Harvey AI and the EU AI Act: A Deployer's Compliance Guide

AI Tool Compliance17 July 2026· 11 min read

Harvey AI is a legal assistant for research, drafting and review — generally minimal risk under the EU AI Act. Real duties: Article 4 literacy, privilege.

Probably not high-risk. Harvey is a legal-AI assistant — built on general-purpose AI (GPAI) models for legal research, drafting, document review and matter Q&A, used by law firms and in-house legal teams. The EU AI Act classifies systems by intended purpose, not product name, and ordinary legal research, drafting and review is in none of the high-risk categories. The realistic compliance work is narrow: Article 4 AI literacy, confidentiality and data governance over privileged client matter, and Article 50 transparency only where output reaches a court or third party. This is the deployer's-eye view of Harvey itself — not a survey of the legal-tech sector, nor the separate case of AI assisting a court.

Harvey runs on GPAI models supplied by a model provider, and the Chapter V obligations on those models (Articles 51–55, in force since 2 August 2025) sit with that provider — not the deploying firm.

Is Harvey AI high-risk under the EU AI Act? Generally not

Under Article 6 and Annex III of Regulation (EU) 2024/1689, a system is high-risk only where its intended purpose falls within a listed category — biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, or the administration of justice. There is no line for "helping a lawyer research the law." Ordinary legal research, drafting and review is in none of them.

Set the expectation honestly. For the typical deployer, Harvey is minimal-to-limited risk: no conformity assessment, no Annex IV technical documentation, no EU-database registration. The duties that apply are Article 4 literacy, data governance over the client matter you feed in, and Article 50 transparency where output goes outside the firm.

One forward-pointer: AI assisting a judicial authority in administering justice is Annex III point 8 — a different actor, covered in the judicial-decision-AI guide. A firm using Harvey to prepare its own work is not doing that.

Where Harvey sits: a risk-tier map by use pattern

The table

The tier follows the workflow, not the tool. The same Harvey licence can sit at minimal risk in one use and pull high-risk duties in another, depending on what the output feeds into.

Use pattern	Likely tier	Governing provision
Legal research / case-law and statute lookup for a fee-earner	Minimal risk	No Annex III category; Article 4 applies
Drafting and document-review aids for internal lawyer use	Minimal risk	Advisory preparatory work; Article 6(3) filter logic
Generative output shown to a client or counterparty (AI-drafted letter, memo)	Limited risk for that path	Article 50 transparency
Client-facing intake / Q&A chatbot built on Harvey	Limited risk	Article 50(1) AI-disclosure
Harvey output wired into a firm HR decision (recruiting, performance)	High-risk workflow	Annex III point 4 (the workflow, not Harvey)
System sold to / deployed by a court to assist judicial decision-making	High-risk, different actor	Annex III point 8 (out of scope; see sibling)

Reading the table: the workflow is classified, not the tool

Stand-alone high-risk obligations under Article 6(2) attach to a workflow engineered on top, not to the assistant. Feed Harvey output into an Annex III decision and you have built something the Act treats as high-risk; use it to draft a clause faster and you have not. The Article 6(3) filter reinforces this: review aids that do preparatory work a lawyer checks and acts on stay outside high-risk, because the lawyer remains the decision-maker.

The real exposure: confidentiality, privilege and data governance over client matter

Why this is the dominant risk for Harvey, not the high-risk stack

For Harvey, the binding compliance work is confidentiality and data governance over the privileged client matter you put into the tool, not the high-risk stack.

Upload client documents, correspondence or privileged advice into a third-party legal-AI assistant and the firm is the GDPR controller, not the processor. Processing basis, minimisation and purpose limitation must be settled before upload, under GDPR Article 5(1)(c) and Article 5(1)(f) — a responsibility you cannot delegate to the vendor.

Professional privilege and bar-association confidentiality duties sit outside the AI Act, but they are the most immediate driver: privilege can be jeopardised if material reaches a vendor without adequate safeguards. This is a GDPR Article 32 and GDPR Article 28 workstream, not an AI Act conformity step — but non-negotiable before you index client matter.

