Four independent accounts of Dreamforce 2026 landed this week, written from different angles for different readers, and they converge on the same unresolved question: Salesforce spent its highest-profile week of the year talking about trust, and none of the coverage found the proof that trust was meant to stand in for.
Dreamforce ran September 15 to 17 in San Francisco, drawing roughly 43,000 attendees across more than 1,600 breakout sessions and 50 keynotes. The theme, “Hey AI, Meet the #1 CRM,” set up a year-two argument for Agentforce: not whether AI agents can work inside a CRM, but whether enterprises can trust them to. This site’s own pre-event analysis argued that trust, not features, would be the gate Salesforce had to clear this year. Four trade outlets that covered the actual event each surfaced a different piece of what clearing that gate would require, and read together, their accounts describe a wider gap than any one of them names on its own.
What each outlet actually reported
MarketScale took the bluntest angle before the keynote even started: pre-event materials for an event with 1,600-plus sessions and 50-plus keynotes contained no customer-reported ROI figures, no adoption metrics, and no cost-to-run numbers for Agentforce. Its framing was procurement-first, aimed at the operations team that has to justify the spend after the keynote ends, not the audience in the room for it. If a rollout stalls in what the outlet called “pilot purgatory,” its argument went, that is because the keynote never handed buyers the audit trail needed to move past it.
Advertisement
UC Today covered the same week from the architecture side, walking through Salesforce’s Trusted Enterprise AI Harness, the six-part framework spanning context, agency, action, governance, security, and model choice that Salesforce positioned as the mechanism for enterprise-scale safety. UC Today’s account treats the Harness largely at face value, as Salesforce’s substantive answer to the trust question: agents can reason flexibly, but high-risk actions such as financial approvals and customer entitlements still run through deterministic, rule-bound controls rather than probabilistic guesses.
Diginomica, in Jon Reed’s weekly column written against the Dreamforce backdrop, pushed the trust question further than either. His argument: buyer trust in an AI vendor is not the same thing as trust in AI itself, and one of its two biggest components is transparency, specifically in pricing, architecture, and how data moves. Most vendor AI pricing, he wrote, is “a gooey (and perpetually shifting mix) of embedded features, add-on credits, consumption pricing shifts, and outcome flirtations.” That is a sharper, more specific bar than either of the other two accounts sets, and it is one no vendor’s Dreamforce week, Salesforce’s included, actually cleared.
Where the accounts split, and where they agree
The disagreement across these three is real, not cosmetic. UC Today reports the Harness as a credible, substantive response to enterprise trust concerns, treating deterministic execution for sensitive actions as the kind of control that makes wider agent deployment safe to approve. Diginomica’s piece, read against that, functions almost as a rebuttal: governance controls address whether an agent behaves inside the rules it is given, not whether a buyer can see what they are paying for it or what it actually returns. MarketScale sits between the two, less interested in governance mechanics than in the plain absence of a number, any number, that would let a buyer weigh cost against outcome before signing.
Where all three agree, and this is the finding the week’s roundup coverage never states outright on its own, is in what none of them found. Not one of the four accounts, across a week when Salesforce put trust at the center of every keynote, cites a customer-reported adoption rate, a production cost figure, or an independently measured return for Agentforce at scale. UC Today’s optimism is optimism about governance, not proof of value. MarketScale’s skepticism is specifically about the missing number. Diginomica’s transparency argument names the exact category, pricing, that the other two accounts each pass over without addressing directly.
The half of trust Salesforce actually shipped
Salesforce’s own Trusted Enterprise AI Harness announcement, published September 10 ahead of Dreamforce and referenced across nearly every keynote that followed, is more specific than any of the four trade accounts individually credits it for. “The Agentic Enterprise won’t be defined by which model a company chooses,” said Rohan Kumar, President, Chief Platform and Engineering Officer at Salesforce, arguing that proprietary customer context, not model selection, is what will differentiate one enterprise AI deployment from another. Rocket Mortgage’s CTO, Shawn Malhotra, framed the appeal in composability terms: the harness let his team avoid betting the company’s AI future on a single closed stack, keeping the freedom to swap models as the market shifts.
Read against the trade coverage, that primary document answers exactly one of the two questions “trust” is being asked to carry. It is a genuinely detailed operational answer: shared context across agents and models, deterministic execution for high-risk actions, a control plane spanning CRM, ERP, contracts, and policy systems that previously had to be stitched together by hand. It says nothing about unit economics. It contains no adoption figures, no cost-per-agent data, no independently verified error or reliability rate for any of the six capabilities it describes. Salesforce built the governance half of trust in real technical depth. It then let the resulting keynote narrative, and a week of largely admiring trade coverage, stand in for the financial half, which nobody, including Salesforce’s own materials, actually supplied. This site made a version of that same point about Koa’s self-reported, vendor-graded benchmark before Dreamforce began; the same pattern held across the entire event, not just one product line.
What it means for the RevOps and sales leader
The practical read for anyone evaluating Agentforce, or any competing agent platform sold on the same trust language, is to price the two halves of “trust” separately in a vendor conversation, and to refuse to let a strong answer on one substitute for the other. Ask for the governance receipts: which specific actions run deterministically rather than probabilistically, what gets logged for every agent decision, and which actions still require a human sign-off before they execute. Then, as a completely separate line item, ask for the financial receipts: adoption rate among customers in a comparable industry and deployment size, cost per resolved task or per qualified lead at production volume, and a benchmark methodology that a third party, not the vendor, designed and graded. A convincing answer to the first question is not evidence for the second, no matter how smoothly a keynote sequences them back to back.
Buyers who cannot get a straight answer to the second question should treat that absence as information in its own right, not as a reason to wait for next year’s keynote to finally supply it. The gap has now been visible, independently, from four different vantage points, for an entire product cycle. A vendor that has had a year to produce the number and still has not is telling a buyer something, whether or not it says so on stage.
That does not make the governance work Salesforce shipped irrelevant. A control plane that can log every agent action, enforce deterministic execution on financial approvals, and share context across systems that previously required custom integration work is a real engineering achievement, and RevOps teams evaluating any agent platform, not just Agentforce, should expect a comparable baseline before they let one touch a live pipeline. The mistake is treating that baseline as the whole answer. Trust, as all four accounts describe it this week, still has a financial half that remains unmeasured, and a buyer who accepts the governance demo as a substitute for it is accepting exactly the substitution the coverage, read individually, never quite catches.
Source: Salesforce
