Cohere vs. Salesforce Einstein / Agentforce: data protection per plan

no (trial keys are subject to the terms of use and privacy policy; the Enterprise Data Commitments and the DPA offer apply to paying SaaS customers. Privacy policy: trial environments are not intended for personal data; for that, switch to an enterprise plan with a contract)

yes for SaaS/Model Vault customers on request (privacy@cohere.com, sent after NDA; the DPA contains the EU standard contractual clauses of 2021-06-04). According to the Enterprise Data Commitments, no DPA is required for private and third-party cloud deployments.

yes (on request: SaaS customers receive the DPA via privacy@cohere.com; according to the Trust Center, sent after a signed NDA)

yes (on request)

yes (on request)

yes (on request)

yes (on request)

yes (on request)

yes (on request)

Privacy policy (as of 2026-05-01, section 2.3): for Trial Users and Researchers, Cohere collects inputs and outputs and may use them for research and development, including 'using inputs/outputs to improve the performance and safety of Cohere Products'; data is de-identified beforehand; users should not enter personal data. An opt-out for trial keys is not described (the dashboard switch 'Data Controls' is described in the Enterprise Data Commitments for paying customers). According to the Enterprise Data Commitments, the Terms of Use and the privacy policy apply to trial keys; the Terms of Use (as of 2022-09-07, section 4.1) allow Cohere to use 'Customer Data', among other things, to 'IMPROVE AND ENHANCE THE COHERE SOLUTION' and other offerings, including by sharing API data and fine-tuning data with third parties.

Enterprise Data Commitments: for private and third-party cloud deployments, Cohere does not receive prompts and generations. Privacy policy: data in the customer's or cloud provider's environment is not accessible to Cohere and is not used for training. Model Vault as a platform managed by Cohere: use for training not separately addressed in the documentation or the Enterprise Data Commitments; for Encrypted Vault, prompts and responses are, according to the documentation, not visible in plaintext even to Cohere.

Enterprise Data Commitments (as of 2025-12-05): opt-out from the use of prompts and generations for training in the dashboard settings ('Data Controls', switch set to 'Off'); the change applies to data generated thereafter. For abuse monitoring, prompts and logs may be reviewed; PII is filtered before any possible use for training. Model Training Privacy Notice (as of 2026-05-01): 'Where a user has given Cohere permission to use inputs and outputs for model training', the data is de-identified beforehand. Default value of the switch for new accounts: not publicly documented.

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7). SPARC and Notices apply to the "Covered Services branded as Agentforce" without differentiating by edition; Salesforce Foundations is not named separately there (q7, q8).

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7).

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7).

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7).

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7).

Documentation does not differentiate by plan. Trust Layer: "prompts and generated responses are never stored or used to train the underlying third-party large language models" (Zero Data Retention). Einstein Platform Notices (Published 2026-05-22, q7): OpenAI, Azure, Google and Anthropic (via AWS) "have contracted to not otherwise retain or use Customer Data sent to the model" (apart from automated safety classification). Use by Salesforce itself: the Einstein Platform SPARC (Published 2026-07-24, q8) permits the use of customer data "on an aggregate basis for purposes such as research, marketing, analysis, and benchmarking, and other purposes reasonably required to develop, deliver, and provide ongoing innovation to the Covered Services", without personally identifiable data; the documentation contains no express statement on training Salesforce's own models. For Einstein Conversation Insights, Salesforce employees may review customer data "for the purposes of improving and training the product" (q7).

Trust Center FAQ: hosting on Google Cloud servers in US-Central; according to the FAQ, Cohere does not use servers outside the USA.

Region depends on the deployment at the customer or the cloud provider. Model Vault (Cohere-managed, single-tenant): documentation 'Compliance' for Encrypted Vault: 'Region selection supports data-residency requirements; contact Cohere to confirm the regions available for your vault.' There is no public list of regions.

Trust Center FAQ: hosting on Google Cloud servers in US-Central. According to the FAQ, in certain cases the offering can be configured so that no customer data is stored at Cohere ('ephemeral').

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

Infrastructure-and-Sub-processors documentation (Published 2026-10-02): Einstein Generative AI Services process at OpenAI in the USA (failover rerouting to Azure, priority to the org region), at Azure OpenAI in several regions incl. the EU (e.g. France, Germany, Sweden), at Google/AWS in the nearest disclosed region (e.g. Germany, Netherlands; failover via US endpoints); batch processing via Azure OpenAI uses global endpoints without geo-routing. According to a footnote, Agentforce Coworker only in certain regions (e.g. USA, Netherlands or Germany, France, Italy, Sweden). Org hosting on Hyperforce/AWS in selectable regions incl. Germany, France, Ireland and others.

