Agentforce safety profile

What it does with your data

Salesforce says the outside AI models behind Agentforce do not train on your data. The vendor pages checked describe a zero data retention policy with third-party model providers such as OpenAI. Under that policy, those providers do not store your prompts or the responses that come back, and they do not use them to train their models or improve their products.

Two limits matter. First, these promises cover only third-party model providers. The pages checked do not say whether Salesforce’s own models train on customer data. Second, Salesforce itself keeps a record. Prompts, responses, trust signals and toxicity scores are logged and stored in Data 360 as an audit trail. The pages checked give no retention period for that log.

The pages checked do not split data handling by tier. They also do not say whether people at Salesforce review prompts or responses.

Settings to change first

The vendor pages checked do not describe a training opt-out setting or a menu path for one. For outside models, Salesforce says training does not happen at all. Salesforce’s own models are not addressed.

The pages do describe protections your Salesforce admin can set up. Before staff start using Agentforce, ask your admin to:

1. Review data masking. It detects sensitive data and masks it before a prompt goes to the AI model, and you choose what must and must not be masked. Put donor and client fields on the must-mask side.
2. Check role-based access and field-level security. Agentforce keeps these standard controls in place when it pulls in your records, so tighten them first.
3. Open the audit trail reports and dashboards so someone knows where prompts are logged.

The pages checked do not give exact menu paths for these settings.

Free and paid tiers

The vendor pages checked do not cover how data handling differs between tiers or plans. None of them describes a free version of Agentforce. None says that training, retention, masking or audit features change from one plan to another.

The admin features in the next section (data masking, secure data retrieval and the audit trail) are described without any plan named. That does not mean every plan includes them. It means the pages checked do not say.

If you are comparing quotes, ask Salesforce to state in writing which of these protections come with the plan you are buying. Also ask whether the zero data retention terms with outside model providers apply to every plan.

Admin controls

The Einstein Trust Layer is Salesforce’s name for the protections around Agentforce. The pages checked describe these admin controls and name no plan for any of them:

• Data masking. You pick which sensitive data is hidden before a prompt goes to the model.
• Secure data retrieval. It keeps Salesforce role-based controls and field-level security in place.
• An audit trail in Data 360, with pre-built reports and dashboards, plus feedback.

On certifications, the Salesforce compliance page for Agentforce & Einstein Platform lists PCI ASV Network Scan documents for Hyperforce regions under PCI DSS. That page points readers to an Audits and Certifications section in separate documentation. That documentation was not among the pages checked.

Using it with donor and client data

If your organization already keeps donor records in Salesforce, Agentforce may read those records directly. Your access settings matter as much as what staff type.

Before anyone uses it with real data:

• Confirm data masking covers donor names, contact details, gift amounts and any payment details.
• Tell staff never to paste in donor names with gift amounts, client case notes, health information, or card and bank details, even with masking on.
• Get answers in writing from Salesforce on the points the pages checked leave open:
  • whether Salesforce’s own models train on your data
  • where your data is stored and processed
  • how long the Data 360 audit log is kept
  • which plan includes which protections

Until those answers arrive, start with low-risk tasks that use no personal data.

Questions people ask

Does Agentforce train AI models on our donor data?

Salesforce says the third-party models it uses do not train on customer data, under zero data retention agreements. The pages checked do not say whether Salesforce’s own models do, so ask before you rely on it.

Does anything we type get stored?

According to Salesforce, outside model providers do not store prompts or responses. Salesforce itself logs prompts and responses in Data 360 as an audit trail. The pages checked give no retention period for that log.

Can we choose where our data is kept?

The vendor pages checked do not cover where data is stored or processed, or whether you can choose a region. The compliance page mentions Hyperforce regions only in its list of scan documents, so get the answer in writing.