You're in a partner meeting when someone asks, "Should we be using Harvey or Claude?"

Your ops team forwarded a Gemini demo invite yesterday. Someone from the executive committee wants to know which AI platform justifies the budget allocation.

You've sat through vendor demos. You've read the case studies. And you still don't have a decision framework that doesn't feel like guessing.

Here's the thing: you're asking the wrong question.

The same model that crushes contract review might fail spectacularly at legal research.

The "best" model for intake automation is completely different from the "best" for due diligence. In my experience, most firms optimize for the wrong variable—they chase the highest-scoring or even the most popular model instead of matching tools to specific workflows.

According to LegalBench benchmarks from VALS.ai updated through December 2025, an 87% accurate model solving 90% of your workflow beats a 90% accurate model solving 30% of it.

Let's start with the most commonly implemented workflow first…

Many firms who deal with lengthy contracts lose hours on contract review not because their AI lacks accuracy, but because it can't see the whole document at once.

Here's the pattern: You've got a 50-page MSA with exhibits and a complex indemnification schedule. 

You feed it into AI. The model chunks the document into sections to process separately. 

It misses how the indemnification clause in Section 12 connects to the liability caps in Section 8. 

You've spent hours reconciling disconnected outputs.

This tends to be the hidden problem many firms don't realize they're solving incorrectly.

My pick: Gemini 3 Pro

The key insight isn't raw accuracy scores. It's a context window (the amount of text the model can process at once without breaking it into smaller pieces).

Gemini 3 Pro's 1 million token context window often matters more than marginal performance differences.

Think of it like this: you can walk the entire document into the room at once instead of bringing pages one at a time. 

That context window means the model sees the whole contract, understands how indemnification connects to liability caps, and catches dependencies human reviewers sometimes miss in first passes.

Real-world results from hybrid AI plus human review: 98% accuracy when using proper verification systems, according to Concord.app deployment data.

Performance: 87% on LegalBench legal reasoning tasks, according to VALS.ai benchmarks from December 2025.

Legal research carries higher risk than contract review. One hallucinated citation in a court filing can trigger sanctions.

Let me explain the corrected data clearly, because this distinction matters and resolves what seems like contradictory findings.

The Three Types of AI Legal Tasks:

When large language models (LLMs, which are basically sophisticated AI trained on massive amounts of text) are asked to recall specific legal facts from memory (like "Who wrote this opinion?" or "What's the holding?"), they can make stuff up - let’s just call it what it is. 

Models frequently invent answers rather than admit uncertainty (Tip: good prompting and safeguards can help protect against this).

When you give an LLM guardrails so that they only retrieve cases from verified databases using RAG, they tend to hallucinate less often versus when they are recalling their general knowledge.

When models reason about information you provide in the query (like "Here's the case facts and law, what's the analysis?"), accuracy hits 87% according to VALS.ai LegalBench benchmarks.

The key insight: Hallucinations tend to be more likely to happen when you ask AI to remember. Accuracy increases when you ask AI to reason with facts you provide.

Process tends to matter as much as model selection.

Model winner: Perplexity (with real-time web search)

This isn't a pure LLM. It's a hybrid retrieval plus generation system. Every citation includes a visible link to the original source. You can verify immediately instead of discovering hallucinations during discovery.

Why this matters: for legal research specifically, verifiable sources often beat raw model accuracy.

Client intake has lower legal risk than research or contracts, but getting it right matters. 

Poor classification delays cases. Missed urgency signals can blow deadlines. And an inappropriate tone can set the client relationship back before you barely started.

The challenge isn't finding the fastest model—any AI is exponentially faster than manual intake. 

What matters is balancing practice area classification accuracy, urgency detection, client communication quality, and cost efficiency for high-volume scenarios.

Model winner: GPT-5.1

For most firms, GPT-5.1 hits the sweet spot. 85-92% accuracy on practice area classification is often sufficient when downstream human review catches errors. 

It excels at urgency detection (reliably flagging "tomorrow," "arrested," "emergency" scenarios), and at the lowest cost per interaction, it makes economic sense for high-volume intake.

Alternative: Claude Sonnet 4.5 tends to be better for practices where client communication quality matters more than cost—think family law, immigration, or criminal defense where distressed callers need empathetic, nuanced responses.

Real-world impact: legal intake AI implementations report 35% higher client conversion rates when using AI classification, according to LawPractice.ai and golawhustle.com legal intake AI analysis from November 2025.

You might be thinking: "Does this mean I need a bunch of different subscriptions? Do my lawyers need to log into Gemini for contracts, Perplexity for research, and GPT for intake?"

Absolutely not. That would be a productivity nightmare.

The most sophisticated firms aren't asking their lawyers to toggle between tools. They are building unified workflows that route the work to the right model automatically.

Here is how it works in practice:

Imagine a single "Contract Review" portal on your firm's intranet.

The lawyer never touches the models directly. They just touch the process.

Why this approach wins:

This is the difference between "buying AI tools" and "building AI capabilities." One adds complexity to your day; the other removes it.

When we help firms and in-house teams implement this, we don't start by buying subscriptions. We start by mapping the workflow, then we wire the right models behind the scenes so the technology feels invisible.

A) We use one model for everything
B) We match different models to different workflows
C) We're still evaluating and haven't committed to any models yet
D) We rely on whatever our enterprise vendor provides
E) We don't use AI for legal work yet

[Reply with your letter choice] - I'll share the results in the next edition.

The firms and in-house teams making model selection decisions deliberately (not just defaulting to the most expensive option) will inevitably get better outcomes.

But the real winners won't just be the ones picking the right models. They will be the ones who make those models invisible.

By building workflows that route work to the best tool automatically, you stop asking lawyers to be "prompt engineers" and let them go back to being lawyers—just faster and more accurate ones.

Of course, getting there requires context. When we build this for clients, we look at your existing tech stack, your budget, and exactly where your current process is bleeding time.

The key is treating AI not as a software purchase, but as a workflow design challenge.

— Liam Barnes

Need help building these invisible workflows?

We help legal teams move beyond "chatbots" to build automated systems that match the right tool to the right task.

Grab some time to chat

(if you don’t see a suitable time, just shoot me an email [email protected])

How Did We Do?

Your feedback shapes what comes next.
Let us know if this edition hit the mark or missed.

Too vague? Too detailed? Too long? Too Short? Too pink?

👍 Loved it | 👎 Not for me | 💡 I Have a Suggestion

OR: [email protected]

Was this week’s newsletter forwarded to you?

Sign up, it’s free.