Best AI for Legal Drafting in 2026: Harvey, Spellbook, CoCounsel, and ABA Opinion 512

The Best AI for Legal Drafting Depends on Your Firm — Not the Hype Cycle

The best AI for legal drafting in 2026 isn't a single product. It's the right product for your firm's scale and matter mix: Harvey for BigLaw enterprise, Spellbook for in-Word contract drafting, CoCounsel for research-backed workflows, and enterprise-tier ChatGPT or Claude (with strict guardrails) for solo and small-firm work.

Pricing tells you most of what you need to know about fit. Harvey runs roughly $1,000+ per lawyer per month with a 20-seat minimum¹. CoCounsel Core starts near $225/month¹. So "best" is a math problem before it's a feature comparison.

Spellbook compares contracts against more than 2,000 industry benchmarks and runs natively inside Word²— useful framing if your associates already live in track changes. Harvey publicly reports a sub-2% hallucination rate, with roughly 0.2% on internal evaluations of its case-law model³. Treat that as Harvey-claimed, not as an independent benchmark.

If you're using enterprise ChatGPT or Claude as your primary drafting layer, the guardrails are non-negotiable:

• A no-training, no-retention contract on the tier you're using
• A documented matter-intake step that decides what data may be entered at all
• A named human reviewer for every output before it leaves the firm
• A logged audit trail (who prompted, what was returned, what was used)
• A written matter-by-matter exception list when a client opts out

Pricing in this space is vendor-reported and moves quarterly. Verify on the vendor's site before you sign. And remember: an AI strategy advisory engagement that maps tools to your real matter mix tends to pay for itself the first time it stops a wrong-tier purchase.

Picking the tool is the easy part. What you owe the client when you use it is where most firms have already failed.

What ABA Formal Opinion 512 Actually Requires

ABA Formal Opinion 512 was issued July 29, 2024⁴ and it does the work most firm policies are still pretending to do. It requires lawyers to disclose generative AI use whenever its outputs influence significant decisions, affect billing, or substantively contribute to research, analysis, or drafting⁵⁶— and it explicitly states that boilerplate consent included in engagement letters is not adequate when client confidences are involved⁷.

"Boilerplate consent included in engagement letters will not be adequate." — ABA Formal Opinion 512⁷

Op. 512 sits on top of Model Rules 1.1, 1.4, 1.6, 1.5, 5.1, and 5.3. In plain terms: competence, communication, confidentiality, reasonable fees, and supervision⁸. All of those are now load-bearing when AI touches a matter.

Disclosure is triggered when:

• AI outputs will influence significant decisions in the representation⁵
• GAI use could affect what the client is billed⁵
• AI substantively contributes to research, analysis, or drafting⁶
• Client information is being entered into a self-learning tool (informed consent required up front)⁹

The 2026 ABA practical guidance draws the line at substance: language refinement and clarity polish don't trigger disclosure; substantive AI contribution does⁶. That distinction is the one most firms haven't internalized. And state bars are layering rules on top — check yours before publishing any policy.

Op. 512 reads informed consent as a comprehension standard, not an exposure standard. A sentence the client never reads isn't consent. Editorial framing, not legal advice.

If boilerplate fails the standard, what passes? Compare what most engagement letters say to what informed consent actually sounds like.

Footer vs. Real Disclosure — A Side-by-Side

A real AI disclosure names the tool, explains how it's used in this matter, describes how client data is protected, acknowledges hallucination risk and the human-review process, and invites the client's questions. Five elements— one disclosure that turns a footer into a conversation.

Look at the difference.

Read those side by side and the footer's problem is obvious. It's a notification. The substantive version is a conversation.

What each line of the substantive version actually does:

1. Names the tool and tier. Some courts already require the specific product, not "AI software"¹⁰.
2. Names the use. Op. 512's substance test runs on what AI is actually doing on this matter⁶.
3. Describes data handling. This is the privilege protection — and the confidentiality duty under Rule 1.6⁸.
4. Names a human reviewer. This is competence (Rule 1.1) and supervision (5.1/5.3) made visible⁸.
5. Acknowledges hallucination. Sets expectation honestly; protects credibility when something has to be corrected.
6. Invites opt-out. Comprehension means the client could say no. That's what makes consent informed⁹.

