AI Privacy Tracker · Version 110426

Harvey

Plain-English resources for lawyers, judges, and government attorneys who use AI, or want to.

Harvey · Harvey

Harvey at a glance

Rated Green

Fit for client confidences after your usual vendor checks.

Rated Yellow

Usable for client confidences only after you deal with the gap named in the entry.

When you delete a chat

Harvey says users can delete their data in the app at any time, and that uploaded files are securely destroyed when the firm's retention period ends. Its public pages do not mention backups.

PlanRatingTrains on your data?How long keptCan the provider read it?
Harvey (core platform)GreenNo (contract)Firm sets; as short as 3 hoursStaff only for support or legal need
Harvey Covered AI ServicesYellowNo (contract)Model company: generally 30 days, sometimes longerModel company may review for safety

Each plan has one color, set by its contract. The settings in each entry lower the risk but do not change the color. Select a plan for its full entry.

Deleting a chat

What deleting does on Harvey

Harvey workspace data: what happens after a chat is deleted
Harvey says users can delete their data in the app at any time. Harvey's public pages do not say how fast deleted data leaves its systems. Harvey says uploaded source files are securely destroyed when the firm's retention period ends. Provider page
Copies kept anyway
  • Audit logs are kept for at least 1 year and up to 10 years.
  • Under the core terms, subprocessors other than cloud storage providers may not keep your content. Harvey says its model partners delete data once a request is complete.
  • Data sent through Covered AI Services may be kept by the model company. See that card.
  • Backups are not mentioned.

In short: deleting removes a chat from view, but it does not undo what has already happened to it. On business plans, the organization's retention settings decide how long chats are kept; each plan below explains how.

Plan by plan

Every Harvey plan in detail

Open a plan for its rating, the settings that lower the risk, and short quotations from Harvey's own policy with links.

Harvey (core platform)GreenRated Green. Harvey's Platform Agreement forbids Harvey to train AI models on your firm's content. It also forbids Harvey's subprocessors to train on it. Subprocessors other than cloud storage providers may not keep the content or log it for human review. A Data Processing Addendum makes Harvey your processor. The firm sets how long workspace data is kept, with options as short as three hours. Harvey staff may access data only to provide or support the service, or to comply with the law. These protections do not cover Covered AI Services, which have their own card.

Harvey platform (Assistant, Vault, Knowledge, Agents, workflows), enterprise. Your firm or legal department signs an order form with Harvey. Harvey's website offers a demo request, not a self-service sign-up.

Default settings Green

The Platform Agreement forbids Harvey to train AI models on your prompts, responses, and uploaded documents. It forbids Harvey's subprocessors to train on them too. Subprocessors other than cloud storage providers may not keep your content or log it for human review. Harvey acts as your processor under a Data Processing Addendum. The firm sets its own retention period. Harvey says options run as short as three hours.

Settings that lower the risk

  1. Workspace retention settingWhat it does: Sets how long workspace data is kept. Harvey says options run as short as three hours, counted from upload or from last use. The Service Terms say the setting does not apply to certain features listed in Harvey's documentation.
  2. Covered AI ServicesWhat it does: An optional feature that customer administrators configure. When it is used, the model company may keep data. Its people may review data for safety. Data may be processed outside your region. It has its own card.
  3. Web Browsing and Knowledge Source featuresWhat it does: Optional features that administrators configure. The Service Terms say the subprocessors for these features may process data outside your selected region.
  4. Ask LexisNexisWhat it does: Brings answers from Lexis+ AI into Harvey. The Service Terms say LexisNexis may process data outside your selected region. Harvey's subprocessor list gives its location as the US.
  5. Email HarveyWhat it does: Lets users send questions to Harvey by email. The email subprocessor does not keep emails or allow human review of them. The exceptions are legal requirements and enforcement of its acceptable use policy. From November 15, 2026, that review is tied to automated scans that detect suspicious activity.
  6. External Collaboration (Shared Spaces)What it does: Lets authorized users invite people outside the firm into a shared space. The Service Terms say shared material stays under the firm's own storage, processing, and retention settings.
  7. Bring Your Own Key (BYOK)What it does: Harvey says customers who use BYOK manage the key that encrypts their stored data. Harvey says revoking the key makes the data inaccessible, including to Harvey. This is on Harvey's blog, not in the published terms.
  8. MemoryWhat it does: Harvey says administrators control whether Memory is available, for all users or a subset. Harvey says each user can review, edit, search, or delete memories, or turn Memory off.

