Salesforce’s Spring ‘26 release added a self-service toggle that lets administrators opt out of sharing customer data for global predictive AI model training. The control is new. The default is not. Einstein has ingested customer data for training since at least 2018 under Salesforce’s Main Services Agreement, and opting out previously required filing a support case. The toggle merely surfaces a long-standing data-sharing posture that compliance teams may have assumed did not exist.
The Toggle Is New; the Default Is Old
The new control lives at Setup → Einstein → Opt Out of Customer Data Access. Before Spring ‘26, an administrator who wanted to prevent their org’s data from training Salesforce’s global predictive models had to open a support case. Now the choice is exposed in Setup, but the default remains enabled for most orgs. Government Cloud customers and organizations that previously opted out through support are excluded from the automatic opt-in.
According to [Salesforce Ben’s coverage]1, the resulting community debate peaked the week of April 28, fueled in part by Salesforce MVP Francis Pindar. The anger is not about a sudden policy change. It is about the realization that the Main Services Agreement has authorized this use since at least 2018. Customers who did not read the MSA closely, or who assumed that “AI” in 2026 meant something different than “Einstein” in 2018, are discovering that their data has been training global models for years.
What the Setting Actually Controls (and What It Doesn’t)
The toggle governs Customer Data Access for global predictive AI models: Einstein lead scoring, opportunity insights, and similar features that rely on patterns across all Salesforce customers. When enabled, Salesforce uses aggregated customer data to train these models, improve existing services, and conduct R&D. When disabled, the org falls back to org-specific models. Einstein does not stop working; it simply loses the global signal.
How to Opt Out and What Happens Next
Disabling the toggle stops new customer data from being shared for global model training. Previously collected data is permanently deleted, but Salesforce specifies that deletion can take up to 30 days. It is not immediate. Government Cloud customers and orgs that already opted out via support case are not affected by the new default.
The Three-Vendor Pattern: GitHub, Atlassian, Salesforce
Salesforce is not the only vendor to surface or tighten a default-on AI training posture this spring. The timing is conspicuous: three major platforms shifted the burden of privacy protection to buyers within roughly five weeks.
[GitHub announced on March 25]3 that Copilot Free, Pro, and Pro+ user interaction data would feed AI model training by default starting April 24, 2026, unless users manually opt out. Business and Enterprise tiers are exempt. Atlassian began rolling out its data contribution settings on April 16, with the policy taking effect on August 17, 2026. As [The Register reported]4, metadata collection is always-on for Free, Standard, and Premium tiers with no opt-out, while in-app data defaults to on for Free and Standard but off for Premium and Enterprise.
| Vendor | Effective Date | Scope | Opt-Out Mechanism | Tier Implications |
|---|---|---|---|---|
| Salesforce | Spring ‘26 toggle surfacing policy since ~2018 | Predictive Einstein global models | Self-service Setup toggle | Default-on; Government Cloud excluded |
| GitHub Copilot | April 24, 2026 | User interaction data for model training | Manual per-user opt-out | Free, Pro, Pro+ affected; Business, Enterprise exempt |
| Atlassian | August 17, 2026 | Metadata + in-app product data | Varies by tier and data type | Metadata always-on for Free/Standard/Premium; in-app data on for Free/Standard, off for Premium/Enterprise |
The pattern is consistent: lower-tier or long-standing customers get the default-on treatment, while enterprise buyers can negotiate or pay for exclusion. Tier-gated privacy is becoming a standard pricing lever.
What Compliance and Procurement Teams Should Do Now
Compliance teams should treat these three announcements as a single signal. The Atlassian pattern is not isolated. Vendors are retroactively surfacing AI training defaults that have existed in legal text for years, then offering self-service opt-outs as a concession while leaving the default unchanged. The practical effect is that the burden of data protection now sits with the buyer’s administrators, not the vendor’s legal team.
Procurement should add explicit AI training clauses to vendor checklists. Do not assume that a platform’s “enterprise” tier handles opt-outs uniformly. GitHub exempts Business and Enterprise from Copilot training entirely; Atlassian leaves metadata collection on for Premium; Salesforce excludes Government Cloud but leaves most commercial orgs opted in by default. The specifics vary enough that a blanket assumption will fail.
For teams already running Salesforce, the immediate task is to audit Setup → Einstein → Opt Out of Customer Data Access, document the decision, and calendar the 30-day deletion checkpoint. For teams evaluating new SaaS, the question is no longer whether a vendor uses customer data for AI training. It is whether the vendor makes exclusion a standard feature or a paid upgrade.
Frequently Asked Questions
Does the Salesforce toggle also opt out Slack workspace data?
No. Slack maintains a separate, email-based opt-out process for its own AI training data sharing. Organizations running both platforms must opt out of each independently — the Setup toggle has no effect on Slack data.
Which of the three vendors provides no opt-out for certain data types?
Atlassian. Metadata collection for Free, Standard, and Premium tiers remains always-on with no opt-out available after the August 17 effective date. Both Salesforce and GitHub offer at least a self-service or manual exclusion path for all affected tiers.
Is the Salesforce opt-out per-user or org-wide, and how does that compare to GitHub Copilot?
The Salesforce toggle is an org-level administrator setting — one decision covers every user. GitHub Copilot’s opt-out is per-user: each developer on Free, Pro, or Pro+ must change their own account settings. There is no admin-level kill switch for GitHub Copilot on affected tiers.
What happens to Einstein accuracy for orgs with limited historical data after opting out?
They fall back to models trained solely on their own records. For small or newer Salesforce deployments with few closed-won opportunities, lead-scoring accuracy can drop noticeably because the global cross-customer model is what compensates for sparse per-org data. The tradeoff is sharpest for orgs with under roughly two years of CRM history.