• Google Cloud (infrastructure): USA
• Fullstory (software delivery): USA
• LaunchDarkly (feature flags): USA
• New Relic (platform monitoring): USA
• Retool (software delivery): USA
• Sentry IO (software delivery): USA
• Segment (event and error processing): USA
• Sendgrid (e-mail communication): USA
• Vercel (web application hosting): USA
• Salesforce (customer support management): USA
• Wiz (cloud security and compliance): USA
• Forethought (customer support management): USA

• Google Cloud (infrastructure): USA
• Fullstory (software delivery): USA
• LaunchDarkly (feature flags): USA
• New Relic (platform monitoring): USA
• Retool (software delivery): USA
• Sentry IO (software delivery): USA
• Segment (event and error processing): USA
• Sendgrid (e-mail communication): USA
• Vercel (web application hosting): USA
• Salesforce (customer support management): USA
• Wiz (cloud security and compliance): USA
• Forethought (customer support management): USA

• Google Cloud (infrastructure): USA
• Fullstory (software delivery): USA
• LaunchDarkly (feature flags): USA
• New Relic (platform monitoring): USA
• Retool (software delivery): USA
• Sentry IO (software delivery): USA
• Segment (event and error processing): USA
• Sendgrid (e-mail communication): USA
• Vercel (web application hosting): USA
• Salesforce (customer support management): USA
• Wiz (cloud security and compliance): USA
• Forethought (customer support management): USA

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

• OpenAI, L.L.C. (generative AI, USA; on failover rerouting to an Azure endpoint)
• Microsoft Corporation (Microsoft Azure / Azure OpenAI; regions incl. USA, Switzerland, France, Germany, Sweden, UK and others)
• Google LLC (generative AI; "disclosed region closest to Customer's Salesforce Org", e.g. Germany, Netherlands; failover via Google endpoint in the USA)
• Amazon Web Services, Inc. (hosting and generative AI provider, including Anthropic models via AWS according to the Notices q7; Agentforce Speech: AWS Polly/Transcribe)
• Bright Data Inc. (web search for generative AI, only if enabled by the customer)
• SuSea, Inc. (dba You.com; web search for generative AI, only if enabled by the customer)
• Temporal Technologies Inc. (orchestration of background workflows; incl. USA, Germany, UK)
• Eleven Labs Inc. (text-to-speech/speech-to-text, Agentforce Speech Foundations)
• Deepgram, Inc. (speech-to-text, Agentforce Speech Foundations; also self-hosted on AWS)

Free (pricing page: calls with a trial key are free, rate-limited)

North and Compass: custom (pricing page). Model Vault (Standard Vault) per instance, in USD, per hour / per month / per year (documentation 'Standard Vault Pricing'): Embed 5 Fast and Embed 5 Pro Small USD 3 / USD 2,000 / USD 20,000, Medium USD 5 / USD 3,250 / USD 32,500; Rerank 4 Fast, Rerank 4 Pro and Rerank 3.5 Medium USD 5 / USD 3,250 / USD 32,500; Rerank 4 Pro Large USD 10 / USD 6,500 / USD 65,000; Parse 5 Medium USD 4 / USD 2,500 / USD 25,000, Large USD 8 / USD 5,000 / USD 50,000, XL USD 7 / USD 4,300 / USD 43,000; Cohere Transcribe Medium USD 3.75 / USD 2,500 / USD 25,000, Large USD 7.50 / USD 4,750 / USD 47,500. Generative models on Standard Vault (mostly via waitlist) per instance and hour, L / XL: Command A USD 40 / USD 48, Command A Reasoning USD 48 / USD 57.50, Command A+ USD 17.50 / USD 32.50, North Mini Code USD 7.50 / USD 10.50. Encrypted Vault priced separately. According to the Terms of Use, prices, except where otherwise indicated, do not include applicable sales taxes ('do not include any applicable sales taxes').

In USD (pricing page): Command R input USD 0.15, output USD 0.60 per 1M tokens; Command R7B USD 0.0375 / USD 0.15; Embed 5 Pro USD 0.12, Embed 5 Fast USD 0.08 per 1M tokens; Rerank 4 Pro USD 2.50, Rerank 4 Fast USD 2.00 per 1,000 searches (1 search = 1 request with up to 100 documents); Parse 5 USD 1.50 per 1,000 pages. Legacy models for existing customers include Command R+ 08-2024 USD 2.50 / USD 10.00; Aya Expanse USD 0.50 / USD 1.50. Billing monthly or when USD 250 is outstanding. Command A+ and North Mini Code free as open-source models (Apache 2.0). According to the Terms of Use, prices, except where otherwise indicated, do not include applicable sales taxes ('do not include any applicable sales taxes').

USD 0 (Salesforce Foundations, incl. Agentforce Builder, Prompt Builder, Agentforce Coworker and others; as of 2026-10-14)

USD 500 per 100,000 Flex Credits (as of 2026-10-14; pricing page shows a USD/EUR converter, EUR 500 per 100k)

USD 2 per conversation; for customer-facing agents, cannot be used in the same org as Flex Credits (as of 2026-10-14)

USD 5 per user/month, requires Flex Credits (as of 2026-10-14)

USD 125 per user/month (Sales/Service/Field Service/Operations add-ons) or USD 150 per user/month (Industries add-ons), unlimited employee agent usage (as of 2026-10-14)

from USD 550 per user/month, includes Agentforce add-on and 2.75 million Flex Credits per org/year (as of 2026-10-14). Sales Cloud pricing page (as of 2026-10-18): Max USD 550 or EUR 550 per user/month, billed annually, no indication whether VAT is included (q9)

yes (trial API key with rate limit, not for production or commercial use)

yes (workplace platform North, search system Compass, dedicated model instances 'Model Vault' incl. Transcribe, deployment in a VPC or on-premises as well as via the cloud platforms of OCI, Azure, AWS, Google Cloud)

yes (Command R, Command R7B, Embed 5 Pro/Fast, Rerank 4 Pro/Fast, Parse 5; older models for existing customers; Aya Expanse via the API)

yes

yes

yes

yes

yes

yes