Which AI tools train on your data? (2026)

Training defaults, opt-outs, retention and audits for 20 tools, plan by plan.

Sellside Stack EditorialAs checked November 202613 min read

Disclosure: no affiliate links on this page. Rankings are independent. Research-based, not hands-on tested. How we research →

A CIM, an NDA, a call transcript and a draft term sheet should never end up teaching someone else's model. We read the published policies of 20 tools dealmakers use every week and recorded the training default, opt-out, plan, retention and audits for each. Where a vendor's own page could not confirm a cell, it says unverified.

The key takeaway

Whether your data trains a model depends on your plan, not the brand. ChatGPT, Gemini, Perplexity, Gamma and Grammarly all train on personal-plan content by default. The same vendors do not train on their business or enterprise tiers. If anyone on your team uses a free or personal account for client work, assume that work can be used for training until someone has switched the setting off.

Four other findings matter for deal work:

  • Some paid business tools still train by default. HubSpot trains its own models on customer data unless a Super Admin switches it off. Otter says it trains its own models on de-identified recordings and transcripts, and its privacy page describes no opt-out.
  • "We don't train on your content" can have a carve-out. PandaDoc says customer content does not train LLMs. Its AI Addendum still gives it a perpetual licence to anonymised AI outputs for its own AI training. The only opt-out is switching AI features off.
  • Opt-outs are rarely retroactive. OpenAI, Perplexity, HubSpot and Anthropic all say that switching training off only applies from that point on. Data already used for training stays in the model.
  • Feedback buttons are a back door. On ChatGPT and Claude, a thumbs-up or thumbs-down can send the whole conversation for training or long-term storage, even when training is otherwise off.

The master table

As checked November 2026. "Default" means what happens if nobody changes a setting. "Opt-out" means training is on by default but can be switched off.

