Skip to content

Rights and contracts

Salesforce Data Cloud training data rights: context is not a license

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Salesforce Data Cloud, now Data 360, sells access to a customer's records as context for AI; it does not create training data rights. Salesforce's agreement gives it no title to customer data, outside model providers keep none of it, and generative models are not currently trained on it, though its help pages say global predictive models may train on it by default, with an opt-out. Licensing records to a model developer stays the customer's decision.

License terms

Salesforce data Cloud

This page covers salesforce Data Cloud training data rights: context is not a license.Illustrative

Key takeaways

  • Data 360, renamed from Data Cloud on October 14, 2025, supplies customer records as context to Salesforce apps and agents.
  • Salesforce's agreement gives it no title to customer data, only a license to process it as needed to provide and run the services.
  • Salesforce's help documentation says customer data may be used by default to train global predictive models unless the organization opts out or uses Government Cloud; generative training would need an opt-in.
  • Agentforce and Data 360 ARR, more than $2.9 billion at Q4 fiscal 2026, measures subscriptions and includes Informatica; it is not a training-rights market.
  • The right to license records to a model developer stays with the customer, under its own contracts, privacy duties and tool terms.

Does Salesforce Data Cloud give anyone the right to train on customer records?

No single grant in Salesforce's published terms gives anyone a right to train AI models on the records held in Data Cloud, now called Data 360. Three permissions have to be read separately: the product access a customer buys, the uses Salesforce reserves for itself, and the right to license the records to an outside model developer, which those terms do not hand to Salesforce at all.

That separation is the investor's decision point. Salesforce reports fast growth in its data and AI products, and it is tempting to read that as a company sitting on a training-data asset. The published terms describe something narrower: a subscription that lets a customer's own records serve as context for AI features, a limited set of training uses Salesforce keeps for its own predictive models, and no route by which the records become a dataset licensed to someone else.

Two definitions keep the rest precise. Context means records supplied to a model at the moment it answers, then discarded, without changing the model. Training means records used to change a model's weights, so their influence stays after the copies are deleted. Most of what Data 360 sells is the first.

Product access: what Data 360 sells

Data 360 sells unified, governed access to a customer's records so that Salesforce applications and agents can use them. Salesforce's help documentation says that as of October 14, 2025, Data Cloud was rebranded as Data 360, with functionality unchanged and existing Data Cloud agreements honored to the end of their term.

The product page describes zero-copy integrations that connect Data 360 directly to platforms such as Snowflake and Databricks without duplicating the data. That design is telling: the value is in reaching records where they already live and putting them in front of an application or agent at the right moment, not in collecting them into a corpus Salesforce controls.

For an AI feature, that access is grounding. An agent that answers a service question pulls the relevant case history, order and policy into its prompt, and the answer depends on those records. The model it calls has not been trained on them, and under the terms covered next, the outside model provider may not keep them.

Illustrative: a regional insurer keeps eight years of claims and service cases in Salesforce. Switching on a service agent lets the agent read those cases as context when it answers. Leaving the default sharing setting on may let Salesforce use them to train its global predictive models, according to its help documentation. Neither step lets the insurer, Salesforce or anyone else hand the case history to a model developer as a training dataset; that would take the insurer's own decision, a review of its customer contracts and privacy notices, and a separate license.

What Salesforce's own terms let it do with customer data

Salesforce's Main Services Agreement keeps ownership with the customer and ties Salesforce's license to providing the service. In the version on salesforce.com, as excerpted on October 28, 2026, Salesforce acquires no right, title or interest in customer data beyond a license to host, copy and process it as appropriate to provide and ensure proper operation of the services, and processing by AI-enabled services follows the customer's documented instructions.

There is one training use by default, and it is easy to miss. A Salesforce help article on managing access to customer data says that, with sharing enabled, a customer consents to Salesforce using its data to train global predictive AI models for the services and features it can use, and that this access is allowed by default unless the organization uses Government Cloud or has opted out; the help pages also say Marketing Cloud Engagement is opted out by default.