A boilerplate sentence is a notification. Informed consent is a conversation. If your disclosure could be copy-pasted into any matter, it isn't disclosure of this matter. Sample language from Clio and Texas Bar resources is a starting point — adapt with bar counsel before deploying.

Disclosure substance is one half. The other half is whether the tool you're using to draft is even compatible with privilege.

The Privilege Problem — Public vs. Enterprise AI Tools

Publicly available AI tools can defeat attorney-client privilege. A New York federal court ruled in February 2026, per Spellbook's summary of the decision, that documents generated using publicly available AI tools are not protected by attorney-client privilege or the work-product doctrine¹¹. Enterprise tiers with no-training, no-retention contracts change that calculus.

Consumer ChatGPT and the enterprise tier are different products with different ethical surface areas. Pretending they're the same is a privilege risk.

Five differences make this a privilege question, not a UI question.

Public vs. Enterprise — What Actually Differs - Training data: Consumer tiers may use prompts to train future models. Enterprise tiers contractually do not. - Retention: Consumer prompts may be stored indefinitely. Enterprise contracts specify retention windows. - Data isolation: Enterprise tiers offer tenant isolation; consumer tiers do not. - Audit logs: Enterprise gives you a record; consumer gives you nothing defensible. - Who sees what: Vendor employees may review consumer prompts for safety. Enterprise locks that down.

Some federal judges, including N.D. Texas's Judge Brantley Starr, now require certifications that name the specific AI tool used in a filing — "AI" as a generic label is no longer enough¹⁰¹². Treat the NY ruling as directional pending appellate path; the privilege framework is moving fast.

Five tool-selection criteria that map directly to privilege defensibility:

• Data residency — where does the data physically sit, and under whose jurisdiction?
• Training opt-out — contractually enforceable, not a setting toggle
• Audit logs — exportable, time-stamped, attributable to a user
• Retention — short, defined, and aligned with your matter retention policy
• Access controls — who at the vendor can see prompts, and under what process

Tool, ethics, privilege — three threads. Here's how to braid them into a workflow you can actually run.

A 5-Step Disclosure Framework Firms Can Use This Week

A workable AI disclosure process has five steps: classify the matter's AI exposure, name the specific tool, draft matter-specific disclosure language, surface it in the kickoff conversation (not just the engagement letter), and document the client's response. This is the operational backbone of an AI implementation roadmap for any firm using the best AI for legal drafting at any scale.

1. Classify exposure. Is AI doing language refinement only, or substantive research, analysis, or drafting? The 2026 ABA practical guide draws the line⁶; your obligation tier follows it.
2. Name the tool. Specific product, tier, and data-handling posture. Some courts already require this in filings¹⁰— start practicing on every matter.
3. Draft matter-specific language. Not a global footer. A paragraph keyed to this matter's AI use, written so a non-lawyer can paraphrase it back to you.
4. Surface it live. Walk the client through the disclosure in the kickoff or status call. Confirm comprehension by asking them to restate what AI will and won't do on the matter⁹.
5. Document the response. Note questions asked and answered. Update the file when AI use changes mid-matter — the 2026 ABA guidance treats AI scope as ongoing, not one-and-done⁶.

The disclosure isn't done when it's written. It's done when the client has read it, asked questions, and answered them in their own words. Engagement letters memorialize consent. They don't manufacture it.

Tool and disclosure aren't two decisions. They're one — and most firms are running them out of sequence.

FAQ — What Lawyers Are Actually Asking

These are the four questions firm leaders ask most when picking a legal-drafting AI and writing the disclosure that comes with it.

Do I have to tell my clients I use AI?

Only when AI use influences significant decisions, affects billing, or substantively contributes to drafting or analysis⁵⁶. Pure language refinement (grammar polish, clarity edits) doesn't trigger disclosure under the 2026 ABA practical guide⁶. Anything beyond polish does.

