SiliconANGLE theCUBE Sep 18, 2026
With George Gilbert, lead data and analytics analyst at theCUBE Research
Among Salesforce customers who have modeled or experienced the financial impact of using Salesforce without its own screen, 75% expect their Salesforce spending to rise rather than fall, on customer research from the firm Qualitate.
The intuitive reading is the opposite. If people stop working inside Salesforce's interface and work inside Claude or Slack instead, the seat-based software business should shrink. The argument in this episode is that the work moves off the screen and onto the platform underneath it.
"The point is the interface can move elsewhere, meaning the next purchase may or may not stay within Salesforce."
George Gilbert was inside the analyst program at Dreamforce 2026 and hosted on theCUBE there, and he built the system-of-intelligence framework this episode maps Salesforce onto. The customer data comes from 20 in-depth interviews Qualitate ran over five business days.
The full episode is covered here so you can skip it.
Here are the 11 insights that matter.
Key Takeaways
75% of customers who modeled headless access expect to spend more, not less, on Salesforce
90% are already working outside the Salesforce interface or plan to, using agents such as Claude Code, Codex or Cursor
A coding agent now generates the screen around the task, which Gilbert called the most useful announcement of the conference
Data 360 has to be installed before Agentforce delivers value, so the addressable market for Agentforce is the Data 360 install base
The Finn acquisition is the light alternative for customers who will not carry that installation
Customers want composability with Databricks and Snowflake, which they say do that job better than Salesforce
Most organizations interviewed are still in pilots or proofs of concept, spending 5% to 15% of their Salesforce budget on Agentforce
18 of the 20 said the Agentforce money was incremental, not shifted out of other Salesforce spend
Salesforce post-trained an Nvidia open-weight model into its own reasoning model for its environment
The unanswered pricing question is whose agent it is when a customer's own agent calls Salesforce and SAP through open interfaces
1. The Interface Moves Out
The episode's premise is that Salesforce's next growth may come from customers spending less time in its interface while agents do more of the work. The progression the show traced runs from command lines to graphical interfaces to browsers and mobile, and now to an agent generating an interface around the task inside Claude, Slack or another client. The starting point stops being the application's screen and becomes a customer outcome.
Generating a screen is not the same as understanding a business, which is where the framework comes in. A system of engagement connects people and agents, a system of intelligence supplies the business context, and a system of agency turns that understanding into action. Salesforce's own name for connecting those pieces is an enterprise AI harness: the data, the business knowledge, the workflows and the controls that let a model do useful work across systems.
The interface is not the same thing as the platform
The point is the interface can move elsewhere, meaning the next purchase may or may not stay within Salesforce.
A host
More consumption is not the same as more customer value, the show added, and pricing has to make sense as pilots scale.
2. The AI Stack Framework
Gilbert walked the framework from the bottom. Analytic data platforms are where customers have historically tried to aggregate data, and he said that still leaves silos, because each modeled cube is a silo of its own. Alongside them sit the traditional systems of record (Oracle, SAP, NetSuite and Salesforce itself), with a governance and metadata layer above that describes people, places and things rather than column and table names, so that both humans and agents can read it.
The layer above that is the one he cares about.
Where he says the value sits for the next 15 years
We keep saying it's the most important piece of real estate in enterprise software for the next 15 years
George Gilbert
At its simplest that layer is metrics and dimensions, the revenue or churn numbers familiar from business intelligence. A full system of intelligence also holds the processes: how quote-to-cash works, how a new employee is onboarded. The system of agency reads that map, reasons about the state of the business, makes a plan and then acts through the workflows beneath it, which become its tools.
What was new at Dreamforce, in his account, was a vendor connecting the idea of an agent harness, normally standalone software around a single agent, to that business model of the company. For Salesforce the harness is Data 360, Agent Fabric from MuleSoft and the Application 360 semantics.
The pitch that goes with it
What they're saying is you take whatever model you want and we make it useful and we allow it to compound in value over time.
George Gilbert
3. The Research Behind It
The customer data in the episode comes from Qualitate, a firm started about three years ago by Sara Kadakia, an early data scientist at ETR. Rather than running multiple-choice surveys, it uses agents to conduct interviews in which the participant answers in ordinary language, then categorizes the answers afterward with AI — an approach the show called schema on read, because the structure is imposed when the data is read rather than when it is collected.