Generative models are treated differently. Salesforce's Agentforce privacy FAQ says Salesforce does not currently use customer data to train generative AI models, and that any future generative training would need the customer's opt-in under the applicable product terms. So the published position is a ladder: context use as part of the service, predictive training unless the customer opts out, generative training only with an opt-in that does not exist today, and licensing to outsiders, which none of the pages read provides for.

A four-rung ladder of uses of Salesforce customer records: context for apps and agents is part of the service, predictive model training is allowed by default with an opt-out, generative model training is not done today and would need an opt-in, and licensing records to an outside model developer is not offered and stays with the customer.
Salesforce's published positions on customer data, from its agreement, help articles and privacy FAQ read on October 28, 2026.

Outside model providers: zero retention, no training

Outside model providers see the records only to answer a request and, Salesforce says, keep none of them. Salesforce's help article on the Einstein Trust Layer says Salesforce partners with all its external model providers, such as OpenAI, to enforce a zero-data retention policy, and that no data is used for model training or product improvement by third-party models.

For investors, this closes the most obvious path by which Salesforce's customer records could become training data for frontier models. A lab serving Salesforce's agents sees each prompt once and, under the contract Salesforce describes, keeps none of it. Any lab that wants those records for training has to obtain them from the company that owns them, under that company's own authority.

Dated evidence table: three data platforms and SourceX, read on their own pages

The pattern is consistent across the three platforms. Each sells processing of the customer's own records, and each publishes limits on training shared models with them, Salesforce's default-on predictive training, as its help documentation describes it, being the notable exception. None of the pages read offers to license a customer's records to an outside model developer, which is the right a training-data business depends on, so their growth figures say little about the value of that right.

The comparison below puts Salesforce next to two data platforms that also run AI on customer data, Snowflake (listed) and Databricks (private), and next to SourceX, using only each company's own pages read on October 28, 2026. It compares positions, not performance: no company is ranked, and a blank or not-read cell means the pages read did not address the point, not that the answer is no.

Dated evidence table: three data platforms and SourceX, read on their own pages
QuestionSalesforce, Data 360SnowflakeDatabricksSourceX
What happens to customer recordsUnified and supplied as context to apps and agents; zero-copy links to Snowflake and DatabricksAI features run on customer data, with each input and output isolated from other customersAssistive features and model serving run on customer dataA defined, de-identified dataset is licensed from the company for AI training
Provider trains shared models on itHelp documentation: customer data may be used by default for global predictive models unless the organization opts out or uses Government Cloud; generative models not currently, opt-in neededNo model offered to other customers is trained on customer data, inputs or outputsAssistive-feature data not used for foundation models offered to others; paid model serving inputs not used to train any modelTraining is the licensed purpose, under terms the company approves and signs before delivery
Outside model providersZero data retention; no training or product improvementNot read for this tablePartner-powered features use zero data retention endpointsMay sublicense or distribute the de-identified dataset as the agreement allows
Right to license records to a model developerNot granted to Salesforce; customer keeps titleNot addressed in the page readNot addressed in the pages readThe company itself grants the license; AI-training exclusivity, its length and any alternatives set in each deal's confirmation
Data or AI metric publishedAgentforce and Data 360 ARR more than $2.9 billion at Q4 fiscal 2026; nearly $3.9 billion at Q2 fiscal 2027 on a wider definitionNot collected for this tableNone used: private company, no figure from its own release relied on hereNone published; shared with investors directly

What Agentforce and Data 360 ARR does and does not measure

Salesforce's headline data and AI figure measures subscriptions to access and AI features, not any market in training rights. The company defines Agentforce and Data 360 ARR as the annualized recurring value of active Data 360 and certain generative AI subscription agreements at the end of the period.

