But the picture has gotten more interesting.
Something real has started to emerge underneath the noise. Not the magic wand version of AI that vendors have been selling. Something quieter, more precise, and in the right hands, genuinely remarkable. To understand it, you have to separate two things that the industry keeps conflating.
Two Very Different Kinds of AI
Most conversations about “AI in legal” treat the category as monolithic. It isn’t. There are two distinct buckets, and they behave very differently in what they promise, in what they deliver, and in where the risks live.
Bucket One: The Chatbot Layer
This is the AI most people have in mind — ChatGPT, Claude, & Gemini. Large language models that can draft, summarize, research, and respond in fluent, confident prose.
The legal industry’s wariness here has been well-founded. Our caution in the first article in this series was directed squarely at this bucket, and it holds. The “easy button” problem is real: these tools produce output that sounds authoritative, that reads as polished and complete, and that can be wrong in ways that aren’t immediately obvious.
That last point is worth sitting with, because it’s the crux of the risk.
The danger with LLM tools isn’t that they’re bad. It’s that they’re confidently wrong — and catching the errors requires the very expertise that organizations are tempted to skip when they reach for AI in the first place. A hallucinated case citation doesn’t announce itself as a hallucination. A subtly mischaracterized legal standard doesn’t come with a warning label. The output looks like the work of someone who knows what they’re doing, which means you need someone who knows what they’re doing to evaluate it.
Here’s what that looks like in practice. A partner at a litigation firm recently described how his team used AI to draft five motions simultaneously — work that would normally take two weeks or more. They did it in three days. But here’s what made it work: a senior associate spent significant time validating every document, researching the relevant case law, and replacing AI-generated examples with stronger, more precise ones. The partner did a thorough final review before anything went out.
The AI didn’t replace the expertise. It created the conditions for the expertise to operate at a different scale and level of efficiency. Without the partner’s judgment and the associate’s research, those drafts would have been a liability. With it, they were a genuine competitive advantage — work that compressed a two-week timeline into three days and saved the client a substantial amount of money.
That story reframes the “confidently wrong” risk in an important way. It’s not an argument against using these tools. It’s an argument for understanding exactly what role human expertise plays in the workflow — and refusing to cut corners on it.
Bucket Two: The Builder Layer
This is the less-discussed, more transformative development. Tools like Claude Code and similar AI-assisted development platforms don’t just generate text — they write functional software. And that shift has changed the economics of custom automation in a fundamental way.
For decades, building custom software was expensive enough that organizations reserved it for large, broadly applicable problems. The cost and time barrier meant that thousands of smaller, solvable inefficiencies just stayed unsolved. A workflow that wasted four hours a week wasn’t worth a six-figure development project.
That math has changed. Significantly.
AI-assisted development has collapsed the cost of building targeted, custom tools. A well-defined problem that a knowledgeable operator understands deeply can now be addressed with a purpose-built solution in a fraction of the time and budget it once required. Every workflow bottleneck, every repetitive manual step, every process that exists because “we’ve always done it this way” is now a potential candidate for a custom fix.
This is not theoretical. It’s happening — in legal ops, in eDiscovery, in document management, and in case administration. The organizations seeing the most benefit aren’t the ones deploying AI broadly. They’re the ones identifying specific, high-friction problems and building precise solutions for them.
What This Looks Like in eDiscovery
The most interesting AI applications in eDiscovery right now are not easy buttons. They are precision instruments — and they are built and operated by people who understand the underlying work at a deep level.
What they actually do is create shortcuts within hybrid workflows. They reduce friction at specific chokepoints. They automate the tedious middle steps of a process that an expert already knows cold. The expert doesn’t disappear from the equation. They move to a higher level of the work.
Think of it less like hiring a junior associate to handle a task, and more like giving a master craftsman significantly better tools. The output is better because the person is better — the AI just allows that person to operate faster, at greater scale, with less wasted motion.
In practice, this looks like AI-assisted culling that a senior reviewer tunes, monitors, and validates. It looks like targeted classification tools built around the specific document types in a particular case — not a generic model, but something calibrated to the matter at hand. It looks like custom scripts that handle the repetitive middle steps of a process, freeing the expert to focus on the judgment calls that actually require their judgment.
The pattern is consistent: the most valuable applications are highly specific, operated by highly knowledgeable people, and embedded in workflows where human oversight isn’t optional — it’s the point.
The Operator Is Still the Variable
The through-line across both buckets is the same: the quality of the output is inseparable from the quality of the person using the tool.
An experienced attorney gets more out of an LLM draft than a novice — because they know what to look for, what to question, and what to fix. A domain expert builds better custom tools than a generalist developer — because they understand the problem at a level that shapes every design decision. A senior reviewer who understands the case produces better AI-assisted culling results than one who doesn’t — because the model needs to be guided, not trusted blindly.
This is why the “AI will replace lawyers and reviewers” framing continues to miss the point. The correct framing is: AI dramatically multiplies what a knowledgeable operator can accomplish. It raises the floor for everyone and raises the ceiling for experts.
The partner who compressed two weeks of motion practice into three days didn’t do it by removing expertise from the process. He did it by deploying expertise more efficiently. The AI was the accelerant. The expertise was still the engine.
Where Lucent Stands
Our position hasn’t reversed — it’s refined.
We were right to be skeptical of AI as a magic wand. We remain skeptical of organizations that deploy these tools without the expert oversight required to catch what they get wrong. The confident-but-wrong problem is real, and it doesn’t go away just because the output looks polished.
But we’re paying close attention to the builder layer — to targeted automation, to custom tools built around specific, well-understood problems, to the precision applications that emerge when deep domain expertise meets AI-assisted development. This is where the genuine value is accumulating, quietly and without much fanfare.
The organizations that will get the most out of AI aren’t the ones who adopt it fastest. They’re the ones who understand it most clearly — who know which bucket they’re working in, what role expertise plays, and where the risks live.
That kind of clarity is what Lucent is built around. It’s what we bring to forensics, to eDiscovery, to managed review, and to the AI questions our clients are increasingly asking us.
The expert in the machine isn’t the AI. It’s still you. AI just lets you do more with it.