Product1 publisher3 min readPublished
Salesforce's AIforce says agents inherit existing permissions to keep access secure and governed
The Dreamforce release says every AIforce request runs on the permissions already configured in Salesforce and every action routes back through it. The reading it enables spans hundreds of records at once, and the page layout is gone.
The Product Desk · Product desk
What happened
- Salesforce unveiled AIforce at Dreamforce on September 16, 2026, calling it a live interface layer that carries Salesforce data, workflows, logic, permissions and governance to wherever people and agents work.
- The release names Claude, Slack and Lightning as places where people can get Salesforce work done without opening the Salesforce app.
- The release says agents can read across hundreds of records at once and pull in connected systems, at a speed and scale it describes as beyond any human.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure An over-broad profile that nobody could exploit through a list view becomes reachable in one prompt by an employee who has never opened a dashboard.
- decision With the screen out of the path and the permission set still in it, admins have to decide whether to re-audit sharing rules before enabling the connection or after the first complaint.
- constraint Without published detail on logging, an admin asked which questions were answered in a third-party assistant has nothing in the release to plan an answer around.
- precedent Once anyone can compose a view by describing it, the layout review that IT used to own stops being the place where data exposure is decided.
A rep who wants every closed-lost reason in a territory for the past six months filters a list view and reads twenty rows, or asks ops for an export and waits a day. Sharing rules are rarely what stops them at twenty rows. The page layout does, and so does the tedium of paging through it.
Salesforce says AIforce removes the tedium, with agents reading across hundreds of records at once and pulling in connected systems [6]. The company also says the permission check stays where it was: "Because every request runs on existing permissions and business rules, every agent sees only what the person asking can see, and every action routes back through Salesforce," according to the release [7].
Salesforce names the intended user. Employees who have never opened a dashboard can ask questions, update records and trigger workflows grounded in Salesforce data and business logic [5]. Until now, the release says, most people could reach that knowledge only by opening the app and navigating the UI, which limited who could use Salesforce and how much they could see at any one time [3]. "As a deeply mobile-first region where people work in conversational channels, ASEAN businesses can unlock the full potential of AI through AIforce," said Paul Carvouni, senior vice president and general manager for ASEAN [11].
Admins now have to be right about the sharing model itself. A profile that can technically read every account in a region was tolerable while the only way out was a list view and a manual export. Read the same profile at the speed the release describes and the tolerance disappears.
"AI is creating an interface revolution," said Marc Benioff, chair and chief executive [9]. The release lists Zero Data Retention as a property of the layer, saying business data is used to answer the question at hand and is not retained [13]. The release is silent on what an admin sees in the logs afterwards [16].
The second admin problem is authorship. Anyone can compose a live interface by describing it, inside platforms including Slack, Claude and Coworker, with MCPs, APIs, plug-ins and skills for teams that want to go further [8][14]. The admin who used to answer for what a page layout exposed now answers for views built outside IT.
A way to sort the work before anyone connects an assistant: for each object, write down two numbers. How many records can this profile technically read, and how many could a person realistically pull through the UI in a working day. Where the first number is large and the second is small, the interface was the limit and the sharing rule is untested. Those objects are where a re-audit belongs first.
AIforce is the fourth layer Salesforce lists in its Agentic Enterprise architecture, after Data 360, Customer 360 and Agentforce [12][15].
What to watch
- Pricing and general availability for AIforce, neither of which the September 16 release states.
- Whether Salesforce ships admin-side scoping and a log of requests issued from third-party assistants.
- How the Zero Data Retention claim applies to transcripts held by the assistant, rather than to Salesforce business data.