In its fourth-quarter fiscal 2026 release, dated February 25, 2026, Salesforce reported Agentforce and Data 360 ARR of more than $2.9 billion, up more than 200% from a year earlier, including $1.1 billion of Informatica Cloud ARR and $800 million of Agentforce ARR. The same release gave fiscal 2026 revenue of $41.5 billion, up 10%. ARR is a run-rate of active subscription agreements at a date; revenue is what was recognized over the year. Neither converts into the other.

The figure also moves for reasons other than demand. Its fourth-quarter total includes a business Salesforce acquired, Informatica, and in the second quarter of fiscal 2027 Salesforce reported nearly $3.9 billion while widening Agentforce ARR, which exceeded $1.5 billion, to include Slackbot and Headless 360 offerings. A growth rate computed across that change compares two different definitions.

The strongest alternative: Salesforce is already a training-data intermediary

The best case against this reading is that Salesforce already trains on customer data and could extend that use. Salesforce's help documentation says customer data may be used by default to train global predictive models, and the Agentforce FAQ leaves the door open to generative training with an opt-in. On that view, Data 360's growth puts more customer records within reach of a future training program, and the subscription is the wedge.

The published terms cut against the strong version of that argument. The training Salesforce reserves is for its own predictive services to customers who have not opted out; the generative path requires an opt-in under product terms; and nothing read here lets Salesforce license customer records to an outside model developer, a step that would collide with the agreement's ownership language and with every customer's own contracts and privacy duties.

Terms can change, so the honest conclusion is conditional. As published on October 28, 2026, Salesforce's data and AI economics are product-access economics. An investor who prices Data 360 as a training-data asset is pricing a right the company's own pages do not claim.

What could make this reading of Salesforce wrong

Each of these would change or weaken the reading above:

  • The agreement and help articles are versioned documents; the agreement excerpt read here may be older than the version a given customer signed, and product-specific terms can add or narrow uses.
  • Help articles describe settings and defaults, not contract language; a reviewer should confirm that the predictive-model consent default appears in the applicable product terms.
  • Every Salesforce, Snowflake and Databricks statement here was read through search excerpts restricted to the company's own domain, not the full page.
  • Agentforce and Data 360 ARR combines Data 360, generative AI subscriptions and acquired Informatica revenue, and its definition widened in fiscal 2027, so it cannot isolate demand for any one product.
  • A customer may still license its own records to a model developer through other channels; whether it can depends on its own contracts, privacy notices and the export terms of each tool, which no vendor page settles.
  • For a company licensing business records, including SourceX, platform terms that restrict export or training could narrow the records that can be offered.

Questions to ask Salesforce, a data platform or SourceX

SourceX works on the top rung of the ladder, the one Salesforce's terms leave with the customer. It licenses business data directly from companies for AI training, and nothing is delivered until the company has approved the scope and terms and signed. Its live legal-framework page names Salesforce among the tools whose own terms govern exported data, says those terms are checked during review, and says data a tool's terms restrict is excluded or the deal is scoped to fit. The supplier agrees not to license the selected data to anyone else for AI training; the length of that commitment, and any alternatives, are written into each deal's Dataset Deal Confirmation. SourceX publishes no commercial metrics; they go to investors directly, and its Rule 506(c) raise is open only to verified accredited investors, with materials on request at /investors.

The questions that separate context from training rights apply to Salesforce, to any data platform and to SourceX:

Checklist · 6 items

  1. Which uses of customer data are part of the service, which are training uses the provider keeps, and which require the customer's opt-in?
  2. Is any training use on by default, and how many customers have opted out?
  3. Do outside model providers retain prompts or outputs, and under what contract?
  4. Does any document grant the provider a right to license customer records to a third party for model development?
  5. Which subscriptions, acquisitions and definition changes are inside the published AI metric?
  6. For a licensing company: who signs, which tool terms were checked, and what was excluded as a result?

0 of 6 ticked. Tick each item as you go.

