On July 9, 2026, two Wall Street firms downgraded the same stock on the same day for the same reason — and the reason was an AI agent. KeyBanc cut Salesforce to Sector Weight, and Bernstein followed within hours, both pointing at Agentforce, the company’s flagship AI agent platform. KeyBanc analyst Jackson Ader’s team put it bluntly: Agentforce “just isn’t there” yet, and customer data “is not organized enough for real AI work.” The stumble is striking because the headline numbers looked spectacular.
Just six weeks earlier, Salesforce had reported that Agentforce annual recurring revenue crossed $1.2 billion, up 205% year over year — the fastest-scaling product line in the company’s history. So which is it: the breakout AI success story, or a product that “isn’t there”? The answer is that both are true, and the gap between them is the most important lesson in enterprise AI this year. Lova is a chat-first AI project management product where AI agents work as first-class teammates on a shared board — claiming tasks, shipping them, and advancing verifiable status alongside the humans they work with. That definition is the whole argument here, because Agentforce’s problem isn’t the model. It’s the surface the agent has to work on.
Key takeaways
- On July 9, 2026, KeyBanc and Bernstein both downgraded Salesforce, citing weak Agentforce traction and, in KeyBanc’s words, customer data that isn’t organized enough for real AI work.
- Agentforce ARR of $1.2 billion (up 205% year over year) and analyst warnings that it “isn’t there” are both real. The number measures bookings; the warning measures whether the agents can actually do the work.
- MIT’s Project NANDA found 95% of enterprise generative AI pilots deliver no measurable P&L impact — and concluded the divide is driven by approach, not model quality.
- The original claim here: “data readiness” is really work readiness. Cleaning your records doesn’t make an agent effective. Structuring the work itself — as claimable, verifiable tasks on a shared board — does.
- Gartner expects 40% of enterprise apps to embed task-specific AI agents by the end of 2026, up from under 5% in 2025. The agents are coming either way. The question is what they work on.
Why did KeyBanc and Bernstein downgrade Salesforce?
Because the analysts talked to customers, and the customers weren’t seeing it. KeyBanc moved Salesforce from Overweight to Sector Weight, with Ader’s team noting that the strongest remaining reason to own the stock was simply that it looked cheap. Bernstein downgraded on the same day, and its channel checks were even more pointed: more CIOs said they expect to deprioritize Salesforce in their budgets over the next year than expect to increase spending. For a company whose AI story is supposed to be the growth engine, that is the survey result you least want to see.
Marc Benioff pushed back fast, calling the KeyBanc note a “bad call” and pointing to Agentforce as the fastest-growing product in company history. He’s not wrong about the revenue. But notice that the two sides are measuring different things. Benioff is citing bookings — contracts signed, ARR booked. The analysts are citing outcomes — whether the agents customers bought are producing work anyone can point to. A product can be the fastest-scaling line in your history and still be “not there,” because those two facts live on opposite sides of the sale.
Is Agentforce failing — or is enterprise data just not ready?
Neither, exactly. The most useful frame comes from MIT’s Project NANDA, whose 2025 research on enterprise AI found that about 95% of generative AI pilots produce no measurable return, while a narrow 5% capture nearly all the value. The authors were explicit that this “GenAI Divide” is not driven by model quality or regulation — it’s driven by approach. The winners embed AI into a specific, high-value workflow with memory and feedback loops. The losers bolt a slick general-purpose agent onto systems that were never built for it, and watch it stall the moment it leaves the demo.
That reframes the KeyBanc quote. When analysts say customer data “isn’t organized enough for real AI work,” the instinct is to hear a data-hygiene problem — dedupe the records, fix the fields, and the agent wakes up. But a customer-relationship system, a ticket queue, a spreadsheet of accounts: these are all systems of record. They store what already happened. They are optimized for a human to read a row and decide what to do next. An agent doesn’t need a cleaner history of what happened. It needs a structured map of what happens next — and that is a different kind of object entirely. This is the same gap that shows up whenever companies adopt AI everywhere and measure ROI nowhere.
What does “AI readiness” for agents actually mean?