Worth knowing

  • Covered AI Services lift the core promises on retention, human review, and data location for the traffic they handle. They have their own card.
  • The Service Terms say the workspace retention setting does not apply to certain features. Harvey lists those features in its customer help center, which requires a login.
  • From November 15, 2026, new Service Terms let Harvey employees or contractors, called Harvey Specialists, help with setup and use. The firm decides what information, if any, they receive.
  • From November 15, 2026, Harvey's subprocessor list adds Applied Compute, an AI provider, and Confluent, a data infrastructure provider.
  • Harvey may update its Service Terms and policies by posting them. The Platform Agreement says an update cannot weaken Harvey's obligations on confidential information, customer data, or security without the firm's written authorization.

What the policy says

On no training by Harvey (binding)
“Harvey will not train any AI models using Your Content or Customer Data.”

Harvey Platform Agreement, §11.8 No Training, Last updated January 9, 2026. Read the source

On no training, retention, or human-review logging by subprocessors (binding)
“Subprocessors, except for cloud storage providers, will not retain or log for human review Your Content or Customer Data.”

Harvey Platform Agreement, §11.8 No Training, Last updated January 9, 2026. Read the source

On processor role
“You are the Data Controller, and Harvey is the Data Processor”

Harvey Data Processing Addendum, §1, Data Processing Roles, Last updated August 2, 2026. Read the source

On confidentiality
“Content specific to You and Your Customer Data are Your Confidential Information.”

Harvey Platform Agreement, §12, Confidential Information, Last updated January 9, 2026. Read the source

On customer-defined retention; in-app deletion
“Customers define their own data retention periods, with options as short as three hours”

How Harvey's Building a Culture of Privacy (Harvey blog), Body, August 19, 2025. Read the source

On staff access limits (binding)
“Harvey personnel will not access Customer Data or Content except (i) to provide or support the Service”

Harvey Security Addendum, §4.2, Last updated December 26, 2025. Read the source

On Harvey Specialists (from November 15, 2026)
“You determine the information (if any) needed for Harvey Specialists to perform their services”

Harvey Service Terms Preview, §8 Harvey Specialists, Last updated October 16, 2026; effective November 15, 2026. Read the source

On Memory controls and no training
“individual memories are never used to train AI models”

Memory is Here: Harvey, Personalized (Harvey blog), You Control What Harvey Remembers, August 26, 2026. Read the source