The Salesforce study was 20 participants in half-hour to hour-long interviews conducted over five business days. Agentforce interest ran 75% interested against 25% uninterested. Organizations were allocating between 5% and 15% of their Salesforce spend to Agentforce.
And the budget question customers are asked first
There's no evidence yet anyway of agent force adoption detracting from the Salesforce budget.
A host
4. Headless 360 in Claude
Headless 360 is Salesforce without a user interface. The functionality is exposed through MCP servers and APIs, plus a set of plugins: a collection of skills inside Claude called Claude Force, and Slack's own Slackbot, so that each knows how to navigate Salesforce. What happens next is the part Gilbert emphasized.
A coding agent builds the screen for the data
So it's really a generative user interface with a coding agent behind the scenes.
George Gilbert
Years of trying to simplify the Salesforce interface so that users see only the functionality they need are answered, he said, by generating that interface on demand. He called it "probably the most immediately relevant and useful and impactful announcement from the entire conference", because Salesforce becomes a semantically harmonized store of customer data that can be served up in whatever context a user needs.
He drew the distinction between the two uses of the same plumbing carefully.
One serves a person, the other serves an agent
So one is custom UI generated by an agent to present to a human. The other is the data and workflows as context to ground the work of an agent operating on its own.
George Gilbert
The Qualitate numbers on the same slide: 90% of the organizations are already outside the Salesforce interface or planning to be, using other agents, whether Claude Code, Codex or Cursor. Developers have consolidated on a preferred model, which in some cases sits inside Salesforce. The verbatim comments ran both ways — "Our developers are pretty standardized on cloud code" against a customer saying they are a Microsoft shop using GitHub Copilot.
5. Slack as the Work Surface
Claude, in Gilbert's description, is a solo environment: one user working with an application, who might generate an artifact to share. Slack is where colleagues and other agents are, so the same headless plumbing with Slackbot generating the screen turns it into a work surface for people and agents together, and one he expects to widen access to Salesforce for new users.
The Slack-specific expectation from the research is more transaction volume and possible API consumption as sales representatives work outside the Salesforce interface. Awareness of Slack code, announced the week before, was thin but not absent: 60% had not heard of it, which the host pointed out means 40% had, days after the announcement. Seven of the respondents reported no interest because of entrenched alternatives.
Why the surface matters beyond Salesforce
But the point is we're in an era where the user interface to your backends is generated and that's the first step in essentially bridging all these silos that we've had.
George Gilbert
Right now the skills navigate Salesforce data. With MCP servers and APIs to other applications, Gilbert said, the same environment could generate a screen over SAP or NetSuite, though harmonizing that data is work that has not been done and the user is left bridging it. Customer reaction split: some found it attractive, others wanted proof on governance and said MCP servers are not fully baked, which he put down to regulated industries wanting more.
6. Composable Data Platforms
Salesforce's own stack slide carries no mention of Databricks or Snowflake, and the customer research says customers will not have it that way. They have standardized on one or both, and they want a customer data platform layered over it, because they think those firms do that job better than Salesforce does, inside or outside Salesforce.
Data 360 is built for that, with a zero-copy capability that points at the relevant data in Snowflake or Databricks, queries it dynamically and either returns or caches it. The obstacle Gilbert named is who Salesforce has to convince.
The credibility problem is on the technical side of the house
Now they still have work to do with earning credibility with the technology side of the house with customers
George Gilbert
Salesforce has been entrenched with the business side of a customer, while Databricks and Snowflake belong to the technology side, and composability means those buyers have to see components that work better layered on what they already run. He said Salesforce is still working on the positioning and on the time to value, and that this matters because of a dependency underneath Agentforce: Data 360 has to be in place first, which makes the addressable market for Agentforce the Data 360 install base. The research also showed hallucination problems, which he treated as a current state rather than a verdict.
7. Determinism Is Illusory
The host brought in work by their former colleague David Floyer on combining deterministic software with probabilistic software, and Floyer's argument that determinism inside most organizations is an illusion. A company may have deterministic processes in human resources, in its customer system, in logistics and in finance; bringing all four together is where it stops being deterministic.