Vendor due diligence before client data goes in

Before privileged material reaches the tool, settle:

A documented no-training-on-customer-data commitment from the vendor.
A GDPR Article 28 Data Processing Agreement governing the processing.
Matter-level access segregation so one client's data is not retrievable across unrelated matters.
EU data residency, or an EU-adequate transfer where processing occurs outside the EEA.
Retention and deletion controls you can exercise.

What you actually owe as a Harvey deployer

Article 4 AI literacy and the hallucination duty

Article 4 AI literacy has been in force since 2 February 2025. Fee-earners and support staff using Harvey must understand that outputs are synthesised, can be incomplete or wrong, and must be verified against the underlying authority before reliance — a duty of competent use, not a certification programme.

The legal-specific failure mode is AI-fabricated case citations. Independent verification of every citation before filing is precisely the competence the duty codifies, and it mirrors the professional-responsibility rules a fee-earner is already bound by; a short module on "verify every citation against the source" discharges the substance of Article 4.

Inventory and record the classification (Article 26)

Enter Harvey in your AI register with the vendor, intended use, the minimal/limited-risk determination and the date. Article 26 sets baseline deployer diligence — use the system per the provider's instructions and keep human oversight — and recording the classification is the audit trail, even for a low-risk tool.

Article 50 transparency — only where output reaches a court or third party

Article 50 transparency applies only where AI output reaches third parties, or could be mistaken for human-authored or interactive content. A Harvey-built client intake chatbot must disclose that it is an AI system under Article 50(1). Content presented externally may need labelling — content-marking duties apply from 2 December 2026 under the Digital Omnibus, adopted by the European Parliament and the Council in June 2026. Output reaching a court or counterparty is the trigger; internal drafting is not.

The penalty anchor: Article 50 breaches sit in the €15 million or 3% of total worldwide annual turnover tier under Article 99(4), whichever is higher; Article 99(6) caps fines on SMEs and start-ups at whichever of the fixed sum or the percentage is lower.

The edge cases that change the classification — and Article 25

Building a decision workflow on top of Harvey

Harvey only approaches high-risk if a deployer wires its output into an Annex III decision — for example the firm's own recruitment or in-employment decisions under Annex III point 4 (4(a) recruitment, 4(b) in-employment). Even then it is the HR workflow that is classified, not the assistant: drafting a job description is not high-risk; scoring candidates is.

When a firm or legal-tech vendor becomes a provider

Article 25 governs the role shift. A deployer becomes a provider — with full provider obligations — if it puts its own name or trademark on a high-risk system, makes a substantial modification, or repurposes a system so that it becomes high-risk. Pointing Harvey's output at an Annex III decision is exactly that repurposing; ordinary research and drafting is not.

A legal-tech vendor building a tool to assist a judicial authority in administering justice is building an Annex III point 8 high-risk system — a separate actor and analysis, covered in the judicial-decision-AI guide, not what a firm deploying Harvey for its own work is doing.

The worked example

Lindqvist & Ríos LLP, a 280-lawyer EU commercial firm, rolls out Harvey for research, drafting and contract review. Classification lands at minimal risk for the internal work and limited risk for one client-facing intake chatbot. The binding work is short: a GDPR Article 28 DPA with a no-training commitment and EU data residency; matter-level access segregation; a 30-minute Article 4 literacy module on citation verification; an AI-register entry with the dated finding; an Article 50 disclosure banner on the intake bot. No high-risk stack and no FRIA.

One timeline note for any high-risk workflow built on top. Stand-alone Annex III obligations (Article 6(2)) now apply from 2 December 2027, deferred from the original 2 August 2026 by the Digital Omnibus, adopted by the European Parliament (16 June 2026) and the Council (29 June 2026), entering into force on Official Journal publication expected before 2 August 2026. "Stop-the-clock" was rejected.

