Latest / Herbert Smith Freehills Kramer Podcasts / Legal Tech Deciphered EP4: The Misdiagnosis Effect: Finding the right solution to solve the right problem
Transcript
- 0:00SPEAKER_00: The fastest gains often come before any AI is introduced at all.
- 0:04SPEAKER_00: Claiming up the data, clarifying roles, redesigning a workflow so that it reflects how work actually moves rather than how it's assumed to move.
- 0:14SPEAKER_00: The technology then has something solid to build on rather than being asked to compensate for foundations that were never in place.
- 0:36SPEAKER_01: Trust, truth, and hallucinations.
- 0:39SPEAKER_01: Herbert Smith free heals Kramer's podcast, cutting through the noise around Legal Tech and Gen AI.
- 0:45SPEAKER_01: I want to start with a thought experiment.
- 0:48SPEAKER_01: A client walks in, they've seen a demo, the technology looks impressive, and then they say, We want to use Gen AI for contract review, or we need something for due diligence, or this is becoming even more increasingly common, we've bought the tool already, now help us make it work.
- 1:06SPEAKER_01: In every one of those moments, the technology has become the answer before anyone has properly defined the question.
- 1:13SPEAKER_01: This is what we're calling in this episode the misdiagnosis effect.
- 1:17SPEAKER_01: I'm Casey Ballard, head of strategic applications team.
- 1:20SPEAKER_01: I spend most of my time at the intersection of legal problems, process design, and technology.
- 1:25SPEAKER_01: Quite often at the point where those things don't line up in the way that people expect.
- 1:30SPEAKER_01: With me today are two colleagues from our digital legal delivery team, Suzanne Cowan, who leads our corporate and construction team and has spent 14 years designing legal delivery solutions that work in practice, not just on paper.
- 1:43SPEAKER_01: And Uchi Duhardener, who leads our data analytics team, turning complex, often unwieldy datasets into clear, actionable intelligence that helps clients make better decisions under pressure.
- 1:56SPEAKER_00: Hi, Kesy.
- 1:57SPEAKER_00: Hi everyone.
- 2:00SPEAKER_01: In our last episode, we talked about why there's no one-size-fits-all approach to AI-assisted workflows and why hybrid models are quickly becoming the standard defensible review.
- 2:11SPEAKER_01: Today, we want to go a level deeper, because what actually happens when those ideas hit real client requests, real timelines, and real pressures?
- 2:20SPEAKER_01: The short answer is things get messy and fast.
- 2:25SPEAKER_01: At the root of most of this is the mess in that misdiagnosis made right at this very start.
- 2:37SPEAKER_02: The first thing to realize is that data problems almost never come in one neat repeatable shape.
- 2:44SPEAKER_02: Contract review can mean a tidy batch of near-identical templates or a chaotic mix of negotiated PDF scans and side letters.
- 2:58SPEAKER_02: Each of these are likely to sit on completely different data solutions and foundations.
- 3:05SPEAKER_02: The quality, the structure, the consistency, and completeness, all of these can vary widely.
- 3:12SPEAKER_02: Sometimes inside the same meta itself.
- 3:14SPEAKER_02: And Genei is incredibly sensitive to that.
- 3:18SPEAKER_01: Exactly.
- 3:18SPEAKER_01: And it's the way that the sensitivity plays out that makes this particularly tricky, isn't it?
- 3:24SPEAKER_02: Exactly.
- 3:25SPEAKER_02: With traditional tools, imperfect data gave you imperfect outputs.
- 3:30SPEAKER_02: And crucially, you could actually see the relationship between the input and the output.
- 3:35SPEAKER_02: You could see where the gaps were, but Geniite changes that entirely.
- 3:41SPEAKER_02: When the inputs are poor, the outputs don't look poor.
- 3:46SPEAKER_02: They look fluent, authoritative, and well constructed.
- 3:50SPEAKER_02: The failure mode is hidden.
- 3:53SPEAKER_02: And that's precisely what makes getting the diagnosis right before you are actually committed to a technology so important.
- 4:02SPEAKER_00: And what we see in practice is that misdiagnosis at the start compounds very quickly.
