Harvey vs Legora 2026: Legal AI After Privilege Ruling

A federal court has confirmed what compliance officers feared: typing into an AI chatbot is not the same as talking to your lawyer.

The Harvey vs Legora decision that law firms have been putting off just got a deadline attached to it. A court ruled that a client’s communications with an AI chatbot were not protected by attorney-client privilege or the work product doctrine — the prompts, the uploaded documents, and the model’s answers are all potentially discoverable. For any business owner whose counsel has started using AI tools, or who drafts contract summaries in a general-purpose chatbot, that ruling converts a technical question (“where does the prompt data live?”) into a billable-risk question with real dollars behind it. The two vendors dominating legal AI software 2026 shortlists — Harvey and Legora — answer that question very differently, and the difference is now the buying criterion that matters most.

What’s new in the Harvey vs Legora decision landscape

The privilege ruling is narrow on its facts and broad in its implications. Attorney-client privilege attaches to confidential communications between a client and their lawyer made to obtain legal advice. A chatbot is not a lawyer. When a client describes their legal problem to a commercial AI service, they describe it to a third party — and third-party disclosure is the classic way privilege gets waived. The work product doctrine, which shields materials prepared in anticipation of litigation, fared no better: the court found the materials were not prepared by or at the direction of counsel.

The second-order effect makes this urgent. Discovery requests already include AI tool logs, prompt histories, and vendor-side retention records. If your firm’s AI vendor retains prompts for 30 days for abuse monitoring, that retention window is a discoverable artifact sitting on someone else’s server, subject to someone else’s subpoena response policy. Does our tool retain prompts, for how long, and under whose control? That is no longer an IT procurement checkbox. It is a litigation exposure line item.

The two leading platforms attack the same problem from opposite directions. Harvey AI for law firms built its reputation on deep model work with elite firms and a strong enterprise security posture — SOC 2, zero data retention arrangements with model providers, and workflows structured around matters rather than free-form chat. Legora — the Swedish challenger formerly known as Leya — built its reputation on the collaborative surface: a tabular review interface where a lawyer runs one question across hundreds of documents and sees every answer linked back to its source text. Any honest Legora review concedes that the citation traceability is genuinely good. The privilege ruling rewards both approaches for different reasons: Harvey for retention discipline, Legora for the auditable paper trail showing a human lawyer reviewed the output.

Why it matters

• Your own chatbot history is now a discovery target. If you described a dispute to a consumer AI product before calling your lawyer, assume that conversation is discoverable. The privilege you would have had by picking up the phone never attached.
• Vendor retention policy is a legal term, not a technical one. “Zero data retention” needs to appear in your contract with a definition, a scope (does it cover the model provider’s infrastructure, or only the vendor’s?), and a remedy. A marketing page claim is worth nothing in a discovery fight over attorney-client privilege AI questions.
• Matter scoping stops being a convenience feature. A tool that keeps every matter in a separate workspace with separate access controls limits the blast radius when one matter gets subpoenaed. A tool with one shared corpus does not.
• Citation traceability is your defensibility record. When a court asks whether a human attorney exercised independent judgment, an output linking each claim to a specific document page is evidence. An unsourced summary is a liability.
• Cost math changes. Firms justify AI document review tools on hours saved. Add the cost of one privilege fight — motion practice, in camera review, potential waiver of an entire subject matter — and the cheaper tool with worse controls stops being cheaper.
• Your outside counsel’s stack is now your problem. Ask your firm what they run, what it retains, and whether your matters share a workspace with other clients. If they cannot answer in one email, that is the answer.

How to use it today: a Harvey vs Legora evaluation you can run this week

1. Inventory what already exists. Before evaluating vendors, find out what your people have already typed into consumer AI tools. Most enterprise accounts expose an export. Pull it, and treat anything legal-adjacent as unprivileged until counsel says otherwise.

