On some level, the media dogpile on Salesforce was inevitable.
Bloomberg reported that some of Salesforce’s polished Agentforce customer stories appeared to run ahead of production reality, including a University of Chicago Medicine example where a promotional video showed patients refilling prescriptions, booking appointments, and getting parking help through Agentforce while much of that experience was still not broadly live. Gizmodo followed with their own version of a similar critique: Salesforce Has an AI Vaporware Problem.
The usual machinery did its thing. Skeptics said told-you-so. Salesforce defenders said you don't understand enterprise software. Buyers asked what's actually live. LinkedIn turned the whole thing into a Marc Benioff effigy.
Fine. That's the internet.
But underneath the pile-on, there's a real story — one that has less to do with Salesforce specifically and more to do with the gap every enterprise AI vendor is about to walk into. It starts where most tales of enterprise woe start: with a demo.
Without context, enterprise AI cannot survive enterprise reality
A demo has every advantage in the world.
The path is clean. The data behaves like a good little child. The user asks the right question and the system returns the right answer. Nobody stops the whole production to ask whether the field feeds seven reports, three flows, and one integration.
A demo gets to show the dream sequence. A live enterprise system does not.
Not exactly controversial take, but an important thing to get out of the way first.
A live Salesforce org carries the full history of the business. From product launches to acquisitions, from new compliance rules to new pricing exceptions, from every buyer behavior change to every peculiar executive request: complexity builds and “we’ll fix it properly next quarter” leaves little residue behind.
Sometimes it leaves a field…
Sometimes it leaves a flow…
Sometimes it leaves a validation rule with archeological signposting like… TEMPFIX17DONOTDELETE_TRY2.
It’s fun poke fun at those field names. It really is.
But honestly, the humor kind of obscures reality itself. (The future is a helluva vantage point to judge folks from.)
In actually, this is business.
Enterprise complexity is not actually accidental mess sitting in the way of transformation. It is the natural result of scale, customization, growth, and most of all.. survival.
The real world is complicated, so the systems that mirror it become complicated too.
That is exactly where AI — or rather, generic generative AI — starts to really struggle.
Generic AI works best in clean, isolated environments where the system is easy to model. Real enterprises don’t exactly work that way. Complexity is unavoidable if you want to grow, and speed without context becomes risk.
That’s the real reason behind this Agentforce backlash: The market has seen what agents can do in controlled environments. And when met against the cold-hard light of reality, context mattered more than anybody ever thought.
Access is not understanding
At earnings, Salesforce reported $800 million in Agentforce ARR and 29,000 Agentforce deals in Q4 FY26. The company is also pushing toward a more agent-first platform through Headless 360, where Salesforce capabilities are exposed across APIs, MCP, and CLI.
Once agents can act across Salesforce without waiting for a human to click through the browser, the important question changes.
It's no longer: Can the agent actually do something?
It's: Does the agent understand what that action will affect?
That is the last mile. Not the interface. Not the prompt. Not the sparkle icon. Not the keynote moment when everyone claps because the agent booked the appointment, updated the record, or built the thing.
The last mile is pushing through the living complexity around the action:
What depends on this?
What fires when it runs?
Which team owns it?
What breaks downstream?
These aren't edge cases. This is the actual job — and it's the job AI inherits the moment it starts acting on enterprise systems.
Greenfield will not transform brownfield
A lot of AI stories assume greenfield conditions. Clean data. Obvious owners. Documented logic. Modern architecture.
That’s adorable.
In Org City, there are roads under roads. There are pipes nobody wants to touch. Buildings with broken foundations. Weird alleys that exist because of a decision made in 2017 by someone who now runs revenue operations at a direct competitor.
That's why "just rebuild it" almost never works the way people think it will. Rebuilding ignores the thing that makes the system valuable in the first place: the accumulated business context inside it.
The goal isn't to nix complexity. It's to slap a harness on complexity by making it legible — legible enough to move through it with both speed and governance.
That's the actual enterprise opportunity. Real-world context has to be part of the AI equation, or the equation just won’t balance out.
The agentic layer carries the context
As Salesforce becomes more agent-accessible, the org itself needs to become more agent-legible. That requires an agentic layer sitting across the systems where work actually happens — the metadata, automations, fields, rules, permissions, workflows, and dependencies that connect them.
Enterprise transformation is a cross-system ordeal. A Salesforce change can affect data pipelines. A ServiceNow process can depend on customer data. A Snowflake model shapes reporting, forecasting, and executive decision-making. The deeper the integrations, the more context matters.
That is why agentic AI cannot sit on top of the enterprise and speculate. It has to understand the system underneath. It has to know the blast radius before the change. It has to carry context from discovery into design, from design into build, from build into ongoing monitoring.
Otherwise, the agent is just moving fast through a dark room full of glass furniture.
Impressive... Still a terrible idea.
An agent that can create a field is interesting. An agent that can tell you where the field is used, what depends on it, what breaks downstream, who owns the process, whether the change violates standards, and how to deploy it safely — that's operational. That's scalable. That's worth the ARR slide.
"Vaporware" is a misdiagnosis. The right diagnosis is: the demo cleared the easy part, and the market is now reckoning with the hard part.
The race is the last mile
The media dogpile will pass. Too much meat on the bone for the news cycle to stay focused on one quarter's worth of customer-story scrutiny.
Salesforce will keep building. Buyers will keep experimenting. The AI narrative will keep swinging between "this changes everything" and "this kinda sucks," depending on the week, the stock price, and which executive said something dramatic on a podcast.
But the shift toward context will keep its steady march. Salesforce is becoming more programmable. Enterprise systems are becoming more agent-accessible. AI agents are moving from answering questions to taking action.
That future doesn't become less likely just because some demos outran production. The distance between a demoed and a deployed is filled with metadata, permissions, dependencies, workflows, integrations, governance, auditability, business logic, and change impact.
In other words: complexity.
The demo — so pure, so simple — was the easy part.