Your data, in plain English

How long it is kept
  • The firm sets a workspace retention period.
  • Harvey says options run as short as three hours, counted from upload or from last use.
  • Harvey says uploaded source files are securely destroyed when the period ends.
  • The Service Terms say the setting does not apply to certain features.
  • Harvey lists those features only in its customer help center, which requires a login.
  • Within 30 days after the contract ends, Harvey must securely delete any remaining customer data and content, unless the firm instructs otherwise.
  • Audit logs are kept for at least 1 year and up to 10 years.
What happens after you delete
Harvey says users can delete their data in the app at any time. Harvey's public pages do not say how fast deleted data leaves its systems. Harvey says uploaded source files are securely destroyed when the firm's retention period ends.
Copies kept anyway
  • Audit logs are kept for at least 1 year and up to 10 years.
  • Under the core terms, subprocessors other than cloud storage providers may not keep your content. Harvey says its model partners delete data once a request is complete.
  • Data sent through Covered AI Services may be kept by the model company. See that card.
  • Backups are not mentioned.
Temporary chat
Not offered as a named mode. The firm can set workspace retention as short as three hours.
Memory
  • Harvey announced Memory on August 26, 2026, in early access, as part of Harvey II.
  • Memory saves a user's preferences when the user asks, or when a prompt signals a preference.
  • Harvey shows a confirmation card for each saved memory.
  • Harvey says memories are scoped to the individual user.
  • Users can review, edit, search, or delete memories, or turn Memory off.
  • Administrators control whether Memory is available.
  • Harvey says memories are never used to train AI models.
  • Harvey does not say how long memories are kept.
Who at the provider can read it
  • The Security Addendum lets Harvey personnel access customer data and content only to provide or support the service, or to comply with the law or a binding government order.
  • Harvey personnel must sign confidentiality agreements.
  • Under the core terms, subprocessors other than cloud storage providers may not log your content for human review.
  • From November 15, 2026, Harvey Specialists, who are Harvey employees or contractors, may help the firm with configuration, implementation, or use.
  • The firm decides what information, if any, they receive.
  • Harvey treats that information as confidential.
  • Harvey limits access to Specialists who need it.
Who else receives it
  • Harvey's subprocessor list (September 21, 2026) names these AI providers: Microsoft, OpenAI, Google Cloud, Amazon Web Services, Anthropic, ElevenLabs, DeepL, Mistral AI, Baseten, and Fireworks.ai.
  • The contract forbids every subprocessor to train on your data.
  • From November 15, 2026, the list adds Applied Compute and Confluent.
  • Web browsing uses you.com and Parallel Web Systems.
  • Ask LexisNexis sends data to LexisNexis in the US.
  • Firms that subscribe to notices get 30 days' warning of a new subprocessor.
  • They may object within 15 days.
  • No advertising use is stated.
  • The DPA requires Harvey to tell the firm about binding law-enforcement requests, where the law allows.
  • Harvey's law-enforcement policy says it sends requesters to the customer.
  • The policy says Harvey discloses only the minimum.
Contract
  • Harvey Platform Agreement (last updated January 9, 2026), Service Terms (September 16, 2026; a new version takes effect November 15, 2026), Security Addendum (December 26, 2025), and Data Processing Addendum (August 2, 2026), plus the firm's order form.
  • The no-training promise binds Harvey.
  • It also binds Harvey's subprocessors.
  • The DPA makes Harvey the processor for customer personal data.
  • Harvey is an independent controller for account and usage data.
  • Your prompts, responses, and uploaded documents are your confidential information under the contract.
  • Use of protected health information requires a signed Harvey Business Associate Addendum.
All sources for this plan (17)

Last checked against Harvey's pages on October 25, 2026.

Not yet confirmed
  • Which features fall outside the workspace retention setting, and how long each is kept. Harvey lists them in its customer help center, which requires a login.
  • How Vault retention works and who sets it. Harvey describes a separate Vault retention rule only in its customer help center.
  • The longest workspace retention period Harvey allows.
  • How long Harvey keeps Memory entries, and whether deleting a conversation removes memories made from it.
  • Whether deleted data is removed from backups, and when.
Harvey Covered AI ServicesYellowRated Yellow. The Platform Agreement forbids Harvey's subprocessors, including the model company, to train on your data. Service Terms §8 does not change that. Harvey remains your processor. The gaps come from Service Terms §8. The model company may keep data. Harvey said in June 2026 that Anthropic generally keeps Fable 5 data 30 days. The model company's people may review data for safety and to enforce its use policy. Data may be processed outside your region. Covered AI Services are off until an administrator turns them on.

Harvey Covered AI Services (formerly Extended AI Features), currently Anthropic Fable 5 and Fable 5.1. Firms whose Harvey administrators have turned on Covered AI Services. Harvey uses this label for AI features whose data handling differs from Harvey's usual commitments. Harvey's announcements name Anthropic's Fable 5 (June 9, 2026) and Fable 5.1 (September 1, 2026).

Default settings Yellow

Covered AI Services are optional. An administrator must turn them on. When they are used, the model company may keep your data. Harvey said on June 9, 2026 that Anthropic generally keeps Fable 5 data for 30 days, or longer when legal or safety reasons require. Anthropic's people may review data for safety. Fable models run only in the United States. The contract still forbids training on your data.

Settings that lower the risk

  1. Covered AI ServicesWhat it does: An optional feature that customer administrators configure. Harvey offers Fable models as an opt-in, not by default.

Worth knowing

  • Harvey said on September 1, 2026 that eligible customers can use Fable 5 and Fable 5.1 with zero data retention under a time-bound exemption, until Anthropic's Enterprise Frontier Safeguards are ready. This is Harvey's statement, not a contract term. Harvey does not say which customers are eligible.
  • Harvey says Anthropic does not offer regional processing for Fable models. Data is processed in the United States even if the firm chose another region.
  • Harvey's June 9, 2026 announcement says Anthropic accesses customer data only through a scoped viewer with no export, copy, or download capability.
  • From November 15, 2026, the Covered AI Services terms move to Service Terms §9. The wording does not change.

What the policy says

On model company may keep data (binding override)
“Subprocessors may retain Customer Data and Content to the extent necessary to provide their services, perform safety reviews, or comply with applicable law”

Harvey Service Terms, §8.2, Last updated September 16, 2026. Read the source

On model company may review data (binding)
“Customer Data and Content may be reviewed by the Subprocessors to perform safety reviews”

Harvey Service Terms, §8.2, Last updated September 16, 2026. Read the source

On processing outside your region (binding)
“may be stored and processed in a location different to the storage and processing locations stipulated in Your Terms”

Harvey Service Terms, §8.2, Last updated September 16, 2026. Read the source

On no training by subprocessors, not overridden (binding)
“Subprocessors will not train any AI models using Your Content or Customer Data.”

Harvey Platform Agreement, §11.8 No Training, Last updated January 9, 2026. Read the source

On 30-day retention by Anthropic; scoped viewer
“Anthropic has informed us that this data will generally be deleted after 30 days”

Fable 5, Now Available in Harvey (Harvey blog), What to Know Before Enabling Fable 5, June 9, 2026. Read the source

On Fable 5.1 added; zero-retention exemption (Harvey statement); US processing
“Eligible Harvey customers can use Fable 5 and Fable 5.1 with zero data retention under a time-bound exemption”

Fable 5.1, Now Live in Harvey (Harvey blog), What to Know Before Enabling Fable 5.1, September 1, 2026. Read the source

Your data, in plain English

How long it is kept
  • Service Terms §8 lets the model company keep data to provide its service, perform safety reviews, or comply with law.
  • Harvey's June 9, 2026 announcement says Anthropic told Harvey that Fable 5 data is generally deleted after 30 days.
  • Anthropic keeps it longer when legal or safety reasons require.
  • Harvey said on September 1, 2026 that eligible customers can use Fable 5 and Fable 5.1 with zero data retention under a time-bound exemption.
  • That statement is not a contract term.
  • Service Terms §8 does not change how long Harvey itself keeps data.
What happens after you delete
Harvey's public pages do not say whether deleting data in Harvey removes the copy the model company keeps.
Copies kept anyway
  • The model company may keep data for safety reviews or as the law requires.
  • Harvey said on June 9, 2026 that Anthropic keeps Fable 5 data longer than 30 days when legal or safety reasons require.
  • Harvey has not published a retention period for Fable 5.1.
Temporary chat
Not offered.
Memory
Harvey's Memory feature applies across the platform. See the core Harvey card. Harvey does not say whether memories are sent to Covered AI models.
Who at the provider can read it
  • Service Terms §8 lets the model company's people review data to perform safety reviews.
  • They may also review it to enforce the model company's acceptable use policies or to comply with law.
  • Harvey's June 9, 2026 announcement says Anthropic accesses customer data only through a scoped viewer with no export, copy, or download capability.
  • Harvey's own staff access follows the core terms.
Who else receives it
Anthropic receives the data. Harvey says Anthropic does not offer regional processing for Fable 5 or Fable 5.1, so all data is processed in the United States. Service Terms §8 allows storage and processing outside the locations in the firm's terms. No advertising use is stated.
Contract
  • Service Terms §8 (September 16, 2026) applies notwithstanding anything to the contrary in Harvey's terms.
  • It overrides Harvey's usual promises on retention, human review, and data location for this traffic.
  • It does not override the Platform Agreement's no-training clause.
  • That clause binds all subprocessors, including Anthropic.
  • Covered AI Services are not part of Harvey's Basic Service.
  • From November 15, 2026, the same text becomes Service Terms §9.
All sources for this plan (9)

Last checked against Harvey's pages on October 25, 2026.

Not yet confirmed
  • Which Harvey customers qualify for the zero-retention exemption on Fable 5 and Fable 5.1, and when the exemption ends.
  • How long Anthropic keeps Fable 5.1 data for customers outside the zero-retention exemption.
  • Whether Harvey has added other models to Covered AI Services. Harvey keeps the current list in its customer help center, which requires a login.
  • Whether Fable 5.1 can be used with protected health information under a Harvey Business Associate Addendum.
  • Whether deleting data in Harvey also removes the copy the model company keeps.
  • Whether Anthropic uses its labels on flagged outputs to improve its safety systems, as Harvey's June 2026 help page said.

Our criteria

How we rate

Each color answers one question: may a D.C. lawyer put client confidences and secrets into this plan? Each plan has one color, set by its contract. The settings in each entry lower the risk but do not change the color. The full method, with the ethics rules behind each test, is on the tracker page.

Green

Fit for client confidences and secrets, after your usual vendor checks. All five are true:

  1. (1)A contract forbids the provider to train AI models on your content or use it for its own purposes.
  2. (2)A contract makes your content confidential and requires notice of a subpoena or government demand.
  3. (3)Staff and subcontractors see it only to run, support, or secure the service. Any outside AI model company is bound by the same terms.
  4. (4)Your organization can delete it or set how long it is kept.
  5. (5)The contract requires notice of a security breach, and an independent audit (SOC 2 Type II or ISO 27001) covers the service.
Yellow

Use for client confidences only after you deal with the gap. A business contract forbids training and limits use, but at least one Green condition is missing. Common gaps:

  1. (1)Your organization cannot delete content or set retention.
  2. (2)An outside AI model company works under its own terms.
  3. (3)The contract lets the provider use your content, even de-identified, for its own purposes.
  4. (4)The contract does not treat your content as confidential.
  5. (5)The contract bars personal data, which client files contain.
Red

Keep client confidences and secrets out. At least one of these is true:

  1. (1)The provider trains AI models on your content, or its terms let it.
  2. (2)Your content is used to choose ads.
  3. (3)Only consumer terms apply, so no contract keeps your content confidential.

Turning training off does not change the color. The only exception is the client's informed consent (D.C. Rule 1.6(e)(1)), and D.C. Bar Opinion 388 says clients typically should not be asked to consent to disclosures this broad.

Last checked against the provider's own pages on October 25, 2026. Quotations are kept short and link to the full page. Where the provider does not say something, we say so rather than guess. Ratings describe each provider's published standard terms; large customers often sign negotiated agreements that may differ. General information prepared by the BADC Young Lawyers Section AI Committee, not legal advice. Spotted an error or a change? Write to YLS@badc.org with the subject "Tracker correction."

Questions about the AI Committee, or want to get involved? Write to us directly.

YLS@badc.org