Methods note: pages read, definitions and limits

Pages read, all on October 28, 2026: Salesforce's help articles on Data 360, the Einstein Trust Layer and managing access to customer data; its Data 360 product page; its Main Services Agreement and Agentforce privacy FAQ PDFs; its fourth-quarter fiscal 2026 results release (February 25, 2026) and second-quarter fiscal 2027 results release; Snowflake's AI Trust and Safety FAQ; Databricks' documentation pages on AI assistive features trust and safety and on model serving; and SourceX's live how-it-works and legal-framework pages. Direct downloads from these hosts are blocked from this environment, so every statement comes from a search excerpt restricted to the publisher's domain; a reviewer must open each page before release.

Method: each company's position on four questions (what happens to records, provider training, outside model providers, licensing to a model developer) was recorded in its own words where the excerpt allowed, then paraphrased; metrics were taken only from a listed company's own release, with the company's definition. No figure is used for the private company, Databricks.

Not measured: opt-out rates, how often customers license records to model developers, prices, the content of any customer's signed agreement, and the terms of products other than those named. This is general information about published terms, not legal advice; counsel reviews a specific contract.

Frequently asked questions

Does Salesforce train AI models on Data Cloud customer data?

Its help documentation says customer data may be used by default to train global predictive AI models unless the organization opts out or uses Government Cloud. Its Agentforce privacy FAQ says customer data is not currently used to train generative AI models, and future generative training would need an opt-in.

Do OpenAI or other model providers keep Salesforce customer data?

Salesforce says its external model providers operate under a zero-data retention policy and that no data is used for model training or product improvement by third-party models. That is a contractual commitment Salesforce describes, not an independent audit.

Is Data 360 the same product as Data Cloud?

Yes. Salesforce's help documentation says Data Cloud was rebranded as Data 360 on October 14, 2025, with functionality unchanged and existing agreements honored through the current term.

Can a company license its Salesforce records to an AI developer?

Salesforce's agreement leaves ownership with the customer, so the question turns on the customer's own contracts, privacy notices and the export terms that apply. This is general information, not legal advice; counsel should review the specific agreements.

What does Agentforce and Data 360 ARR measure?

Salesforce defines it as the annualized recurring value of active Data 360 and certain generative AI subscription agreements at period end. It includes Informatica Cloud ARR and, from Q2 fiscal 2027, a wider Agentforce definition, so it is not comparable across that change.

Sources

  • Data Cloud was rebranded as Data 360 as of October 14, 2025; functionality unchanged; existing agreements honored through the current term. Source
  • Zero-copy integrations connect Data 360 directly to platforms such as Snowflake and Databricks without duplicating data. Source
  • Salesforce acquires no right, title or interest in customer data beyond a license to host, copy and process it to provide the services; AI-enabled processing follows the customer's documented instructions. Source
  • With sharing enabled, customers consent to Salesforce training global predictive AI models on their data; access is on by default unless the org uses Government Cloud or opts out. Source
  • Salesforce does not currently use customer data to train generative AI models; future generative training would require customer opt-in under product terms. Source
  • Salesforce enforces zero data retention with external model providers such as OpenAI; no data is used for model training or product improvement by third-party models. Source
  • Q4 fiscal 2026: Agentforce and Data 360 ARR more than $2.9 billion, up over 200%, including $1.1 billion Informatica Cloud ARR and $800 million Agentforce ARR; fiscal 2026 revenue $41.5 billion, up 10%; ARR definition. Source
  • Q2 fiscal 2027: Agentforce and Data 360 ARR nearly $3.9 billion; Agentforce ARR over $1.5 billion, now including Slackbot and Headless 360. Source
  • Snowflake does not use Customer Data or Usage Data, including AI inputs and outputs, to train any model offered to other customers; inputs and outputs are isolated. Source
  • Databricks does not use data, prompts or responses from assistive features to train foundation models offered to third parties; partner-powered features use zero data retention endpoints. Source
  • For all paid accounts, Databricks Model Serving does not use inputs or outputs to train any models or improve Databricks services. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies