AWS and Salesforce's Zero Copy: What It Reduces and What You Still Need to Check

One common way to let AI work with business data is to copy that data into a separate environment. This can speed up searches and responses.

But the more copies there are, the more work it takes to confirm that the viewing and editing rules set in the original system still hold wherever the data now sits.

The expanded AWS–Salesforce partnership announced on September 15, 2026 includes new ways to reduce the need for data replication. This article looks at the part of that announcement that concerns where data sits and who may use it, and sets out what to check when you are considering adoption.

What was announced

The central change is that Salesforce Data 360 has widened the scope of its “zero copy” connections. Data 360 is a data platform for combining data from inside and outside a company so it can be used in business processes and by AI.

According to the announcement, zero copy now covers Apache Iceberg tables managed by AWS Glue, Apache Iceberg tables on S3, Amazon Aurora, Amazon RDS, and SageMaker Lakehouse (Amazon Redshift and Amazon S3 were already supported). We refer to these below as the databases and storage services on AWS.

The announcement also covers an integration between Amazon Quick and Salesforce. Amazon Quick is an AWS AI assistant for business users, and the integration is described as letting people look up and update information held in Salesforce CRM — the system that manages customer records, opportunities, and support cases — from within Quick. The connection uses MCP (Model Context Protocol), a shared standard that lets AI applications work with outside data and functions.

The announcement further includes an integration that brings AWS AI agents into Slack, which we do not cover here.

What zero copy reduces, and what remains

Salesforce describes zero copy as a way to read and use data held elsewhere without first replicating it into Data 360. At the same time, there is a configuration that caches external data temporarily so that processing runs faster. The term “zero copy” alone does not tell you whether the system accessing the data retains any of it.

If you can reduce the number of separate places data is copied to, you may be able to reduce the work needed to synchronize those copies and maintain their access controls. We see this as a potential benefit. We also recommend checking the access controls on each connection and how retrieved data is handled.

When you adopt it, check what data is stored and where under the chosen connection method, and under what conditions it is refreshed or deleted.

Maintaining access controls on copied data

Why do copies become a permissions question in the first place?

Say customer information is visible only to the account owner and their manager, and that same information is copied, unchanged, into another database. If the same restriction is required there but the copy is configured so the whole department can see it, the intended access control is not being maintained. In this example, the problem is that the required access controls were not applied to the copied data.

If the copy carries the same restriction, and permission changes are reflected in it, the existence of the copy is not itself the issue. A direct connection to the original data can also allow unintended access if it has broader permissions than the task requires.

Check whether permission changes are reflected in the copy, as well as changes to the data itself. If you revoke someone’s access in the original system but a separately granted permission on the copy remains in force, that person may still be able to access the copied data.

The more places data sits, the more of these checks there are.

Check who can do what, connection by connection

Using data requires both confirming who is making the request and deciding what that person or system is allowed to do. The first is authentication; the second is authorization. OWASP’s Authorization Cheat Sheet also separates the two, noting that a user who has been authenticated is often not authorized to reach every resource or perform every action that is technically possible in a system.

Seen through that distinction, the two integrations in this announcement call for separate checks.

The Amazon Quick listing on Salesforce’s app marketplace states that every operation runs under the user’s own Salesforce permissions and sharing rules, and that write operations are presented to the user for explicit review and approval before they are carried out.

The authentication and access-control settings for Data 360’s AWS connections depend on the connected service and the connection method. Check the relevant connector’s setup documentation.

In doing so, we would not judge safety by the type of account alone. Even when something runs under a person’s own account, not every permission that person holds is necessarily required for the task. A system account can also be restricted to specific resources.

We focus on which account is used, what actions are permitted, and where those limits are enforced.

Five things to check before adoption

The following are the items MIF suggests checking. Make them specific to your own operations, the data you handle, and the rules that apply to you.

What to check	What that means in practice
Whose account and permissions are used?	Does access use the user’s own account or a system account? Where are permissions to access data and perform actions checked and enforced?
What is allowed, and when is approval required?	Is access read-only, or are record creation and updates also allowed? Can you specify which operations require approval and who approves them?
Can actions and approvals be traced?	Can records show when an action occurred, who requested it, and which process accessed or changed which data? Can approvals be linked to the resulting actions?
How are permission changes and stored data handled?	When does revoking access take effect? Where are retrieved results and cached data stored, and when are they refreshed or deleted?
How can access or execution be stopped?	Can access or execution be suspended for a particular user, agent, or connection? Who decides when to suspend and resume it?

The fourth item is there because, even with zero copy, some methods involve temporary storage. Stopping further access to the source data is separate from managing results and cached data that have already been stored.

After reviewing the published specifications, test the integration in your own environment. Check access for users with different permissions, behavior after permission changes, and whether the logs let you trace what happened.

Conclusion: plan data storage and access together

Zero copy is one option for reducing the copying and synchronization needed to use data. When evaluating it, also check where retrieved data is stored and which viewing and editing permissions apply.

At MIF, we recommend defining what data AI may access, which actions it may take, and when human approval is required as part of evaluating product features and accuracy.

Review the published specifications for each integration and check how the stated controls apply to your configuration. Then test behavior after permission changes, check that actions can be traced in the logs, and confirm how to suspend access or execution when a problem occurs.

Use the five checks above as a starting point.

Sources and Verification

This article draws on the following material, reviewed on September 22, 2026.

The announcement published by Amazon’s press center on September 15, 2026, “AWS and Salesforce Put CRM Data, AI Agents, and Model Choice into the Tools Teams Use Every Day.”

Salesforce’s explanatory page on zero copy connectivity, “Why Data 360’s Zero Copy Makes Business and Technical Sense.”

The “Amazon Quick” product listing on Salesforce’s app marketplace.

OWASP Cheat Sheet Series, “Authorization Cheat Sheet.”

Product information in this article is drawn from vendor-published materials rather than independent test results.

Availability reflects what the announcement stated at the time; the announcement notes that availability can vary by region and is governed by customer agreements. Feature details and timing may change.

The framing of the issues and the list of items to check in this article are MIF’s own analysis, based on these published materials.