How Confir helps

Add Harvey to your inventory and Confir's classification engine asks how it is used — its role, the data it touches, who sees the output — then maps research, drafting and review use to the correct tier. For the normal case it returns minimal or limited risk with the short, real obligation set (Article 4 literacy, a register entry, the confidentiality workstream) plus a dated, auditable record. Describe wiring output into an Annex III decision and the engine flags the Article 25 role shift instead.

The engine is deterministic and rule-based — no model inference, no hallucination. The same intake produces the same finding, and the rule that fired is written in plain language for any reviewer to check. GPAI handling is the upstream model provider's responsibility, not a Confir offering.

Frequently asked questions

Is Harvey AI high-risk under the EU AI Act?

Generally not. The Act classifies systems by intended purpose, not product name (Article 6, Annex III), and ordinary legal research, drafting and document review is not a listed Annex III category. For a typical law firm or in-house team, Harvey is minimal-to-limited risk — no conformity assessment, no Annex IV documentation, no EU-database registration. The realistic duties are Article 4 literacy, data governance over client matter, and Article 50 transparency where output reaches a court or third party.

What is the real EU AI Act risk of putting client matter into Harvey?

The dominant practical risk is confidentiality and data governance, not the high-risk stack. When you upload client documents or privileged advice you are the GDPR controller, so processing basis, minimisation and purpose limitation must be settled first (GDPR Article 5(1)(c) and 5(1)(f)). Professional privilege and bar confidentiality duties sit outside the AI Act but are the most immediate driver. Before indexing client data, secure a no-training commitment, a GDPR Article 28 DPA, matter-level access segregation and EU data residency.

Do I need to do anything under the EU AI Act before deploying Harvey?

Yes, but the list is short. Provide Article 4 AI literacy so fee-earners understand Harvey's outputs are synthesised and must be verified — including independent checking of every case citation (in force since 2 February 2025). Record Harvey in your AI register with its intended use and minimal/limited-risk classification as baseline Article 26 deployer diligence. Settle the confidentiality and data-processing terms before client matter goes in.

Does Article 50 transparency apply to Harvey?

Only where output reaches third parties or could be mistaken for human-authored or interactive content. A client-facing intake or Q&A chatbot built on Harvey must disclose it is AI under Article 50(1). AI-generated content presented externally — to a court or counterparty — may need labelling, with content-marking duties applying from 2 December 2026 under the Digital Omnibus, adopted by the European Parliament and the Council in June 2026. Purely internal drafting for your own lawyers does not trigger Article 50, and breach sits in the €15M/3% tier under Article 99(4).

Who handles the GPAI obligations for the models behind Harvey?

The general-purpose AI model provider, not your firm. Chapter V obligations (Articles 51–55) — technical documentation, downstream information, copyright policy and training-data summary — sit with the model provider and have applied since 2 August 2025. As a deployer of the Harvey application you are an Article 26 deployer, not a GPAI provider, so you benefit from upstream compliance without inheriting those provider duties.

Could using Harvey ever make my firm a provider of a high-risk system?

Only if you build a high-risk workflow on top of it. Under Article 25 a deployer becomes a provider if it puts its name on a high-risk system, makes a substantial modification, or repurposes a system so it becomes high-risk. Ordinary research and drafting does not do this — but wiring Harvey's output into your own recruitment or in-employment decisions (Annex III point 4) could make that HR workflow high-risk and pull provider obligations onto you. AI used by a judicial authority (Annex III point 8) is a separate actor and analysis.

When do high-risk deadlines apply, and do they affect Harvey?

For normal Harvey use they do not — there is no high-risk obligation to meet. If you engineered a stand-alone Annex III workflow on top, obligations under Article 6(2) now apply from 2 December 2027, deferred from the original 2 August 2026 by the Digital Omnibus, adopted by the European Parliament and the Council in June 2026 and entering into force on Official Journal publication expected before 2 August 2026. The deferral is fixed calendar dates, not a stop-the-clock.

Manage your EU AI Act compliance in one place

Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.

Start free trial →

Confir EU AI Act Platform