- 4:07SPEAKER_00: We've worked with clients who've skipped the diagnostic work entirely, invested significantly, and ultimately had to walk away.
- 4:15SPEAKER_00: Not because the technology was wrong, but because the problem was never properly defined before it was selected.
- 4:21SPEAKER_00: By the time that becomes clear, the costs, financial, time, reputational, have already stacked up in a way that's very hard to unwind.
- 4:31SPEAKER_01: And I think that's the crux of it.
- 4:32SPEAKER_01: The wrong answer, reached early, tends to get embedded, which is why when that happens in those very first conversations, it means so much more than people realize.
- 4:43SPEAKER_01: So Suzanne, practically speaking, when a client comes to you, where should they actually be starting?
- 4:50SPEAKER_00: Well, I'd say, first of all, not with the technology.
- 4:54SPEAKER_00: That's the direct answer.
- 4:55SPEAKER_00: The moment you go straight to this tool, you skip the question that matters most.
- 4:59SPEAKER_00: What problem are we actually trying to solve?
- 5:02SPEAKER_00: And it sounds obvious and it is obvious, but the pull in the opposite direction is enormous.
- 5:08SPEAKER_01: It really is.
- 5:09SPEAKER_01: And we see two things driving that consistently.
- 5:13SPEAKER_01: First, it's the pressure, senior stakeholders pushing hard to invest in AI, a compelling demo in a boardroom, and suddenly that technology is leading rather than the problem.
- 5:24SPEAKER_00: And second, the second is habit.
- 5:27SPEAKER_00: As lawyers, we're used to doing things in a particular way.
- 5:31SPEAKER_00: And unless someone challenges that from the ground up, challenges the process, the assumptions, the way work actually moves, you may never unlock what the technology can genuinely offer.
- 5:42SPEAKER_00: You end up doing the same thing slightly differently.
- 5:46SPEAKER_01: So I guess what does the right order actually look like?
- 5:50SPEAKER_00: Problem first, process second, technology third.
- 5:55SPEAKER_00: Define what you're trying to achieve, map how the work actually moves, and only then decide where technology fits within that workflow.
- 6:03SPEAKER_00: When you do get to technology, Gen AI is 100% one option within a broader toolkit, but it is not the automatic starting point.
- 6:14SPEAKER_02: And just to add to that, when clients actually approach it in that way, what tends to surprise them is where the early gains actually come from.
- 6:23SPEAKER_02: And quite often it's not the AI.
- 6:26SPEAKER_00: Yes.
- 6:27SPEAKER_00: And this consistently catches people off guard.
- 6:30SPEAKER_00: The fastest gains often come before any AI is introduced at all.
- 6:35SPEAKER_00: Claiming up the data, clarifying rules, redesigning a workflow so that it reflects how work actually moves rather than how it's assumed to move.
- 6:44SPEAKER_00: The technology then has something solid to build on rather than being asked to compensate for foundations that were never in place.
- 6:52SPEAKER_01: I think that's something worth sitting with.
- 6:54SPEAKER_01: If you're listening to this and you're in one of those missing conversations, the vastest game might come before AI even enters the picture.
- 7:01SPEAKER_01: That's a significant shift in how to frame this whole exercise.
- 7:06SPEAKER_00: Exactly.
- 7:07SPEAKER_00: And the clients who get the most from their technology treat it as a design problem, not a procurement decision.
- 7:13SPEAKER_00: The question we always push them to start with is what is the decision or outcome you're actually trying to improve?
- 7:21SPEAKER_00: Get clear on that, and the technology question becomes much clearer.
- 7:25SPEAKER_00: Getting clear on it often means having some early, sometimes quite uncomfortable conversations, which is actually where we want to go next.
- 7:35SPEAKER_01: What we call the messy conversation.
- 7:38SPEAKER_01: And it's called that for a reason.
- 7:40SPEAKER_01: Our job isn't to say no, but it is sometimes to slow things down just enough, and ideally as early as possible, to understand why a client thinks a particular tool is the answer they're searching for.
- 7:51SPEAKER_01: And what problem are they actually trying to solve?
- 7:54SPEAKER_02: And teams need to genuinely get comfortable with that stage because the discovery phase is where the misdiagnosis gets prevented.