2. Send both vendors the same written questionnaire. Do not accept a demo answer. Put this in an email and require written responses you can attach to the contract:

   RETENTION AND PRIVILEGE QUESTIONNAIRE — [VENDOR]
   
   1. Prompt/response retention period, in days, by default: ____
   2. Can retention be set to zero at the tenant level? Y/N
   3. Does zero retention extend to your upstream model provider(s)? Y/N
      - Name each provider and the contractual instrument: ____
   4. Are uploaded documents used for training? Y/N. Ever? Y/N
   5. Data residency options (list regions): ____
   6. Are matters isolated at the storage layer or only by UI permissions?
   7. Your policy on responding to third-party subpoenas for tenant data:
      - Do you notify the customer before producing? Y/N
      - Notice period: ____ days
   8. Can we export a complete audit log of prompts, users, and matters? Y/N
      - Format: ____ Retention of the log itself: ____ days
   9. Sub-processor list and change-notification period: ____
   10. SOC 2 Type II date and ISO 27001 status: ____
   

3. Run an identical bake-off task. Same document set, same question, both tools, same reviewer. Use a real but non-sensitive matter — an expired NDA set, closed transaction docs. This prompt surfaces the difference in citation quality:

   You are assisting a licensed attorney. Review the attached
   agreements and produce a table with one row per document.
   
   Columns:
   - Document name
   - Governing law clause (quote verbatim)
   - Assignment restriction: yes / no / silent
   - Change-of-control trigger: yes / no / silent
   - Source citation: document name + page + section number
   
   Rules:
   - Quote the operative language verbatim. Do not paraphrase.
   - If a provision is absent, write "SILENT" — do not infer.
   - If the language is ambiguous, write "FLAG FOR REVIEW" and
     explain the ambiguity in one sentence.
   - Every factual claim must carry a citation. No citation, no claim.
   

   Score on three axes: how many rows are correct, how many citations resolve to the right page on click, and how many hallucinated provisions appear. The last number should be zero. If it is not, that vendor is out.

4. Verify retention behavior rather than trusting the toggle. If the vendor exposes an API, set retention to zero, run a query, then attempt to retrieve the conversation. It should be gone.

   # Set tenant retention to zero, then verify a prompt is unrecoverable
   curl -X PATCH https://api.vendor.example/v1/tenant/settings \
     -H "Authorization: Bearer $VENDOR_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{"prompt_retention_days": 0, "training_opt_out": true}'
   
   # Run a query, capture the id, then try to fetch it back
   CONV_ID=$(curl -s -X POST https://api.vendor.example/v1/chat \
     -H "Authorization: Bearer $VENDOR_API_KEY" \
     -d '{"matter_id":"TEST-001","message":"retention canary"}' \
     | jq -r '.conversation_id')
   
   curl -s https://api.vendor.example/v1/chat/$CONV_ID \
     -H "Authorization: Bearer $VENDOR_API_KEY" | jq '.messages'
   # Expect: 404 or empty. Anything else contradicts the sales deck.
   

5. Write the policy before you write the check. A one-page internal rule, distributed firm-wide, is what you point to later when someone asks whether the organization exercised reasonable care.

   AI USE POLICY — LEGAL MATTERS (v1)
   
   APPROVED TOOLS: [Harvey | Legora] tenant only.
   PROHIBITED: consumer chatbots for any matter-related content,
   including hypotheticals with facts drawn from a live matter.
   
   BEFORE ANY MATTER CONTENT IS ENTERED:
     [ ] Matter workspace exists and is scoped to this matter
     [ ] Retention set to zero; training opt-out confirmed
     [ ] Engagement letter includes AI-use disclosure
   
   EVERY OUTPUT:
     [ ] Attorney reviewed every citation against source
     [ ] Review recorded in the matter file with reviewer name + date
     [ ] Output labeled: DRAFT — ATTORNEY WORK PRODUCT,
         PREPARED AT DIRECTION OF COUNSEL
   
   NEVER: paste privileged communications into any tool not listed above.
   

   That labeling step is not decoration. Work product protection depends on materials being prepared at the direction of counsel in anticipation of litigation, and documenting that at creation time is far more persuasive than asserting it two years later.

6. Negotiate the subpoena-notice clause. The single most valuable contract term you can add: the vendor must notify you before producing tenant data to any third party, with enough lead time to move to quash. Vendors agree to this during procurement and rarely afterward.

