You've been told that enterprise legal AI platforms are the only serious option.

That open-source is for small-firms and budget-constrained solos.

That Harvey's upfront implementation costs and ~$1,000+ per lawyer monthly fees and Lexis+ AI's per-use fees are justified by their performance edge.

Here's what recent performance tests actually show: specialized open-source models now score 8.2-8.5 out of 10 on legal tasks, compared to Harvey-equivalent platforms at 8.7 out of 10.

That's a performance gap of less than 1%.

Meanwhile, the cost difference over three years can exceed ~$400,000 (see the table below).

The expensive platform you've been pitched isn't necessarily the wrong choice.

But it's no longer the only credible choice for your firm.

And if you're making this purchase decision based on outdated assumptions about performance, cost structure, and strategic control, you might be overpaying for overhead you don't need.

As you can see in the chart above, the performance gap is minimal, but the cost and control differences are substantial.

Let me walk you through what's driving this shift.

Let me be clear: Harvey has genuine data advantages.

Multi-jurisdictional case law, specialized legal training sets, and millions in R&D investment.

That's real, and I'm not naive enough to dismiss that advantage.

A July 2024 study published on arXiv showed early signs that fine-tuned Llama 3 models could outperform GPT-4 on specific classification tasks.

But the real turning point came 6 months ago, in June 2025.

The VLAIR (Vals Legal AI Report) Benchmark revealed that AI tools had finally surpassed human lawyer baselines in 4 out of 7 substantive legal tasks, including Document Q&A and Summarization.

Then, just last month in November 2025, the Legal AI Benchmark (Phala) tested 12 leading models across 163 legal tasks.

The results confirmed a new reality: multiple non-proprietary models now score in the same performance band as leading enterprise models, with differences of only a few tenths of a point on 10-point scales.

That's a negligible gap for most practical purposes.

On your firm's specific workflows - contract review, discovery, research memo drafting, does that 0.2-point edge justify a $155K-$465K three-year premium?

Maybe.

But it's a question worth asking.

As recently as early 2024, this wasn't a choice, open-source had significant quality gaps.

The June 2025 and November 2025 benchmarks prove the inflection point happened this year.

Law firms and in-house teams have two main approaches to building with open-source models:

Approach 1: Fine-tune an open-source model exclusively on your own documents (we briefly covered this topics in last week’s newsletter).

This creates a private, firm-specific AI that understands your templates, clauses, and workflows.

Think of it like training an associate who only learns from your firm's past work product.

Enterprise AI is undoubtedly more expensive to build than open-source deployments, but not by the price premiums you're paying.

I’ve worked in SaaS for over 20 years and I can tell you that a significant portion of subscription fees covers overhead: sales, marketing, customer success, and profit margins - not just technology.

As of the time of publication, market estimates suggest Harvey pricing ranges from $1,000-$1,200 per lawyer per month, with additional onboarding fees.

For a 50-lawyer firm, the total three-year cost is approximately $1.8 million to $2.2 million.

Lexis+ AI typically operates as a premium subscription add-on, often costing an additional ~$300-$500+ per user per month on top of your existing research fees.

Open-source eliminates those overhead costs.

This doesn't mean Harvey isn't investing in AI and product development.

It means their business model structurally requires pricing that covers services and overhead you may not need.

That's a legitimate trade-off, not a criticism.

But it's a trade-off you should evaluate deliberately.

With performance gaps now negligible (8.8 vs 8.6), this isn't about performance versus budget.

It's about convenience versus control.

Here's the comparison laid out clearly for a 50-lawyer firm:

Enterprise offers turnkey convenience.

Open-source offers control and ownership.

And if you work with a reliable partner to plan, build, and manage this, it's largely hands-off for the firm or in-house team….Cyberaktive is one such company ;).

Different firms need different things.

A 25-attorney litigation boutique with no IT capacity might genuinely need Harvey's managed service.

A 95-attorney corporate firm with an operations team might be better served building in-house capabilities that become institutional competitive assets.

The question isn't which option is objectively better.

The question is: which trade-offs matter most to your firm's specific context, risk tolerance, and strategic priorities?

Your confidential client data is your most valuable asset.

When you process M&A documents, litigation strategy memos, or sensitive regulatory filings through a vendor platform, you're introducing third-party processing into your confidentiality chain—even with strong contractual protections.

Harvey's Platform Agreement states they "may collect and use Usage Data to develop, improve, support, and operate its Service," while also clarifying "Your data stays yours. We don't use inputs/outputs to train models."

That's transparent and better than many alternatives I’ve read.

But for your most sensitive matters, on-premise processing eliminates the question entirely.

This isn't paranoia.

I’d say it’s just responsible risk management.

For routine contract review, vendor convenience might be perfectly acceptable.

For your client's $500M acquisition or regulatory investigation defense, keeping processing fully in-house makes strategic sense for many Managing Partners & Board members.

Firms building in-house today won't be vendor-dependent in three years, that infrastructure becomes a competitive moat.

Clifford Chance announced in February 2024 that they built their generative AI platform on Azure OpenAI, maintaining full control over their AI infrastructure.

Wilson Sonsini built their proprietary Neuron platform, creating differentiation in M&A that's difficult for competitors to match.

Vendor-dependent means recurring fees forever.

In-house can mean compounding value and strategic control.

Vendor contracts are designed to make switching expensive.

Three-year terms, proprietary integrations, price escalation clauses, and migration complexity all create exit friction.

Before you commit to any platform, understand what it would cost to leave.

That determines whether you're making a flexible choice or a long-term bet you'll be locked into regardless of how the market evolves.

A) 100% rented from major vendors (Harvey, Lexis+ AI, etc.)
B) A hybrid mix: Rented platforms + some internal tools
C) Mostly in-house: We will own our own models and data infrastructure
D) We are still waiting to see what happens

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

The inflection point happened in 2024.

The performance gap that justified enterprise premiums six months ago no longer exists for most legal workflows.

What remains is a strategic choice: pay for managed convenience or build institutional capabilities that compound over time.

The firms making this decision deliberately—not defaulting to the expensive option—will own their AI future.

The ones that don't will rent it.

What's your take on this shift?

Hit reply and let me know where your firm stands.

— Liam Barnes

Not sure where to begin automating your workflows? Or the best to leverage AI in order to show an ROI?

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.

Last Week’s Reader Poll Results