Gilbert's addition was about what used to do the bridging. Procure-to-pay and order-to-cash were hardcoded into applications, but the joins between them were made by people, which is what made the end-to-end result non-deterministic — the same charge now leveled at agents. An agent needs the map more explicitly than a person does: it has to be told that customer 103 in one system is the same as Acme in another, and that revenue calculated one way maps to revenue calculated another.
The harness is what he says supplies that. The bridges in it do not dictate every step; they put guardrails and controls around what an agent must and must not do.
8. Why Agentforce Lagged
The host put it to Gilbert that practitioners find AI good at mundane data entry and much weaker at higher-level agency and workflows, and that Salesforce ought to have an advantage there because it holds the application logic and the process knowledge. Gilbert agreed and said he had expected Salesforce to be among the first big winners with agents.
What got in the way was the metadata. Even with harmonized definitions of customer data and the processes it moves through, there were gaps in the flow definitions describing how things actually worked. A person looking at a screen could bridge those gaps without noticing; removing the ambiguity for an agent took much more work.
Which is why the screen is the nearer-term win
It's much easier to generate a UI for a single person on a small set of back-end functions than it is for an agent to go to town across all these workflows.
George Gilbert
He said that makes the generative interface more relevant to Salesforce customers in the near term than end-to-end workflows executed by agents.
9. Role Agents and Models
The system-of-agency slide showed role-specific agents rather than a build-your-own toolkit: Piper for inbound pipeline generation, Hunter for outbound sales, KC Help for service and Page for IT service management and human resources. The intent, Gilbert said, is faster time to value — configure and tune, with enough back-end knowledge behind it, instead of custom-building each one in Agentforce.
Model choice was the other change. Customers can use Claude or OpenAI, and both Dario Amodei and Sam Altman appeared at the keynote. Salesforce also introduced a reasoning model of its own, taking an Nvidia open-weight model and post-training it on Salesforce so that it knows how to work in a Salesforce environment.
The economics of a smaller trained model
This gets back to the point that if you take a midsize model and you post-train it on an environment, it is much more cost efficient and performant in that environment.
George Gilbert
The pattern he expects is a family of models: the frontier for the hardest reasoning, a Salesforce-native model for everyday work. Around it sits a governance fabric for discovering, orchestrating, governing and observing agents, which he tied to how agents improve — reading back the reasoning traces and tool calls, seeing which led to good outcomes, then improving the prompt, the context or the model itself.
The new exhaust of the agent era
Now your agent traces are going to be what make your agents smarter.
George Gilbert
His comparison was the consumer internet, where user clicks trained the matching and recommendation engines.
10. Finn as the Light Option
The weakness of the enterprise harness is the installation it demands, and the Finn acquisition is the answer for customers who will not do it. For a customer that needs access to a couple of Salesforce objects and some limited data, whether from Salesforce, Databricks or Snowflake, Finn is the quick route to a pilot.
It removes the platform prerequisite
So now you don't need to install that whole enterprise harness.
George Gilbert
Gilbert spent time at the booth and described what is in it.
What the product actually is
It's really just a bunch of scripts, a natural language development environment and then just easy connections.
George Gilbert
He compared connecting it to attaching a tool to Claude or ChatGPT: a few API endpoints. Its competitive position is against Sierra and Decagon, the specialist agent companies, rather than against Salesforce's own platform.
11. What Salesforce Charges
The pricing discussion ran off a Boston Consulting Group maturity model, from usage based on resources, meaning token consumption, through agent-based pricing that works as a proxy for seats, then interactions, then outcomes such as help requests resolved, and finally a share of the financial upside. Marc Benioff has discussed all of them, including the gain-sharing end.
Gilbert's reading is that reality is still on the left of that chart, because the measurement is not in place and neither is agreement on sharing financial outcomes. What he did see for the first time was outcome pricing on jobs completed, conversations or resolutions, which is what Sierra offered from the start. The host's caution on enterprise license agreements is that they sound like all-you-can-eat in the marketing and carry restrictions procurement finds later.
The open question Gilbert raised is about ownership rather than the price sheet.