- 8:04SPEAKER_02: It is also the phase most often skipped or rushed simply because everyone feels the pressure to move quickly or their budgeting constraints.
- 8:15SPEAKER_00: And as lawyers, that instinct runs deep.
- 8:18SPEAKER_00: We're trained to move forward, progress matters.
- 8:21SPEAKER_00: That's generally the right instinct.
- 8:23SPEAKER_00: But when you're designing a solution involving AI, moving fast before you understand the current state of play creates real expensive problems down the line.
- 8:33SPEAKER_00: Being honest early means surfacing things like the actual condition of the data, the systems already in place, the client's internal processes that the solution needs to sit alongside or feed into, and the supervision, quality control, and approval steps that will need to exist.
- 8:51SPEAKER_00: Those things don't disappear just because Gen AI has come into the picture.
- 8:56SPEAKER_02: And if you don't surface these realities upfront, you can end up with a design that looks great on paper, but simply doesn't hold up.
- 9:05SPEAKER_02: You might discover you need to rebuild an entire workflow to introduce the proper oversight, or that there is an additional stage needed just to get the data into a usable state.
- 9:19SPEAKER_02: Or that outputs need to be transformed before the clients' systems can actually ingest them before manual rework.
- 9:29SPEAKER_01: And all of that affects the costs and the timelines, but also something that often gets underestimated, which is how people actually judge whether the project was a success.
- 9:39SPEAKER_01: When a project is perceived to have underdelivered, it can set back appetite for technology investment much more broadly than just this individual scenario.
- 9:48SPEAKER_01: And it's not just one project, it's the trust.
- 9:52SPEAKER_00: And that's the thing.
- 9:53SPEAKER_00: Having those messy conversations early doesn't slow down delivery.
- 9:57SPEAKER_00: It protects it, it sets realistic expectations, puts the right guardrails in place, and reduces the risk of costly course corrections once you're already committed.
- 10:08SPEAKER_00: The discomfort of that conversation is a feature, not a bug.
- 10:12SPEAKER_02: And those guardrails have to hold across all four layers: people, process, technology, and data.
- 10:20SPEAKER_02: That last one can't be simply treated as pumping in the background.
- 10:24SPEAKER_02: With Gen AI, data is part of the operating model.
- 10:28SPEAKER_02: And that's non-negotiable.
- 10:30SPEAKER_02: If it's messy, outdated, fragmented, that mess shows up downstream.
- 10:37SPEAKER_01: Exactly.
- 10:38SPEAKER_01: And it is 100% the most overused slide in legal tech delivery, but it is still the foundation.
- 10:44SPEAKER_02: Fair enough.
- 10:45SPEAKER_02: But it comes much less of a cliche once Gen AI enters the picture.
- 10:50SPEAKER_02: With traditional tools, you could work around the data quality issues to some degree.
- 10:54SPEAKER_02: The systems were relatively forgiving.
- 10:56SPEAKER_02: Gen AI isn't.
- 10:58SPEAKER_02: And as we said earlier, the failure mode is silent and hidden.
- 11:02SPEAKER_02: The outputs look authoritative even when the inputs aren't.
- 11:07SPEAKER_02: The scenario we often see plays out like this: a genuine, well-identified use case.
- 11:14SPEAKER_02: The process has been mapped, the technology has been selected, the team is ready to go.
- 11:19SPEAKER_02: You get into the data and it hasn't been maintained consistently.
- 11:24SPEAKER_02: Suddenly, you face an uncomfortable choice.
- 11:28SPEAKER_02: Invest unplanned time and resource getting it into shape, which hits your timeline and budget budget, or proceed on an unstable foundation.
- 11:37SPEAKER_02: Neither of these options is what the client actually signed up for.
- 11:42SPEAKER_01: And more often than not, the tool gets blamed.
- 11:46SPEAKER_02: Every time.
- 11:57SPEAKER_02: The technology isn't the problem, but the technology takes the reputational hit.
- 12:02SPEAKER_00: And the same can be said of the people layer, which works in exactly the same way.
- 12:07SPEAKER_00: You can have clean data, a sound process design, and the right technology.