Is a single AI clause in my engagement letter enough?

No. Op. 512 explicitly calls boilerplate consent inadequate when client confidences are involved⁷. Pair the engagement-letter clause with a matter-specific disclosure conversation — a footer doesn't satisfy a comprehension standard.

Is Harvey worth $1,000+ per lawyer per month?

For BigLaw enterprise teams running large transactional and diligence workflows, the ROI case exists. For mid-market and below, Spellbook or CoCounsel typically deliver more per dollar¹². Pricing is vendor-reported; verify before purchase.

Does using ChatGPT violate attorney-client privilege?

Public-tier ChatGPT can. A February 2026 NY federal ruling, per Spellbook's summary, held that documents generated using publicly available AI tools aren't protected by privilege or work product¹¹. Enterprise tiers with no-training contracts materially change the analysis — but only if the contract is actually in place.

One Decision, Not Two

Picking the best AI for legal drafting and writing the disclosure that goes with it is one decision, not two. The firms that treat it that way produce better work and better client trust. AI as intellectual augmentation— not artificial replacement— is what the standard actually rewards: tools that make the lawyer sharper, with disclosure that makes the client smarter about what's happening on their matter.

If running the tool decision and the disclosure decision in parallel feels like a stretch, an implementation partner can help. Dan Cumberland Labs offers fractional AI advisory, built for founder-led firms — peer-to-peer, with exactly these decisions in scope.

Editorial framing, not legal advice. Consult your state bar before adopting any disclosure language.

References

1. Spellbook, "CoCounsel vs Harvey: Pricing, Features & More (2026 Review)" (2026) — https://spellbook.com/briefs/cocounsel-vs-harvey
2. Spellbook, "9 Best Legal AI Tools Trusted by Lawyers (May 2026)" (2026) — https://spellbook.com/learn/legal-ai-tools
3. Harvey, "Introducing Harvey's Legal Agent Benchmark" (2026) — https://www.harvey.ai/blog/introducing-harveys-legal-agent-benchmark
4. American Bar Association, "ABA issues first ethics guidance on a lawyer's use of AI tools" (2024) — https://www.americanbar.org/news/abanews/aba-news-archives/2024/07/aba-issues-first-ethics-guidance-ai-tools/
5. ABA, "Formal Opinion 512: Generative Artificial Intelligence Tools" — disclosure trigger (2024) — https://www.lawnext.com/wp-content/uploads/2024/07/aba-formal-opinion-512.pdf
6. American Bar Association, "When Should Lawyers Disclose AI Use? A Practical Compliance Guide" (2026) — https://www.americanbar.org/groups/law_practice/resources/law-technology-today/2026/when-should-lawyers-disclose-ai-use/
7. ABA, "Formal Opinion 512: Generative Artificial Intelligence Tools" — boilerplate inadequacy (2024) — https://www.lawnext.com/wp-content/uploads/2024/07/aba-formal-opinion-512.pdf
8. National Conference of Bar Examiners, "Generative Artificial Intelligence Tools: ABA Formal Opinion 512" (2024) — https://thebarexaminer.ncbex.org/article/fall-2024/generative-artificial-intelligence-tools/
9. ABA, "Formal Opinion 512: Generative Artificial Intelligence Tools" — informed consent for self-learning tools (2024) — https://www.lawnext.com/wp-content/uploads/2024/07/aba-formal-opinion-512.pdf
10. Spellbook, "When Must Lawyers Disclose AI Use? Court Requirements for Legal Work" — named-tool requirement (2026) — https://spellbook.com/learn/ai-disclosure-requirements-legal-work-product
11. Spellbook, "When Must Lawyers Disclose AI Use? Court Requirements for Legal Work" — NY February 2026 privilege ruling (2026) — https://spellbook.com/learn/ai-disclosure-requirements-legal-work-product
12. Spellbook, "When Must Lawyers Disclose AI Use? Court Requirements for Legal Work" — Judge Brantley Starr certification (2026) — https://spellbook.com/learn/ai-disclosure-requirements-legal-work-product