TL;DR
The story this cycle is the gap between the demo and the ledger: Salesforce put 30,000 customers and its first testable ROI numbers on stage at Dreamforce, while three new surveys show measurable value actually getting harder to demonstrate as pilots hit production costs. Meanwhile, governance is crystallizing around a "bounded autonomy" consensus — no one, anywhere, is running agents with unrestricted autonomy.
The ROI gap widens even as adoption becomes universal
The adoption funnel is filling from the top while value metrics drain from the bottom. Deloitte's survey of 3,200+ executives found 75% of organizations already use some agentic AI and 95% expect to within two years — but only 2% use it extensively and just 1% have it fully integrated into operations (CIO.com via enterprise-agent-adoption-production-gap-2026). The sharpest reversal: Accenture's July survey found the share of companies reporting widespread, sustained AI value fell from 32% to 23% as deployments moved into production and real costs surfaced.
"23% of companies report widespread and sustained business value from AI, which is down from 32% earlier this year." — enterprise-agent-adoption-production-gap-2026
The economics explain the slide: Gartner expects inference costs per agentic workflow to rise more than fivefold through 2028, and McKinsey finds spending rises nearly fourfold moving from isolated use cases to widespread adoption enterprise-agent-adoption-production-gap-2026. This extends last cycle's Dreamforce critique — the event had 1,600+ agent sessions and no customer ROI figures — except this cycle Salesforce finally supplied some (see below), and they arrived pre-discounted by a Bloomberg investigation finding Agentforce outcomes "materially miss what Salesforce's marketing promised" (TechTimes via agentic-ai-market-size-growth-2026
).
What to watch: Whether the Bloomberg investigation's named customers and methodology surface in full, and whether the value-demonstration share keeps falling as more pilots convert to production.
Dreamforce delivered numbers — with asterisks
The thread opened last cycle on Dreamforce ROI disclosures resolved, and the answer is: partial disclosure, real numbers, published caveats. Fulton Bank was the first customer proof point specific enough to test — 80,000 hours saved, $389 million in attributed loans and deposits, and a four-day reduction in complaint resolution time — while Siemens described lead-qualification agents handling ~2,500 unqualified leads per month across ~18,000 sellers in 132 countries (beri.net via agentic-ai-market-size-growth-2026).
"Treat the $389 million with the caution the word 'attributed' deserves — attribution in a lending pipeline is a modelling choice, not a measurement, and no methodology was published." — Dreamforce 2026 Shipped Agent Operations, Not the Headline Model (beri.net)
The buyer-side teardown's verdict cuts to the pattern: "Salesforce showed a coherent architecture and shipped the least glamorous third of it" — the agent-operations layer (Observability, Optimizer, Router) is live, while Koa, the marquee CRM reasoning model, is pilot-only with a "3x fewer errors" claim sourced from Salesforce's own benchmark agentic-ai-market-size-growth-2026. An independent review pegs Year 1 TCO for complex rollouts at $270,000–$540,000, and only 27% of commerce organizations have fully unified customer data — the precondition for agents acting correctly rather than confidently wrongly agentic-ai-market-size-growth-2026
.
What to watch: Koa's winter GA — whether it ships on time, and whether Salesforce publishes the benchmark methodology or submits to independent verification.
Orchestration is the new moat — and the new bottleneck
The deployments that work share one trait: they redesigned workflows around agents instead of bolting agents onto old ones. Talkdesk's survey of 252 CX/IT leaders found 98% of organizations have deployed AI somewhere in the customer journey, but only 15% combine agentic AI with cross-departmental orchestration, and just 5% can quantify AI's business impact (ERP News via enterprise-agent-adoption-production-gap-2026).
"As organisations shift from deploying AI tools to managing AI as part of their workforce, orchestration has become the defining capability separating experimentation from measurable business impact." — Talkdesk CEO Tiago Paiva, via enterprise-agent-adoption-production-gap-2026
AT&T is the lighthouse: 1,000+ agentic workflows in production, 45 billion tokens consumed daily, and an AI gateway routing requests to smaller models — "we're saving 90% of the cost" enterprise-agent-adoption-production-gap-2026. Contrast that with New York Life, which is deliberately avoiding embedded vendor AI from Salesforce and others, building its own stack on AWS plus LangChain — a notable vote of no confidence in platform bundles from a regulated buyer enterprise-agent-adoption-production-gap-2026
.
What to watch: Whether the build-your-own-stack approach among sophisticated regulated buyers spreads, pressuring platform vendors on the bundle economics flagged last cycle.
Governance hardens into "bounded autonomy"
The governance vacuum is now quantified — and the emerging operating model is agents that act only inside tight, auditable boundaries. In healthcare, 72% of surveyed managers report AI tools or agents deployed without formal IT approval "at least occasionally," and just under eight in ten cite security, identity or governance issues as a top scaling barrier (Medical Device Network via enterprise-ai-agent-security-incidents-governance-2026).
"No participating organisation reported deploying agentic AI with unrestricted autonomy." — OECD practitioner interviews, via enterprise-ai-agent-security-incidents-governance-2026
The OECD's interviews across 25 organizations in 11 countries found the same safeguards everywhere: least-privilege access, separate agent authentication, approved-agent registries, human approval for consequential actions like payments and data deletion — but "no settled framework for evaluating or assuring agentic AI" enterprise-ai-agent-security-incidents-governance-2026. This confirms last cycle's trajectory: governance is moving from policy documents to operational controls, because the population of agents is growing faster than the population of governed ones.
What to watch: Whether agent registries and identity standards consolidate into a de facto procurement requirement — the Imprivata control stack reads like a preview of one.
What surprised us
- Value metrics went backwards. The share of companies reporting widespread, sustained AI value fell from 32% to 23% in a single year — the opposite of what you'd expect if the technology were maturing on schedule enterprise-agent-adoption-production-gap-2026
.
- The Fulton Bank number everyone quoted is the weakest one. $389 million in "attributed" loans and deposits is a modelling choice, not a measurement; the 80,000 hours and four-day resolution gain are the figures you can actually use agentic-ai-market-size-growth-2026
.
- Zero unrestricted deployments, worldwide. The OECD couldn't find a single organization running agents with full autonomy — the "autonomous agent" market is, in practice, a recommend-then-approve market enterprise-ai-agent-security-incidents-governance-2026
.
- 45 Salesforce employees joined Anthropic since the start of 2026 — an odd talent flow for two companies announcing a headline partnership agentic-ai-market-size-growth-2026