- 12:12SPEAKER_00: But if the people involved haven't been brought along in the journey, adoption fails almost immediately.
- 12:18SPEAKER_00: And we're typically operating across large, diverse teams: lawyers, project managers, client-side contacts, sometimes third-party providers across multiple jurisdictions.
- 12:29SPEAKER_00: Getting genuine alignment across all of those groups requires deliberate, ongoing effort.
- 12:35SPEAKER_00: Not a briefing and a go live.
- 12:38SPEAKER_00: People need to understand not just how to use the system, but why things are changing and what their role is within it day to day.
- 12:47SPEAKER_01: Because when that investment isn't made, people default to what they know.
- 12:52SPEAKER_00: They find workarounds, they revert to manual steps, the process was designed to eliminate.
- 12:57SPEAKER_00: And the efficiency gains that justified the whole investment don't materialize.
- 13:02SPEAKER_00: The people layer is not the soft bit, it's structural.
- 13:07SPEAKER_02: And if I can just add to that, partial adaptation is often worse than no adaptation at all.
- 13:13SPEAKER_02: You lose consistency, control, so on and so forth.
- 13:17SPEAKER_02: And that's when risk creeps back in.
- 13:21SPEAKER_01: These workflows only really work when everyone is involved.
- 13:25SPEAKER_01: Legal teams, delivery teams, clients, and third parties.
- 13:28SPEAKER_01: They all have to understand their role and play their part.
- 13:32SPEAKER_01: Otherwise, all you've really built is a very clever tool, which never quite leaves the starting block.
- 13:38SPEAKER_02: Which brings us to tool selection, because the right tool for the job really matters here.
- 13:44SPEAKER_02: Gen AI is powerful, but it has real limits, particularly when volume, structure, and precision collide with one another.
- 13:53SPEAKER_02: We recently worked through an exercise for one of our legal teams involving 616 million lines of data.
- 14:02SPEAKER_02: It was spread across multiple files with widely different structures.
- 14:06SPEAKER_02: Some of them had a handful of fields while others had more than 100.
- 14:10SPEAKER_02: A classic case where Genei simply wouldn't work, not because of any failure in the technology, but simply because of the fundamental nature of the problem.
- 14:23SPEAKER_02: And yeah, if you're listening to this and thinking that sounds like an extreme example, it isn't.
- 14:28SPEAKER_02: That's the kind of skill that comes up in real transactional work.
- 14:34SPEAKER_01: It definitely does, which is why we almost never think about the tools in isolation either.
- 14:39SPEAKER_01: The real work is in designing the workflow and the connections first.
- 14:43SPEAKER_01: That means understanding the requirements, mapping how the work moves, connecting the data, the people, the process, and being clear about what output you actually need.
- 14:53SPEAKER_01: From there, you can then decide where standardization makes sense, where flexibility has to remain to make it operational, where automation can remove low-risk, repetitive tasks, and where analytics or search can bring structure, and where Gen AI can help with the more language-heavy parts of the work.
- 15:14SPEAKER_02: Casey, just going back to the previous example I used, 660 million lines of data.
- 15:20SPEAKER_02: Just let that sink in for a moment.
- 15:24SPEAKER_02: 660 million lines.
- 15:26SPEAKER_02: In settings like that, you often spend more time normalizing the inputs than you actually save on the outputs.
- 15:34SPEAKER_02: Where the task needs verbatim precision, exact extraction, exact close matching, consistently structured fields, so on and so forth, Gen AI is considerably less reliable than many people assume.
- 15:51SPEAKER_00: Exactly, Uchi.
- 15:52SPEAKER_00: And we see this in DD style work constantly.
- 15:55SPEAKER_00: The documents aren't neat, they're negotiated, inconsistent, and can often live in contract families.
- 16:02SPEAKER_00: Without strong processes wrapped around Gen AI in those settings, you don't reduce the quality control burden, you add to it.
- 16:10SPEAKER_00: And without the right guardwheels, pushing GNAI in can actually increase risk rather than reduce it.
- 16:17SPEAKER_01: Which is why we almost never think about tools in isolation.
- 16:21SPEAKER_01: We thread them together.
- 16:22SPEAKER_01: Workflow first to standardize the intake.