Tool and plan Trains on your data by default? Opt-out, and how No-training tier Retention note SOC 2 / ISO 27001 Source
ChatGPT Free / Plus / Pro Opt-out (on by default) Settings → Data controls → turn off "Improve the model for everyone". Temporary Chats are never used. Thumbs-up/down feedback can still send the whole conversation for training. Business, Enterprise, Edu Temporary Chats kept up to 30 days. Opt-out is not retroactive and does not delete chats. SOC 2 Type 2 is stated for business products only. ISO 27001: unverified OpenAI data controls
ChatGPT Business / Enterprise No Not needed. Training only happens if you opt in, for example through feedback. This is the tier Admins control retention. Deleted chats leave OpenAI's systems within 30 days. SOC 2 Type 2. ISO 27001: unverified Enterprise privacy
Claude Free / Pro / Max Your choice, made at sign-up via the "Help improve" setting. Which way the setting starts: unverified Settings → Privacy → "Help improve our AI models". Incognito chats are never used. Chats flagged for safety review may still be used. Team, Enterprise, API 30-day retention if training is off. Up to 5 years (de-identified) if it is on. Feedback is kept 5 years. Consumer-plan scope: unverified Consumer training, retention
Claude Team / Enterprise No Not needed. Owners can disable the thumbs-up/down feedback ("Rate chats"), which otherwise stores the conversation for up to 5 years. This is the tier Chats are kept in the product until deleted. Deleted chats leave back-end systems within 30 days. SOC 2 Type I & II, ISO 27001:2022, ISO 42001 Commercial training, certifications
Microsoft 365 Copilot No Not needed. Prompts, responses and Microsoft Graph data are not used to train foundation LLMs. All M365 commercial Stored like other M365 content. Admins set retention in Purview, and users can delete their Copilot history. ISO 27001 and ISO 42001 listed. SOC 2: unverified (not named on the page) Copilot privacy
Google Gemini consumer app Yes, while "Keep Activity" is on. Whether it starts on: unverified Turn off Keep Activity at myactivity.google.com/product/gemini, or use Temporary Chats. Human reviewers see a sample of chats. Workspace editions Kept 18 months by default. Human-reviewed chats are kept up to 3 years, even after you delete them. 72 hours when Keep Activity is off. Consumer app: unverified Gemini Apps Privacy Hub
Google Gemini in Workspace No (not without your permission) Not needed. No human review. This is the tier Admins set it: 90 days to indefinite for Gemini in Workspace apps, up to 36 months for the Gemini app. Gemini Notebook and Gemini in Chrome are outside the SOC/ISO scope. SOC 1/2/3, ISO 27001, 27017, 27018, 27701, 42001 Workspace AI Privacy Hub
Perplexity Free / Pro / Max Opt-out ("AI data retention" is on by default) Account → Preferences → turn off "AI data retention". Not retroactive. Enterprise Pro, Enterprise Max Enterprise: files deleted after 7 days. Custom session retention needs a Max seat or 50+ Pro seats. Enterprise: SOC 2 Type II. ISO 27001: unverified Data collection, Enterprise Max
Otter.ai Yes: trains its own models on de-identified recordings and transcripts None described on its privacy page. Its third-party AI providers do not train. None stated. Whether an Enterprise contract can exclude training: unverified Deleted conversations are purged from Trash after 30 days. Custom retention and audio removal on Enterprise. SOC 2 Type 2. Policies are "based on" ISO 27001, not certified. Privacy & security
Fireflies.ai No Not needed. "Zero-day" retention agreements with its AI vendors. All plans Custom retention and transcript/summary-only mode on Enterprise SOC 2 Type II, ISO 42001. ISO 27001 not listed. Trust Center
Krisp No: meeting content is used only to deliver features. Krisp has opted out of training with every third-party AI provider. Not needed All plans Desktop transcription runs on your device. Nothing goes to the cloud unless you save it. Saved data is kept until you delete it, then hard-deleted. On-device private mode on Enterprise. SOC 2 Type II, PCI DSS. ISO 27001 not listed. AI Meeting Assistant security
Zoom AI Companion No: audio, video, chat and attachments are not used to train Zoom's or third-party models Not needed All plans Follows each Zoom service's retention. Optional zero-retention mode for meeting summaries. Third-party models are on zero data retention. SOC 2 Type 2 and ISO 27001:2022 both cover AI Companion Zoom AI whitepaper, ISO 27001
Gamma Free / Plus / Pro / Ultra Opt-out (training allowed by default) Avatar → Settings → Data Controls Team, Business, Enterprise (always off, locked) Unverified SOC 2 Type II. ISO 27001 not listed. Gamma data controls, Trust Center
Beautiful.ai No for third-party models ("will not train public LLM models"). Whether Beautiful.ai trains its own models: unverified Generative AI can be switched off for your company on request Not stated Generative-AI providers keep data up to 30 days for abuse monitoring SOC 2 Type II, PCI. ISO 27001 not listed. Beautiful.ai security
Notion AI No: neither Notion nor its AI providers train on customer data by default Not needed All plans AI providers keep data 30 days or less on Free, Plus and Business. Zero retention on most models for Enterprise. Embeddings deleted within 60 days after a page is deleted. SOC 2 Type 2 and ISO 27001 both cover Notion AI Notion AI security
Grammarly individual and online-bought teams Opt-out (on by default) Individuals: account.grammarly.com/security/privacy. Team admins: account.grammarly.com/admin/data_settings. Non-content writing statistics are still used. Plans bought through Grammarly's sales team, and Education (off by default) Unverified SOC 2 Type 2, ISO 27001, 27017, 27018, 27701, 42001 Training control, Security
HubSpot (Breeze) Opt-out: HubSpot trains its own models, some across many customers' accounts Super Admin: Settings → Account management → AI → Access → turn off "AI Model Training". Not retroactive. Third-party AI providers never train. Accounts with Sensitive Data turned on (opted out automatically) Zero retention with third-party AI providers "wherever possible" SOC 2 Type II, SOC 3. ISO 27001: unverified AI model training, Security
Pipedrive AI No ("not without your permission"). But its terms let Pipedrive train on your inputs and outputs when you use Beta AI features. Avoid Beta AI features. Admins can turn AI features off. All plans (for non-beta features) Kept for the life of the subscription. Deleted within 180 days after the account closes. SOC 2, SOC 3, ISO 27001, ISO 27701 Nova FAQ, Terms, Security
Papermark No: AI is off until you turn it on. It runs on OpenAI's API, which does not train on API data. Not needed All plans. A self-hosted option is also available. Embeddings are deleted when you disable AI for a document. Hosted in the EU (Frankfurt) by default. SOC 2 Type II, ISO 27001 Papermark AI data note, Security
DocSend (Dropbox) No: Dropbox says it will not build generative AI models from customer content without consent AI features can be turned off in settings All plans Unverified SOC 2: yes. ISO 27001: no (per Dropbox's audited-features list) Dropbox AI principles, Audited features
PandaDoc Opt-out, in part. Document content does not train LLMs. Anonymised AI outputs, metadata and usage data may train PandaDoc's own models. Admins: Settings → Data privacy settings → turn AI features off. That is the only opt-out. AI is on by default (off by default for HIPAA accounts). None stated (the opt-out means losing AI features) PandaDoc keeps the derived "AI Improvement Data" and has no obligation to delete it, even after the contract ends SOC 2 Type 2. ISO 27001: not claimed for PandaDoc itself, only its AWS hosting. AI policy, AI Addendum, AI settings
Spellbook No: third-party LLMs on zero retention, with no training. Personalisation stays within your organisation. Not needed. De-identified, aggregated usage data is used for product improvement. All plans Quality-assurance copies of inputs and outputs kept up to 90 days SOC 2 Type II. ISO 27001 not listed. Security, Terms §4
Reclaim.ai No for generative AI without consent. But its business terms let it use anonymised customer data to train its machine-learning algorithms. Third-party LLM features run only if you opt in Not stated Unverified (stores primary-calendar data; hosted on AWS in the US) SOC 2 Type II. ISO 27001 not listed (applies only to AWS). AI disclosure, Privacy, Business terms §4.3, Security

