2025 was the year legal AI moved from “pilot theatre” to everyday usage — but not in the way many predicted. Our Moderator Horace Wu observed that adoption has become broad but shallow: lots of firms using generative AI in some form, but comparatively few changing how legal work is truly delivered.
2025: What we learned (the reality check)
Looking back over the year that was, Wu shared significant insight on what he has seen in practice.
1) Widespread adoption, limited depth
In short, Wu felt that he underestimated willingness to adopt, but overestimated the impact. It is clear that, many firms have access to tools, but relatively few have embedded them into repeatable workflows that change cost, speed, or risk meaningfully.
2) “Deep” use cases clustered in three places (for AM Law 200)
The discussion highlighted three areas where adoption has been most significant at the top end of the market:
- Grid / matrix-style review across large document sets (popularised by tools built around structured extraction at scale).
- Productivity copilots that “grease the wheels of commerce” (e.g., translation, drafting, email assistance).
- Litigation + eDiscovery support, which Wu described as a bit behind transactional use cases — but powerful where it lands.
3) One market, many markets
A repeated theme throughout the session was that legal is highly fragmented, and product-market fit varies dramatically between AM Law 100/200, mid-market firms, boutiques, in-house teams, and government. A workflow that’s valuable for a global transactions practice may be irrelevant for a small firm that simply doesn’t handle that kind of volume or complexity.
4) Selling legal AI depends on who holds power inside the firm
For Legal tech vendors Wu’s take was highly pragmatic on the best approach for sales:
- Selling to Partners tends to favour narrow, practice-specific solutions (because Partners optimise for their own matters and margins).
- Selling to innovation teams favours platforms (because they want fewer tools, less vendor risk, and broader coverage).
He pointed to “power broker” dynamics too- when a rainmaker Partner insists on a tool, procurement tends to move fast.
5) Funding didn’t vanish — but scrutiny increased
2025 saw the highest level of investment ever into Legal Tech and, Wu noted continued VC interest in legal AI, with growing pressure on vendors to prove ROI and durable revenue (unless they’re viewed as category-dominant bets). As a concrete signal, he discussed Hebbia’s Ryan Samii moving to Harvey as Head of Product Innovation, interpreting it as a sign of strategic repositioning in the market.
2026 – The Year Ahead and Wu’s forecast
what will happen, won’t happen, and might happen
What will happen (soon)
- More sales of “agent” capabilities, even if vendors increasingly rebrand them as workflows.
- New “frontier” models (GPT-6, Gemini 4, Claude 5, etc.) — but Wu argued they’ll make almost no difference to most legal outcomes because models are already “good enough” for many tasks.
- More acquisitions/shutdowns across legal AI.
- Early experimentation by firms with subscription-style products for clients.
What won’t happen
- AGI (Wu called it plainly).
- Fundamental change to law firm business models at speed.
- Generative AI “solving” access to justice (the problem is bigger than the tooling).
- The removal of humans from most legal practice: legal remains human-led.
What might happen (and could reshape behaviour fast)
- A high-profile negligence/malpractice claim linked to AI use, against a globally recognised firm. Wu’s point wasn’t about who wins — it’s that the mere existence of such a case could drive firms to reduce public AI claims and tighten governance.
- A new line item: “legal AI audit” — where in-house teams do more work using AI, then ask external counsel to sign off (with a premium charged for that assurance).
- A “mischievous” move: someone open-sourcing a workflow platform that commoditises parts of the market — potentially shifting value back to law firms because they can uniquely combine tech with trust.
The skills shift: better lawyers, not just faster lawyers
The discussion emphasised that, in a tech-heavy era, the differentiator is the human side: listening, context, judgement, and client connection. All attendees agreed that in 2026 the human element won’t disappear — but it will be applied differently.
A practical framework: what AI changes in legal work
When describing new methodlogy for Legal work Wu cited three ways AI will affect delivery:
- Eliminate humans from certain routine tasks (e.g., straightforward NDA review).
- Accelerate tasks dramatically (hours down to minutes).
- Enable work that historically wasn’t feasible (e.g., reviewing every document in diligence instead of using strict materiality cut-offs).
He flagged the biggest unresolved issue as pricing: if competitiveness rises and time falls, what exactly are clients paying for — and what does “value” mean?
“Accuracy” vs “uncertainty” (and why it matters)
One point that Wu argued was that the key challenge isn’t simply hallucinations — it’s uncertainty. Generative AI can increase throughput, but it can also increase doubt about what is right. His prediction: we’ll see stronger demand for more deterministic tooling alongside generative AI, so lawyers can reduce uncertainty faster and more reliably.
Advice for smaller firms: start simple, prove value
The session closed with the focus on the needs on smaller Law Firms. Wu’s practical path for firms without Harvey/Legora-level budgets was:
- Use ChatGPT / Claude / Gemini at an enterprise-appropriate tier first.
- Pick tasks you would normally give a paralegal or first-year associate.
- Give the AI the same instructions — and use the speed of turnaround as the internal proof-point.
In this way respective Ai promoters within firms can demostrate the valuable use cases and encourage adoption.