- 16:26SPEAKER_01: Automation to strip out predictable, low-risk tasks, analytics and search to structure the dataset, and then Gen AI, where the language understanding genuinely helps.
- 16:38SPEAKER_01: Summarising, clustering themes, drafting first-pass review issue descriptions for a human to then validate, and then take the unnecessary weight out of the model, build the right process around it, and you'll see how much better it actually performs.
- 16:55SPEAKER_00: And there's actually a really good illustration of that in the delivery tooling we use for verification work, where the efficiency doesn't come from the Gen AI tool at all.
- 17:05SPEAKER_00: It comes from strong processes and project management.
- 17:09SPEAKER_00: Tools that structure work, manage dependencies, surface issues early, and give visibility across large teams and timelines.
- 17:17SPEAKER_00: The value is driven by clarity, coordination, and control.
- 17:22SPEAKER_02: Once that foundation is in place, you are in a much stronger position to layer Gen AI ineffectively and to start thinking seriously about a gentic AI beyond that.
- 17:35SPEAKER_02: Gen AI becomes a component in a toolkit, not the whole toolkit.
- 17:40SPEAKER_01: And I think that's really the message here.
- 17:43SPEAKER_01: It's not about holding Gen AI back.
- 17:45SPEAKER_01: It's about using it deliberately, understanding where it adds genuine value, where it needs structure and support around it, and where human judgment still has to lead.
- 17:55SPEAKER_01: When you get that balance right, it's a powerful part of the solution, not just a magic wand.
- 18:01SPEAKER_01: So pulling all of that together, the common thread running through everything today is that the most important work should happen before any technologistic decision is made.
- 18:11SPEAKER_01: Getting clear on the problem, having those early, sometimes uncomfortable conversations about the data, process, and people, being honest about what the current state actually looks like.
- 18:22SPEAKER_01: Because the wrong diagnosis, made early, tends to then compound.
- 18:27SPEAKER_01: Decisions get layered on top of it, timelines get committed to, expectations get set, and then when the cracks appear, and they will, the cost of fixing them is always higher than it would have been if those questions had been asked at the very beginning.
- 18:42SPEAKER_00: I 100% agree, Casey.
- 18:44SPEAKER_00: And the practical thing I'd say to anyone listening in one of those early conversations right now, whether you're on the client side or the delivery side, is this.
- 18:54SPEAKER_00: The discomfort of that conversation is not a sign that something's going wrong.
- 18:59SPEAKER_00: It's a sign it's working.
- 19:00SPEAKER_00: The teams that navigate this best don't have fewer messy conversations, they have better ones and earlier.
- 19:09SPEAKER_02: And if there's one thing I would like to leave people with it, is this none of what we've described today is possible without a solid data foundation.
- 19:19SPEAKER_02: People process technology.
- 19:22SPEAKER_02: All of it ultimately depends on data being in the right shape, in the right place, and properly understood before you build anything on top of it.
- 19:32SPEAKER_02: Which is exactly why our next episode is dedicated to that topic alone, because data itself isn't part of the equation.
- 19:41SPEAKER_02: Increasingly, it's the equation.
- 19:45SPEAKER_02: I feel like we could keep that one going on.
- 19:47SPEAKER_00: We really could.
- 19:48SPEAKER_00: There's a whole episode just in the messy conversation.
- 19:52SPEAKER_01: Fair enough.
- 19:53SPEAKER_01: So our next episode is from Chaos to Clarity: Building the Data Layer That Unlocks AI.
- 19:58SPEAKER_01: And Uchi, I think you're the lead on that one.
- 20:02SPEAKER_01: It is.
- 20:02SPEAKER_01: And I'll see you then.
- 20:04SPEAKER_01: Thanks, Susanna Nucci, and thank you for listening.
- 20:07SPEAKER_01: We'll see you next time on Legal Tech Deciphered.
- 20:09SPEAKER_01: Trust, Truth, and Hallucinations.
- 20:12SPEAKER_00: Thanks everyone.
- 20:13SPEAKER_00: Goodbye.
- 20:14SPEAKER_02: Thanks all.
- 20:15SPEAKER_02: Goodbye.