"Not listed" means the vendor's own security page names other certifications but not this one. "Unverified" means we could not find the answer on a vendor page.

Safest defaults for confidential deal work

Based only on published policy, these setups do not train on customer content by default and have independent audits behind them:

  • General assistant: ChatGPT Business or Enterprise, Claude Team or Enterprise, Microsoft 365 Copilot or Gemini in Workspace. Choose the one that matches the email and document stack your firm already runs. Copilot and Gemini inherit your existing M365 or Workspace permissions and retention, which is usually the easiest case to make to compliance.
  • Meeting notes: Fireflies (no training on any plan; Enterprise adds custom retention and transcript-only mode), Zoom AI Companion if you already meet on Zoom, or Krisp if you want transcription to stay on your device. Avoid Otter for client or MNPI calls: it says it trains its own models on de-identified customer recordings.
  • Documents and data rooms: Notion AI for internal knowledge, and Papermark or DocSend for sharing documents outside the firm. Papermark's AI stays off until you switch it on, and you can self-host it.
  • Contracts: Spellbook has zero-retention terms with its model providers and no cross-customer training. On PandaDoc, read the AI Addendum before using its AI features on live deal documents.
  • CRM: Pipedrive has no training without permission, as long as you skip Beta AI features. On HubSpot, ask a Super Admin to switch off AI Model Training on day one.

Two rules apply whichever tool you pick:

  1. Use firm-managed accounts only. A banker's personal ChatGPT Plus or Gemini account is the most common way client data reaches training, because both default to training.
  2. Turn off the feedback buttons, or train people not to press them. On ChatGPT and Claude, a thumbs-up or thumbs-down sends the whole conversation for training or 5-year storage. Claude Team and Enterprise owners can disable feedback with the "Rate chats" setting.