Today the assumption is that it is Salesforce's agent
You're charging because it's your agent on your platform.
George Gilbert
And that assumption may not hold
But what this does not reckon with is what if you have headless 360 from Salesforce and eventually say headless 360 from SAP and you're building your agent with data bricks or snowflake then it's their agent
George Gilbert
In that case Salesforce is one data feed among several, and what it can charge for is unsettled.
The spending data is where the episode ended. Flex credits drew interest, following the success CrowdStrike has had with flexible pricing. The 5% to 15% going to Agentforce was incremental in 18 of the 20 interviews rather than shifted from other Salesforce spend, 75% of those who had modeled headless access expected higher spending as a result, and 44% planned spending increases. The host's caveat is that procurement will watch whether new AI spending is offset later by lower legacy spending or retired features. Salesforce has guided to revenue moving out of the 50 billions into the 60 billions, which is the number the story eventually has to show up in.
Gilbert's verdict on the week
I think Salesforce has told a very compelling story and pulled together the pieces to show where the industry is going
George Gilbert
Bonus Insights
The host is a Qualitate customer as well as a subscriber to its data. Onboarded on the Monday, he had launched two surveys by speaking to the platform, one on ServiceNow control-tower customers and one on CoreWeave, and was halfway through the results by Friday.
The software apocalypse is not binary. The host's distinction is between a customer deeply embedded in Salesforce workflows, who is unlikely to leave, and one using it only as a system of record, whose data can be accessed and whose functionality can be recreated elsewhere. Parts of the software industry will feel that and parts will not.
Bret Taylor was Salesforce's chief product officer, then chief operating officer, then co-chief executive before co-founding Sierra, the company Finn is positioned against, and he is now prominent at OpenAI.
The Nvidia model Salesforce post-trained is Nemotron, which the host said Nvidia has pushed to shore up its open-source position, and which his sources describe as still a little behind.
Salesforce's earlier acquisitions were read as a collection of assets. The host's point was that Tableau, Slack and the rest looked like unrelated purchases at the time, while the engineering work that became Data 360 was going on underneath, and that this is what the agent era is now built on.
The episode's bottom line is that Salesforce has connected its data, semantics and workflows into something an agent can use from outside its own screen, that customers who have modeled the effect expect to pay it more rather than less for that, and that the unresolved questions are the Data 360 installation that gates Agentforce and who owns the agent when the customer builds it somewhere else.
Products, Companies & Tools Mentioned
Salesforce (Headless 360, Data 360, Agentforce, the Application 360 semantics and a new Salesforce-native reasoning model; guided to move from the 50 billions of revenue into the 60 billions)
Qualitate (The research firm behind every customer number in the episode, founded by Sara Kadakia, an early data scientist at ETR; conducts interviews with agents and categorizes the answers afterward)
Anthropic (Claude is the client where a coding agent generates a Salesforce interface, through a plugin set called Claude Force; Dario Amodei appeared at the keynote)
Slack (The collaboration surface where Slackbot generates the same interface; 60% of those interviewed had not heard of Slack code, so 40% had within days)
MuleSoft (Agent Fabric, one of the three pieces Salesforce now calls its agent harness)
Databricks and Snowflake (What customers have standardized on, and what they want Data 360 layered over rather than replaced by)
Nvidia (Supplied the open-weight model Salesforce post-trained; the host identified it as Nemotron)
OpenAI (A model choice inside Agentforce, with Sam Altman at the keynote; Codex is one of the coding agents customers have standardized on)
Cursor and GitHub Copilot (The other coding agents named in the customer verbatims, one of them by a self-described Microsoft shop)
Sierra and Decagon (The quick-to-deploy agent companies the Finn acquisition is meant to answer)
Finn (Salesforce's acquisition for customers who want a pilot without installing Data 360 — scripts, a natural-language development environment and a few API connections)
Oracle, SAP and NetSuite (The systems of record in the framework, and the ones a generated interface could eventually span)
CrowdStrike (The precedent for flexible credit pricing that Salesforce customers responded to)
Boston Consulting Group (Source of the pricing maturity model, from resource-based metering to sharing the financial upside)
Watch the full episode:
More George Gilbert interviews and podcast appearances
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: