📋 The Outline For Today

This week, I'm tackling a question that I believe is only going to get louder over the next 12-24 months: do we buy a specialist legal AI platform that looks purpose-built for our needs, or do we build on top of the LLMs our teams already use every day?

Here's the reality many legal ops directors won't say out loud: your associates are already using ChatGPT or Claude. They're drafting motions at 11 PM, summarizing depositions on their lunch break, and asking Gemini to punch up client emails before they hit send.

Meanwhile, the firm just spent $240K on a specialist platform that three people have logged into twice.

The question isn't if AI—it's whether you fight where your lawyers already work, or build workflows around the tools they're using anyway.

Most legal ops directors I talk to are stuck in the same loop.

Partners want "enterprise-grade" platforms with certifications. IT wants compliance and vendor contracts.

But nobody wants to spend 40 hours training lawyers on yet another system when ChatGPT significant proficiency can be achieved in a 2-3 hour workshop (and they're already using it daily).

This week: let’s take a look at why many legal ops teams have stopped fighting that reality and started building privacy-first workflows in ChatGPT, Gemini and Claude instead.

Plus, I’ve put together what I’ve found to be a useful and broadly applicable decision matrix to help you figure out when to skip the specialist platform, and when you shouldn't.

Let’s roll.

Here's what many smart legal ops teams figured out: you don't need lawyers to adopt new software if you build workflows where they already work.

ChatGPT, Claude, and Gemini can be good enough (with proper planning, guardrails, infra and governance) at many legal tasks and lawyers are already using them 40 times a day.

Juro, a contract AI platform, just announced they're meeting lawyers inside ChatGPT instead of forcing them to learn new software.

Translation: even specialist vendors admit lawyers already live in general LLMs.

In my experience, LLM-first typically wins when three things are true.

First: your volumes swing unpredictably. Commercial litigation teams might handle 15 document reviews one month and 50 the next.

Paying $1,000+/month per seat for specialist legal AI doesn't typically make sense when usage fluctuates and half your associates use it sporadically.

Second: your data lives across multiple systems. If you're pulling from iManage, email, Excel, SharePoint, and matter management platforms, a specialist tool that only reads one DMS creates more friction than it solves.

Third: speed-to-value beats niche features. Most legal AI workflows-contract review, clause extraction, legal research, due diligence checklists-don't require specialized training (or at least not much) if you structure the prompt correctly and filter data before it reaches the model.

Think of it like this: a universal remote instead of juggling multiple specialist platforms, or even one complex system with dozens of modules.

ChatGPT and Claude can orchestrate document review, draft pleadings, summarize depositions, analyze case law, and extract key contract terms without switching interfaces or learning new workflows.

The trade-off? You're responsible for the plumbing—building data security, retrieval accuracy, and human review checkpoints.

Specialist platforms do that for you - and regularly charge $12K–$15K per seat annually for the convenience.

Now, this approach isn't turnkey—you'll still need prompt engineering, API connections, data governance & more (plan for ~40 contractor hours).

But for high-volume, repeatable workflows like contract review and legal research, it's achievable without the complexity of specialist platforms.

Here's a handy way to decide - without a 37-slide deck.

Start with LLM-first for high-frequency, low-risk workflows—legal research, contract review, client intake summaries.

Prove ROI in 30 - 90 days and track where you hit accuracy or compliance ceilings.

Then, based on measured gaps, selectively add specialist tools for the 10-20% of work that demands it - e-discovery platforms for complex litigation, contract analytics for M&A due diligence.

This "prove, then buy" approach gives you firm-wide AI access from day one (~$6K/year + relatively minimal dev costs) without committing to $240K in platform seats upfront.

When you do buy specialist tools in month 6-12, you're purchasing 3-5 seats for proven use cases, not 20 seats for "maybe someday" workflows.

Anthropic, the makers of Claude, just came out with some pretty cool (aka nerdy) news.