Checklist: vetting a new AI tool before you upload a CIM or NDA

Ask these questions, and get the answers in writing. A sales deck does not count.

  1. Does the vendor train any model on our content, including its own models, by default and on our exact plan? Get the answer in the contract or DPA, not just an FAQ.
  2. Is "content" defined narrowly? Ask about AI outputs, metadata, usage data, embeddings and "anonymised" or "de-identified" data. PandaDoc, Reclaim and Spellbook all draw lines here.
  3. Who switches training off, and is the setting at user level or account level? Can it be locked so individual users cannot turn it back on?
  4. Is the opt-out retroactive? If not, switch it off before anyone uploads anything.
  5. Do the vendor's AI providers (OpenAI, Anthropic, Google and others) have no-training and zero-retention terms? Get the subprocessor list.
  6. How long is content kept, and can we set our own retention? Ask separately about deleted content, backups and human-reviewed samples.
  7. Do humans review content? Ask who, why, and whether you can switch it off.
  8. What audits does the vendor hold, and is the AI feature in scope? A SOC 2 report that covers the core app but not the AI feature, as with Gemini Notebook, is a gap.
  9. Do feedback buttons, beta features or "labs" features change the rules? On ChatGPT, Claude and Pipedrive, they do.
  10. Where is data processed? Check the region and any cross-border transfers, especially for EU or UK targets and for MNPI.
  11. Can you turn the AI off without losing the core product?
  12. Does the tool fit your NDA? Many NDAs limit disclosure to "representatives". Check whether a third-party AI processor counts as one, and whether the counterparty's process letter restricts AI use.

Methodology

  • Scope: 20 tools that bankers, VC/PE, founders, IR and corp dev teams commonly use, with separate rows where the consumer and business tiers differ.
  • Sources: Only the vendor's own privacy policies, terms, AI addenda, trust centres, security pages and help-centre articles. We did not use third-party reviews or search-result snippets for any cell. Each source page was retrieved directly in November 2026, and the link is in the table.
  • Definitions: "Trains by default" refers to training or improving AI or machine-learning models with customer content (prompts, files, recordings, outputs) when nobody has changed a setting. Where a vendor separates its own models from third-party models, or content from usage data, we say so in the cell.
  • Certifications: We list only what the vendor states for itself. Certifications that belong only to the vendor's cloud host (for example, AWS ISO 27001) are not counted. We did not inspect audit reports, which are usually available only under NDA.
  • Unverified cells: Where the vendor's own pages did not answer the question, the cell says "unverified". It does not mean the vendor lacks the control.
  • Limits: Vendor policies change often, and negotiated enterprise contracts can override published terms. This guide describes published policy, not actual practice. It is not legal advice.