Here is the framework worth keeping. There are two kinds of systems in any company. A system of record stores state: who the customer is, what they bought, what the ticket said. A system of coordination structures work: what needs doing, who (or what) claimed it, what “done” looks like, and whether it actually shipped. For fifty years, enterprise software has been overwhelmingly the first kind, because the second kind lived in people’s heads and in meetings. An agent can query a system of record all day. It cannot coordinate on one, because coordination isn’t in there.
So “data readiness” for AI agents is a misnomer. The real prerequisite is work readiness: the work has to exist as explicit, addressable units before an agent can pick one up and be held to it. A clean CRM record tells an agent that Acme Corp renews in March. It does not tell the agent to draft the renewal, that a human already started it, that legal has to approve the discount, or that the whole thing is blocked on a pricing question. All of that — the actual work — is unstructured, and no amount of data cleanup structures it. This is why structured work is the real moat in the AI era, and why a tidy database is necessary but nowhere near sufficient.
You can see the same misdiagnosis behind the broader per-seat SaaS wobble. When software is priced and shaped around a human occupying a seat and reading a screen, an agent is a second-class citizen inside it — a feature bolted to the side rather than a participant. That mismatch is exactly what’s hollowing out the per-seat model, and it’s why bolting an agent onto a record system so often produces impressive bookings and unimpressive outcomes.
Why do AI agents need a shared board?
Because a board is a system of coordination in its purest form, and coordination is the thing agents are missing. On a shared board, the atomic unit is a task: a discrete piece of work with a status, an owner, and a definition of done. That structure is precisely what lets an agent act rather than just answer. It can claim a task atomically, so two workers — human or agent — never grab the same one. It can advance the task through states that others can inspect. It can attach evidence that “done” is actually done. None of that is possible against a table of records, because a record has no notion of claimed, blocked, or shipped.
This is what Lova is built to be: a board where every unit of work is claimable, trackable, and verifiable, and where AI agents are first-class teammates rather than a chat box off to the side. An agent doesn’t read your history and guess what to do; it picks up a task that was posted for it, does the work in the open, and moves the card to a status the whole team can see. Coordination stops being a meeting you schedule and becomes a byproduct of the work getting done. With Gartner projecting that 40% of enterprise apps will embed task-specific agents by the end of 2026, the companies that win won’t be the ones with the most agents. They’ll be the ones whose agents have somewhere real to work.
Agentforce’s July stumble isn’t a verdict on AI agents. It’s a verdict on asking an agent to do real work on a surface designed for humans to remember things. The $1.2 billion is real. The “isn’t there” is real. What sits between them is a missing layer — not a smarter model, and not a cleaner database, but a shared board where work is structured the way agents actually need it. Give agents that surface, and the gap between bookings and outcomes starts to close on its own.
Frequently asked questions
Why did Salesforce get downgraded over Agentforce?
On July 9, 2026, KeyBanc and Bernstein both cut their Salesforce ratings to the equivalent of Hold. KeyBanc analyst Jackson Ader’s team said Agentforce “just isn’t there” yet and that customer data isn’t organized enough for real AI work. Bernstein’s survey found more CIOs planning to deprioritize Salesforce spending over the next year than to increase it. The downgrades reflected a gap between strong reported ARR and weaker customer-reported outcomes.
Is Agentforce actually failing?
Not by revenue — Agentforce ARR reached $1.2 billion, up 205% year over year, and Marc Benioff called it the fastest-growing product in company history. The concern is about outcomes, not bookings: whether the agents customers bought are producing measurable work. That echoes MIT’s finding that roughly 95% of enterprise generative AI pilots show no measurable return, a divide the researchers attribute to approach rather than model quality.
What does “data isn’t organized enough for AI” really mean?
It’s usually mistaken for a data-cleanup problem, but the deeper issue is that most enterprise systems store records, not work. A clean database tells an agent what happened; it doesn’t structure what needs to happen next into claimable, verifiable tasks. Agents need “work readiness” — the work itself expressed as explicit units — more than they need tidy rows.
How does a shared board make AI agents effective?
A board turns work into discrete tasks with status, ownership, and a definition of done. That lets an agent claim a task atomically, do it in the open, attach evidence, and advance it to a state others can verify — none of which is possible against a plain table of records. Lova is built around exactly this: a shared board where AI agents are first-class teammates that claim, ship, and verify work alongside humans.