Data 360
Data 360 Notebook AI: research documents and verify the answers
Use Notebook AI for research across documents and enterprise content. Compare it with agent grounding, check access, and measure consumption before rollout.
An account plan can start with an RFP, meeting notes, and a folder of conflicting product documents. Notebook AI gives a Salesforce user a place to investigate that material and produce a report with sources to check.
Use it when a person needs to research a question, review the evidence, and decide what to share. Keep your customer-facing Agentforce grounding design as a separate responsibility. A useful research report still needs review before it becomes an approved knowledge article or an answer sent to a customer.
A notebook starts with a research question
Salesforce describes Notebook AI as a workspace for personal, enterprise, and public content, available inside Data 360 and through a web application. A notebook is private to its creator by default. Its documented examples include researching recurring support issues and preparing an RFP response.
Deep Research can combine uploaded documents, CRM records, enterprise search results, and web sources into a report with citations. That is the product capability. The quality check belongs to the person using the report.
For an RFP, start with a specific question such as which requested capabilities the approved product documentation supports. Name the documents that count as authoritative, their dates, and any exclusions. A broad request to write the whole proposal makes unsupported assumptions harder to spot.
Read the cited passage behind each material claim. A citation may point to a relevant document without proving the sentence beside it. Record conflicts instead of allowing a fluent summary to settle them. If two sources disagree about availability, assign that question to the product owner before copying the answer into the proposal.
Choose the document tool by its output
Salesforce has several ways to work with documents. The useful distinction is what you need to produce and who will rely on it afterward.
| Requirement | Starting point |
|---|---|
| Research a question and review a written analysis | Notebook AI |
| Supply approved material to an agent during a conversation | Agentforce Data Library or a configured retriever |
| Extract fields into a structured processing workflow | Document AI |
| Diagnose whether retrieval finds the right passages | Retriever quality evaluation |
For a support investigation, a notebook could help compare incident notes and product documentation. The approved resolution can then become maintained knowledge through your normal publishing process. Treat that publication as a deliberate handoff, with an owner and review date.
Don’t assume uploading a file to a notebook automatically makes it available to every agent or colleague. Test the actual access and sharing behavior for the intended user, and document the separate process for promoting approved content into your agent’s knowledge source.
Access and billing need their own pilot
Salesforce’s setup guidance calls for enabling Notebook AI for data spaces and granting user access through custom permission sets. In a multi-org design, it identifies setup in both home and companion orgs. Web access requires authorization of the application connection.
Start with a small group and approved sample documents. Test with an ordinary user’s permissions. Include a source that the user should see and one they should not, then check both retrieval and the reports shared afterward.
Human review doesn’t make untrusted uploads harmless. Public pages and uploaded documents can contain incorrect claims or instructions aimed at the model. Treat their text as evidence to examine, and avoid treating a generated recommendation as permission to change CRM data.
The Notebook AI billing reference describes consumption across Data 360 and Agentforce usage types. It also notes that some usage types aren’t metered under particular licenses or editions. Check the active Digital Wallet cards and your contract instead of assuming every question has one universal price.
Measure a repeatable workload: the same document set, a defined research task, and the resulting reports. Record consumption alongside the time spent checking and correcting the output. The result should show the cost of producing a reviewed artifact, rather than just the cost of generating text.
Make the review criteria concrete
Before the pilot, decide what would make a report usable. For an RFP response, require a supporting source for each capability claim, a visible unresolved section for conflicting evidence, and a reviewer responsible for commitments made to the customer.
For a support article, test whether another engineer can reproduce the proposed resolution from the cited material. An answer that sounds plausible but cannot be verified needs more investigation.
Keep the original question, the source inventory, and the approved report together. When a product changes, the next reviewer should be able to see which statements need another check without repeating the entire investigation.
