legal-tech · AI · Harvey

Custom AI vs. Off-the-Shelf Legal AI Tools: Harvey, Lexis+, and When to Build Your Own

An honest comparison of buying Harvey, Lexis+ Protégé, or CoCounsel versus building custom legal AI. When off-the-shelf wins, when custom wins, and how to decide.

Evgeny Smirnov ·

The market has exploded. Harvey raised over $100M and hit recurring revenue milestones faster than most enterprise SaaS companies. Thomson Reuters spent $200M acquiring Casetext and integrating CoCounsel into Westlaw. LexisNexis launched Protégé. Luminance, LegalOn, and dozens of smaller players compete for every corner of the legal workflow.

If you’re a technology leader at a law firm, legal publisher, or legal services company, you’ve got more options than ever. The question isn’t “should we use AI?” — it’s “should we buy, or should we build?”

We’ve done both. We’ve built custom legal AI systems — ChatBook tools for the AAA, the PlanYourSunset estate planning platform, AI engines for law firms — and we’ve also evaluated off-the-shelf options for clients and recommended them when they were the better fit. So our perspective here is pragmatic, not self-serving.

What the off-the-shelf leaders actually do well

Harvey excels at general legal research and drafting. It’s trained on a broad legal corpus and handles common workflows — memo drafting, contract review, case research — competently. For a large firm doing high-volume, varied work, it reduces research time meaningfully.

Lexis+ Protégé and Westlaw’s CoCounsel sit on top of the world’s largest legal databases. When your queries map to published case law and statutes, these tools deliver well-cited results. The integration with existing Westlaw/Lexis workflows is seamless.

Luminance focuses specifically on contract intelligence. For organisations processing high volumes of contracts, it’s purpose-built and effective.

These are real products solving real problems. We recommend them to clients when they’re the right fit. The question is: when aren’t they?

Five factors that determine the answer

The first is proprietary data. If your AI needs center on publicly available legal information — case law, statutes, regulations — buy. If your competitive advantage comes from proprietary content — internal research archives, unpublished arbitration awards, firm-specific precedent banks, specialized industry knowledge — build. Off-the-shelf tools don’t have your data. Some let you upload documents, but they treat uploads as supplementary, not primary. This is exactly why we built a custom system for the AAA rather than deploying Harvey — their journal archive and arbitration materials represent decades of proprietary knowledge that no off-the-shelf tool could leverage.

The second is workflow specificity. If your workflows align with common patterns — research, drafting, contract review — buy. If you need AI integrated into specialized workflows — custom intake processes, proprietary scoring models, specific compliance procedures — build. PlanYourSunset is a clear example. Estate planning involves jurisdiction-specific rules (NY only, for now), complex family dynamics, and workflows that combine legal document generation with client-facing questionnaires and an AI assistant named Larry who explains legalese in plain English. No off-the-shelf legal AI tool handles this.

The third is cost at scale. At fewer than 50 users with standard usage, buying makes sense. At 100+ users, the math changes fast. Harvey and similar tools charge $100–$300/user/month for enterprise plans. For 200 attorneys, that’s $240K–$720K annually — every year. A custom solution costs $100K–$200K to build, with $30K–$100K in annual infrastructure and maintenance. By year two, custom is cheaper. By year three, dramatically so. And the custom solution appreciates — it gets better with data and user feedback. The SaaS subscription stays the same.

The fourth is accuracy requirements. Off-the-shelf tools optimise for breadth across all practice areas. Custom solutions optimise for depth in your specific domain. A custom RAG system built on your content, with domain-tuned embeddings and a verification layer, will outperform a general tool on your specific use cases — often dramatically. If general accuracy (85–90%) is acceptable, buy. If you need near-perfect accuracy for specific content types, build.

The fifth is control and compliance. If standard data handling meets your requirements and you’re comfortable with data flowing through a third-party provider’s infrastructure, buy. If you need full control over data residency, model behavior, audit trails, or regulatory requirements demand it — build.

“The build vs. buy decision isn’t really about technology. It’s about where your competitive advantage lives. If AI is a productivity tool for your team, buy it. If AI is your product or directly enables your product, build it. The difference matters enormously.”

— Evgeny Smirnov, CEO and Lead Architect:

The hybrid approach most organisations miss

The smartest organisations don’t choose one or the other — they do both.

Use off-the-shelf tools for general legal work. Let your attorneys use Harvey or CoCounsel for routine research and drafting. These tools are good at this, and the per-user cost is justified by time savings.

Build custom for your differentiator. Invest development resources in the AI capabilities that set your organisation apart — making proprietary content searchable, automating specific workflows, embedding AI into client-facing products.

This approach gets you 80% of the value from off-the-shelf tools while building defensible competitive advantages with custom development. And it lets you start custom projects with narrower scope, reducing risk and cost.

Decision framework

Answer honestly: Does your AI need depend on proprietary data? Are your workflows specialised or industry-specific? Will you have 100+ regular users? Do you need domain-specific accuracy above 95%? Do you have strict data control or compliance requirements?

Three or more “yes” answers point toward custom development. One or two: start with off-the-shelf and build custom only for specific differentiating use cases.

If the framework points toward building: 6–10 weeks for an MVP covering your core use case. 3–6 months for a production system. This assumes a team that’s done it before — first-timers should double these estimates.

Budget: $40K–$80K for MVP, $80K–$200K for production. Ongoing costs of $2K–$8K/month for infrastructure and maintenance.

What you need from your side: access to your content in digital format, 3–5 domain experts for requirements and testing, a designated product owner who can make scope decisions, and clear success criteria.

What you get: a system tailored to your content, your workflows, and your accuracy requirements, running on infrastructure you control, with no per-user licensing fees.


Not sure which approach is right? Book a consultation — we’ll review your use case, evaluate off-the-shelf options honestly, and recommend the path that makes the most sense. No obligation to build with us.