How Harvey vs Legora compares

Criterion	Harvey	Legora	General-purpose chatbot
Core strength	Deep workflows built with large firms; enterprise posture	Collaborative tabular review; strong source linking	Speed and familiarity, nothing else
Matter scoping	Matter-centric by design	Workspace and project scoping	Effectively none
Citation traceability	Citations to source documents	Cell-level links back to source passages	Frequently fabricated
Bulk document review	Strong, workflow-driven	Strong, spreadsheet-style across large sets	Limited by context window
Retention controls	Enterprise agreements including no-training terms	Enterprise agreements; EU data residency emphasis	Consumer terms; often retained
Privilege risk posture	Lowest, with contract terms in place	Lowest, with contract terms in place	Highest — assume discoverable
Best fit	Firms standardizing across practice groups	Teams whose bottleneck is document review volume	Non-legal drafting only

The honest read: for most buyers this is no longer a capability contest. Both platforms handle contract analysis competently and both cite their sources. The decision turns on which one will sign the retention and notice terms your risk profile requires, and which interface your people will actually adopt. Run the bake-off with your own documents; vendor benchmarks are marketing.

What’s next

Expect the discovery playbook to formalize quickly. Within a year, standard document requests will name AI tool logs, and courts will start distinguishing between AI use directed by counsel — which has a real work product argument — and AI use by a client acting alone, which does not. The firms that come out ahead will produce a clean audit trail showing which attorney reviewed which output on which date. That is a records-management problem more than a technology problem, and most organizations are not ready for it.

On the vendor side, watch legal AI data retention become a headline feature rather than a footnote in the security whitepaper. Likely developments: verifiable zero-retention with cryptographic attestation, customer-managed encryption keys as a standard tier rather than an enterprise upsell, and privilege-aware audit logs that record attorney review as a first-class event. Watch the regional split too — Legora’s European posture wins anywhere data residency is a hard requirement, while Harvey’s depth with large US firms keeps it the default in that market. Pricing pressure between the two is real and getting better for buyers.

The wider thing to watch is bar guidance. State bars have been issuing competence and confidentiality opinions on AI use, and this privilege ruling will accelerate the specific ones — rules about disclosure to clients, about supervision of AI output, and about whether AI use requires informed consent in the engagement letter. Get ahead of it: add an AI-use paragraph to your engagement letters now. It costs nothing and it will look like foresight rather than scrambling.

Frequently Asked Questions

Does this ruling mean my lawyer can’t use AI?

No. It means AI use needs to sit inside the attorney-client relationship rather than outside it. Work performed by a tool at the direction of counsel, within a matter-scoped workspace, reviewed by an attorney, and documented as such has a strong protection argument. A client typing facts into a consumer chatbot before ever contacting a lawyer does not.

Which is better for a small firm, Harvey or Legora?

For teams whose main pain is reviewing large document volumes, Legora’s tabular interface tends to produce faster adoption because it looks like a spreadsheet and behaves like one. For firms standardizing workflows across multiple practice groups, Harvey’s matter-centric structure fits better. Ask both for pricing at your actual seat count — the published enterprise positioning is not the only tier available.

Is “zero data retention” enough to preserve privilege?

It helps considerably but is not sufficient on its own. Privilege depends on the communication being between client and counsel for legal advice; retention policy affects what a third party can be compelled to produce. You want both: zero retention to limit the discoverable footprint, and a workflow structured so the AI work is genuinely attorney-directed.

What should I ask my outside counsel right now?

Four questions, in one email: which AI tools do you use on our matters, is our matter isolated from other clients’ data, what is the retention period on prompts and uploads, and will you notify us if a third party requests that data. A firm with a mature AI program answers all four immediately.

Do I need to disclose AI use to my clients?

Increasingly, yes — bar guidance trends clearly toward disclosure. Even where it is not yet required, an AI-use paragraph in the engagement letter is cheap insurance. It sets expectations, documents consent, and demonstrates the kind of care that matters if the question ever comes up.

What about documents I already put into a chatbot?

Talk to counsel before doing anything else, and do not delete anything — if litigation is reasonably anticipated, deletion creates a spoliation problem far worse than the original disclosure. Inventory what was entered, when, and by whom, then let your lawyer assess whether privilege was waived and whether the waiver extends to related subject matter.