say they use AI tools in practice. That number was 37% a year ago. If you take it at face value, you'd assume AI-powered knowledge management has swept through the mid-market like a wave - every firm with 150+ fee earners running some kind of intelligent system that surfaces the right precedent at the right moment and matches the perfect lawyer to every new instruction.
The reality is quite a bit more modest. And honestly, that's fine - as long as you know it.
Because the problem right now isn't that AI knowledge management tools don't work. Some of them work very well. The problem is that the gap between what vendors demonstrate in a polished 45-minute demo and what actually produces measurable value inside a mid-market law firm is significant enough to trip up any knowledge manager, IT lead, or operations partner trying to make a sensible buying decision. I've sat through enough of those demos - and then worked with enough firms trying to operationalise what they bought - to know that the gap is where most of the wasted money lives.
We've looked at the AI KM tools and the vendor demos are impressive, but we can't tell what's real and what's marketing. We don't want to be an expensive beta tester.
That's the right instinct. So let me try to give you something vendor marketing won't: an honest picture of what's been deployed, what's producing results, what's still more promise than performance, and where to start if you want to move this quarter without betting the farm.
Let me be specific about what I'm seeing across mid-market UK law firms - by which I mean firms with roughly 100 to 500 fee earners. Not Magic Circle. Not sole practitioners. The firms where the knowledge management function is real but stretched, and where the budget for experimentation is limited.
The most common AI KM deployments right now fall into three buckets. First, AI-enhanced search layered on top of existing document management systems - iManage and NetDocuments, primarily. This is the workhorse. Least glamorous, most consistent results. Second, experience databases with AI-assisted matter tagging and lawyer profile enrichment - essentially making it easier to answer "who in this firm has done something like this before?" Third, automated document classification and metadata tagging for new documents entering the system, which sounds dull until you realise how much of the data quality problem in law firm KM comes from inconsistent tagging at the point of filing.
That's the reality. ILTA's most recent survey data suggests AI adoption in legal knowledge management specifically has moved from around 15% in 2023 to 37% in 2024, with roughly 50% of firms now having some form of AI task force or working group. But having a task force and having a production deployment are very different things.
What's less common than vendor marketing suggests? Autonomous document drafting that meaningfully reduces fee earner time. Predictive analytics for matter outcomes. And "intelligent legal research" that produces reliable conclusions without someone senior checking the sources. I'm not saying these don't exist in any form. I'm saying that in the mid-market, at the kind of firms you and I are talking about, they're largely still pilot territory or confined to very specific, narrow use cases.
Let's start with the good news, because there is genuine good news here.
AI-enhanced document search is the standout. Specifically, natural language query against a well-structured document library - the ability for a fee earner to type something like "shareholders' agreement with drag-along provisions for a tech company, completed in the last three years" and get useful results rather than a wall of noise. The evidence for this is strong. Where firms have deployed it against a reasonably well-organised DMS, the workflow improvement is immediate. Fee earners who previously spent 20 to 30 minutes hunting through folders - or, more commonly, just asked the person sitting next to them - are finding relevant documents in under two minutes. That's not a marginal gain. Over the course of a week, across a team, it adds up to hours.
The conditions required: your document library needs to be in a searchable state. I'll come back to what that means, because it's the bit most vendors gloss over in the demo. But if your iManage or NetDocuments environment is reasonably well-maintained, this application works now, with current technology, and you can see results within four to six weeks of deployment.
Precedent identification - finding the most relevant prior work for a new matter from the firm's historical archive - is the second strong performer. I worked with a firm last year where the knowledge team had been manually maintaining a precedent bank for years. Good people, diligent work. But the bank covered maybe 15% of the firm's actual output, because that's all the team had capacity for. An AI layer over the full document archive surfaced relevant precedents they didn't even know existed. One partner told me - and I'm paraphrasing slightly - "I've been here seventeen years and I didn't know we'd done that deal." That's the kind of moment that makes the investment case self-evident.
But - and this is important - precedent identification works well where the archive is reasonably well-tagged by matter type and practice area. Where tagging is inconsistent or incomplete, the results degrade noticeably. Not to zero, but enough that fee earners lose trust in the tool quickly, which is a harder problem to fix than the technical one.
Experience databases are the third application producing measurable results. AI-assisted matching of client needs to lawyer expertise using structured matter history. If a prospective client comes in with a cross-border M&A question involving German regulatory approvals, the system can identify which lawyers in your firm have handled comparable matters - and surface that information in seconds rather than relying on someone's memory or a round of emails. I've seen this produce real improvements in pitch quality and staffing decisions. One firm I'm thinking of reduced their average pitch preparation time by a third, not because the AI wrote the pitch, but because it found the right people and the right experience to put in front of the client. The BD partner told me afterwards that they'd been doing it the hard way for years - ringing round, asking who'd done what, waiting for people to dig out old files. Turns out the answer was sitting in the DMS the whole time.
Again, the caveat: this works when the underlying data is reliable. If your matter history is incomplete or your lawyer profiles are out of date - which, let's be honest, they usually are - the output will reflect that.
I want to be careful here, because "not working" doesn't mean "will never work." It means that the gap between what the vendor showed you in the demo room and what happens when you try to run it against your actual data, with your actual fee earners, in your actual workflows, is still wide enough to be a problem.
