Last week, KeyBanc analyst Jackson Ader downgraded Salesforce. Bernstein followed the same day. The headline is a stock rating. But the substance is something more important, and more useful to anyone running a Salesforce org trying to make AI actually work.
Customer feedback, per KeyBanc, was consistent in two areas: customers' data is not organized to do meaningful AI work, and Agentforce as a product is not ready.
Salesforce's answer: Why it's half right
Salesforce will respond to this downgrade exactly as they have been responding for the past two years: buy Data 360. Unify your data. Give your agents a trusted, clean, connected foundation to reason against.
That's the right answer to part of the problem.
Data 360, what used to be called Data Cloud, rebranded at Dreamforce 2025 is a valuable capability. It ingests data from across your systems, harmonizes it into unified customer profiles, and exposes that data to Agentforce as grounded context. When Benioff said "without clean, connected, trusted data there is no intelligence, only hallucination," he was right. And Data 360 is built to solve that.
But here's what the downgrade is actually describing and what Data 360 does not solve…
Two meanings of "metadata." Two different problems.
When Salesforce says Data 360 makes your data AI-ready, they mean it unifies your data assets: customer records, profiles, segments, semantic definitions. That is data-layer readiness.
What agents actually execute against is something different: the operational logic of your Salesforce org. The Flows. The Apex triggers. The validation rules. The field dependencies. The cross-system automation behavior. The live state of an org that has accumulated years of customizations, integrations, and changes that nobody fully mapped.
Salesforce's own Atlas Reasoning Engine ( the "brain" of Agentforce) confirms this. Salesforce says agents "reason over metadata" and that "every field, label, entry, and automation built on the platform is tagged with relevant metadata that Agentforce can read and understand." That is operational metadata. And Data 360 does not govern it. Salesforce's own Metadata API Developer Guide defines this layer as "the metadata model, not the data itself" — Apex classes, triggers, Flows, validation rules, workflow rules, approval processes.
These are two different meanings of the word "metadata." They describe two different layers of the stack. And they require two different solutions.
Data 360 governs what data exists. It does not govern how the systems that hold it actually behave.
What agents actually inherit
When an AI agent is deployed on top of a real enterprise Salesforce org, it doesn't just read unified customer profiles. It executes actions. And every action fires a chain reaction: a field update calls a Flow, the Flow triggers Apex, Apex runs a validation rule, the validation rule fails silently, and somewhere downstream (in Salesforce, in Snowflake, in ServiceNow) something breaks.
An agent doesn't compensate for that the way a careful admin would. It acts on what it finds. If what it finds is ungoverned, it produces what looks like a confident, correct action… but is actually a fast, autonomous mistake built on a stale mental model of the org.
This is not a data quality problem. This is a system-logic readiness problem. And it's the second half of what KeyBanc's customer feedback is actually describing.
Gartner projects that more than 40% of agentic AI projects will be canceled by end of 2027. Forrester's data suggests 88% of agent pilots never reach production. Deloitte found only 11% of organizations are actively using agents in production. The data-unification narrative doesn't explain those numbers. The ungoverned-logic narrative does.
The gap no one is governing
The most important thing I've learned across a decade of watching enterprise AI stall — at Salesforce, at Coastal, and now at Sweep — is that the problem is almost never the technology. The models are capable. The platforms are powerful. The teams are smart.
The problem is that metadata complexity has outpaced what any team can manually understand, manage, and keep current. Every new integration widens the gap. Every acquisition adds a layer. Every automation creates consequences somewhere downstream that nobody mapped. And when AI agents are deployed on top of that environment, they inherit all of it.
Data 360 unifies the data. That's necessary. But the ungoverned logic underneath — the operational metadata that agents actually reason and act against — stays ungoverned.
That is the layer Sweep governs. Not as a one-time audit. As persistent infrastructure that maintains a continuously current map of how every field, flow, dependency, and automation actually behaves — across Salesforce orgs and the connected systems they touch.
Discovery is the entry point. Understanding is the product.
If your Agentforce rollout is stalling, you may have already solved the data problem. The answer might be in the layer underneath it… and Data 360 wasn't built to go there.