Sources

  • OpenAI, Data controls in ChatGPT: https://help.openai.com/en/articles/7730893-data-controls-faq
  • OpenAI, Enterprise privacy: https://openai.com/enterprise-privacy/
  • Anthropic, Is my data used for model training? (consumer): https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training
  • Anthropic, Is my data used for model training? (commercial): https://privacy.anthropic.com/en/articles/7996868-is-my-data-used-for-model-training
  • Anthropic, How do I change my model improvement privacy settings?: https://privacy.anthropic.com/en/articles/12109829-how-do-i-change-my-model-improvement-privacy-settings
  • Anthropic, How long do you store my data? (consumer): https://privacy.claude.com/en/articles/10023548-how-long-do-you-store-my-data
  • Anthropic, How long do you store my organization's data?: https://privacy.claude.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data
  • Anthropic, Updates to Consumer Terms and Privacy Policy: https://www.anthropic.com/news/updates-to-our-consumer-terms
  • Anthropic, What certifications has Anthropic obtained?: https://support.claude.com/en/articles/10015870-what-certifications-has-anthropic-obtained
  • Microsoft, Data, Privacy, and Security for Microsoft Copilot: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy
  • Google, Gemini Apps Privacy Hub: https://support.google.com/gemini/answer/13594961
  • Google, Generative AI in Google Workspace Privacy Hub: https://support.google.com/a/answer/15706919
  • Perplexity, Data Collection at Perplexity: https://www.perplexity.ai/help-center/en/articles/11564572-data-collection-at-perplexity
  • Perplexity, What is Enterprise Max?: https://www.perplexity.ai/help-center/en/articles/12310544-what-is-enterprise-max
  • Otter.ai, Privacy & Security: https://otter.ai/privacy-security
  • Otter.ai, Pricing (Enterprise retention features): https://otter.ai/pricing
  • Fireflies.ai, Trust Center: https://trust.fireflies.ai/
  • Fireflies.ai, Pricing (Enterprise retention features): https://fireflies.ai/pricing
  • Krisp, Security for AI Meeting Assistant: https://krisp.ai/security-for-ai-meeting-assistant/
  • Zoom, Models, Processing, Storage, and Usage: https://library.zoom.com/ai/ai-whitepaper/models-processing-storage-and-usage
  • Zoom, ISO 27001: https://www.zoom.com/en/trust/legal-compliance/iso-27001/
  • Zoom, SOC 2 Type 2: https://www.zoom.com/en/trust/legal-compliance/soc2/
  • Gamma, Does Gamma use my content to train its AI features?: https://help.gamma.app/en/articles/12281928-does-gamma-use-my-content-to-train-its-ai-features
  • Gamma, Trust Center: https://trust.gamma.app
  • Beautiful.ai, Security: https://www.beautiful.ai/security
  • Notion, Notion AI security & privacy practices: https://www.notion.com/help/notion-ai-security-practices
  • Grammarly, Product Improvement and Training Control: https://support.grammarly.com/hc/en-us/articles/25555503115277-Product-Improvement-and-Training-Control
  • Grammarly, Security: https://www.grammarly.com/security
  • HubSpot, Understand HubSpot's AI model training: https://knowledge.hubspot.com/account-management/hubspot-ai-mode-training
  • HubSpot, Security and Compliance: https://www.hubspot.com/security-and-compliance
  • Pipedrive, Nova FAQ: https://support.pipedrive.com/en/article/nova-faq
  • Pipedrive, AI CRM FAQ: https://www.pipedrive.com/en/products/ai-crm
  • Pipedrive, Terms of Service (Beta AI services): https://www.pipedrive.com/en/terms-of-service
  • Pipedrive, Privacy and security: https://www.pipedrive.com/en/features/privacy-security
  • Papermark, AI data note: https://www.papermark.com/news/papermark-ai-on-product-hunt
  • Papermark, Security: https://www.papermark.com/security
  • Dropbox, AI Principles: https://www.dropbox.com/ai-principles
  • Dropbox, Audited Features List: https://assets.dropbox.com/documents/en-us/legal/audited-features-current.pdf
  • Dropbox, AI organize on DocSend: https://help.dropbox.com/organize/ai-organize-space-data-room
  • PandaDoc, AI Policy: https://support.pandadoc.com/en/articles/13794735-pandadoc-ai-policy
  • PandaDoc, AI Addendum: https://www.pandadoc.com/legal/ai-addendum
  • PandaDoc, AI data privacy settings: https://support.pandadoc.com/en/articles/16106373-ai-data-privacy-settings
  • PandaDoc, Compliance: https://www.pandadoc.com/security/compliance/
  • Spellbook, Security: https://www.spellbook.legal/security
  • Spellbook, Terms of Service: https://www.spellbook.legal/legal/terms-of-service
  • Reclaim.ai, AI Disclosure: https://help.reclaim.ai/en/articles/13178785-reclaim-ai-disclosure
  • Reclaim.ai, Privacy Policy: https://reclaim.ai/privacy
  • Reclaim.ai, Business Terms of Service: https://reclaim.ai/business-terms
  • Reclaim.ai, Security: https://reclaim.ai/security