Automated drafting from precedents is the one I get asked about most. The demo is always impressive - the system pulls a relevant precedent, generates a first draft adapted to the new matter's specifics, and the fee earner just reviews and refines. In practice, for complex documents, the review required is so extensive that the time saving is marginal. I've seen implementations where senior associates were spending almost as long checking and correcting the AI-generated draft as they would have spent drafting from a template themselves. For simple, highly standardised documents - NDAs, basic engagement letters - there's genuine value. But for the complex, bespoke drafting that makes up the bulk of high-value legal work, we're not there yet. The technology will improve. But right now, if a vendor is selling you autonomous complex drafting as a current capability, that's marketing, not reality.
Predictive analytics for matter outcomes is the one that sounds most exciting and is furthest from being useful. The data required to train a reliable predictive model - consistent, structured outcome data across comparable matters - simply doesn't exist in most mid-market firm document archives. Your matters are stored in documents, emails, file notes, and the memories of the people who worked on them. Until that data is structured, standardised, and comprehensive enough to train a model, the predictions won't be reliable enough to influence decisions. I've yet to see a mid-market firm where this is producing actionable insight. Not one.
Autonomous legal research sits somewhere in between. The AI research tools are genuinely impressive at surfacing relevant sources quickly. But the source verification step - checking that the cases cited actually exist, say what the tool claims they say, and are still good law - currently consumes a significant chunk of the time you saved on the research itself. For straightforward questions, the net time saving is real. For complex, multi-jurisdictional research questions, it's marginal. And in a profession where citing a hallucinated case is a career-defining embarrassment, the verification step isn't optional. I've spoken to associates at two different firms who'd quietly stopped using their AI research tool for anything complicated, because the checking took longer than just doing the research properly. That's not a ringing endorsement.
Here's the section that might frustrate you, because it's the least exciting part of the whole conversation. But it's the one that determines whether everything above actually works for your firm.
AI KM tools perform in direct proportion to the quality of the data they operate on. Full stop. A document management system with inconsistent tagging, incomplete matter history, poorly named files, and limited structured metadata will produce poor AI KM results regardless of how sophisticated the AI layer is. I've seen firms spend six figures on an AI KM tool and then wonder why the results are disappointing, when the answer was sitting in their DMS the entire time - thousands of documents filed as "Draft v3 FINAL (2).docx" with no matter type, no practice area tag, and no structured metadata.
But our DMS has been in place for fifteen years and the data's fine.
Is it though? Because in my experience, "the data's fine" usually means "nobody's looked at it recently." The accumulated debt from years of inconsistent filing, departed fee earners who had their own naming conventions, and matters that were never properly closed out shows up the moment you try to put an AI layer on top of it. I've written separately about why data quality matters more than fancy dashboards - the argument applies here with particular force.
The minimum viable data foundation for productive AI KM looks something like this: a document library that is consistently tagged by matter type and practice area, lawyer experience data that is structured enough to be searchable, and a data governance process that maintains quality as new documents enter the system. That last one is the kicker, because it's not a one-off project. It's an ongoing discipline.
And yes, this is almost always the most significant investment before any AI KM tool can deliver its claimed capability. Not the licence fee. Not the integration work. The data.
I don't want to leave you with a counsel of perfection that feels like "sort out your entire data estate and come back in eighteen months." That's not practical, and it's not necessary.
Based on what I've seen produce the fastest visible value, here's how I'd sequence it.
Start with AI-enhanced search over your existing document library. This is the application that requires the least data preparation, works with your current document management system, and produces a visible, measurable result within four to six weeks. It won't be perfect - the results will reflect whatever state your DMS is in - but it will be useful enough that fee earners start using it. And that adoption is what creates the momentum for everything else.
Use that deployment to assess what data quality improvements you actually need. Once people are using AI search, the gaps in your data become obvious very quickly. "Why didn't it find the Henderson matter?" "Because it wasn't tagged as a property dispute." That conversation is worth more than any data audit, because it makes the case for data quality in language that fee earners understand and care about.
Build the data governance process as a parallel workstream, not a prerequisite. Start with new documents - make sure everything entering the system from now on is properly tagged and classified. Then work backwards through the archive in priority order, starting with the practice areas where AI KM will have the most impact. I've written about three things a 300-person law firm can do with AI this quarter - use case two covers the knowledge search quick start in more detail.
One more thing worth flagging: the biggest risk to any AI KM deployment isn't the technology. It's that fee earners have a perfectly functional informal knowledge network already - it's called asking the partner in the next office - and your AI tool needs to be easier and faster than that, or it won't get used. That's a design challenge as much as a technology one. There's a companion piece on designing for adoption that covers this in more detail.
I'm genuinely optimistic about AI-powered knowledge management in law firms. The applications that work - search, precedent identification, experience matching - are producing real, measurable value right now. Not theoretical value. Not "potential" value. Actual hours saved, actual precedents found, actual pitches improved.
But some of what's being marketed to you is ahead of where the technology reliably performs in a mid-market firm with real-world data quality. That doesn't make it vapourware - it makes it early. And "early" is fine, as long as you know that's what you're buying.
If you want to understand what your firm's current document archive and data governance would actually support in terms of AI KM capability - and what the minimum viable improvement path looks like - book a KM readiness assessment. It's a structured two-week assessment covering document library tagging consistency, matter history completeness, structured lawyer experience data, and data governance processes. You'll get a clear picture of which use cases your firm is ready for now, which need foundation work first, and what that work actually involves. It'll save you from becoming someone else's expensive beta test.
For firms that want a broader view beyond knowledge management, we've published a detailed piece on how we approach AI readiness that covers the full assessment framework.