This is the breakthrough that makes LLM-first viable for hesitant firms: privacy-preserving code execution with MCP (Model Context Protocol).

Think of it as a safe side-room where sensitive data gets cleaned and masked before the AI sees anything.

Traditional legal AI workflows send raw client documents to the model, e.g. ChatGPT. Names, case numbers, financial details-everything. Even with encryption, that makes risk-averse partners nervous.

It’s called ‘Code Execution’ and it literally flips the script.

The process runs locally (or in a sandboxed environment you control), masks sensitive fields, runs transforms, and only sends aggregated or tokenized snippets to ChatGPT, Gemini, or Claude.

Example: instead of uploading 500 client contracts to analyze indemnification clauses, the code execution layer extracts only the indemnification sections, replaces client names with tokens (Client_A, Client_B), and sends those masked excerpts to the model.

The AI never sees who the clients are. Your audit logs show exactly what data moved where. And you stay in control of the entire pipeline.

Anthropic recently published a deep dive on why "filter first, then prompt" changes the privacy and efficiency math for firms that avoided open LLMs.

Translation: you can use a general LLM for legal workflows with dramatically reduced exposure of raw client data.

Let's make this concrete.

Here's a workflow you can set up with off-the-shelf tools and a few contractor hours.

Use case: Analyze 200 settlement agreements to identify non-standard clauses.

Old way: Upload all 200 agreements to a specialist platform ($15K/year per seat), wait hours for processing, and trust the vendor's security holds.

Privacy-first way:

Time saved: ~12 hours of manual review compressed to 90 minutes.

Data exposed: Zero. The model never sees who the clients are or what they paid.

Cost: ~$50 in API fees (+ developer hours) for 200 documents vs $1,250/month for a specialist seat.

This is the type of legal AI workflow that gets partners nodding and gets some all important ‘points on the board’.

Second workflow: AI e-discovery without uploading privileged documents to third-party servers.

Use case: Review 5,000 emails for privilege and relevance in a commercial litigation matter.

Privacy-first approach:

Result: 5,000 emails screened in 4 hours instead of 20 hours. Full documents never leave your environment. Audit trail shows exactly what snippets were analyzed.

This is how firms that can't afford $150K>$300K/year e-discovery platforms still benefit from AI-powered document review.

Don't skip this part. Even, in fact especially LLM-first workflows need guardrails

Sandboxed execution: Code runs in isolated environment with resource limits—no access to full network.

Audit trails: Log every query, every retrieved document, every masked field, every model response. Store for 7 years.

Approval gates: No AI output goes to a client without associate review. Tag outputs as "Draft-Human Review Required."

Data residency: If you're in Canada or EU, confirm your code execution environment and API calls stay in-region.

Red-team testing: Before you scale, run 10 adversarial prompts to see if the masking layer can be bypassed. Fix holes before partners find them.

Policy-before-pilot. Lock down governance, then launch workflows. Reversing that order kills adoption when the first security question tanks momentum.

Want the one-page decision matrix and a governance checklist you can hand to IT?

Anthropic on code execution: Why "filter first, then prompt" changes the privacy math for law firms that avoided open LLMs. The breakthrough: data never leaves your control, but you still get GPT-4-class reasoning.

Read the post →

Juro meets lawyers in ChatGPT: Contract AI platform ditches proprietary UI and integrates via MCP, validating LLM-first as a credible legal AI adoption path.

Read the story →

In my experience adoption has a better chance of winning when you meet lawyers where they already work - ChatGPT, Gemini, Claude, email with privacy-first plumbing under the hood.

LLM-first isn't right for every legal team.

But if budget, speed, and flexibility matter more than vendor hand-holding, it's often the fastest path from "we should try AI" to "we're using AI daily."

Got a curveball scenario or strong opinion on this?

Please do reply, I read them all of course.

Catch you next week, I hope it’s a great one.

— Liam Barnes

P.S. Want our team to guide you on whether subscribing to Harvey or building inside of Claude is right for your team? Grab time for an introductory call.

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]