McKinsey: No More Than 10% Are Scaling AI Agents in Any Business Function. Salesforce Just Got a Double Downgrade That Shows Why
Over the past four weeks, we've written twice about companies betting on deployment, not the model: Ode (Anthropic) and Presence (OpenAI). Both are betting big — hundreds of millions of dollars and, at OpenAI, an agent already resolving 75% of calls on the company's own support line. But the enterprise customers named there are testing partners, not full-scale deployments, which leaves the essential question open: how often does an ordinary company get from pilot to production? McKinsey measured exactly that, at the market level: 88% of companies already use AI somewhere, 62% have started experimenting with agents — but in no single business function does scaling clear 10%. And the world's largest CRM vendor, Salesforce, just got proof that it isn't immune: two research firms cut their rating on Salesforce stock on the same day, citing exactly this gap.
- McKinsey (global survey, nearly 2,000 companies): 88% use AI somewhere, 62% are experimenting with agents — but in no single business function does real scaling clear 10%.
- Salesforce, the CRM market leader, was downgraded by two research firms on the same day (KeyBanc + Bernstein, July 9), both citing weak customer feedback on Agentforce. KeyBanc goes further and names the cause: customer data isn't ready for real AI work.
- The common thread between independent research and the market diagnosis: the average company still hasn't solved the operational foundation — clean data, governance, a production-tested product, a stable commercial model.
88%, 62%, 10%: Where Adoption Breaks Down
McKinsey publishes "The State of AI" every year — the largest company-level AI adoption survey out there, with nearly 2,000 respondents across 105 countries (the current edition, fielded June-July 2025, published in November 2025). The numbers aren't news from this week — they're the structural backdrop against which every product announcement plays out, including Salesforce below.
Three thresholds, three different realities. 88% of companies report regular AI use in at least one business function, up from 78% a year ago. 62% have reached at least the experimentation stage with AI agents — of which 23% report already scaling an agentic system somewhere in the company, and another 39% have started experimenting without reaching scale yet. At the top rung, the numbers collapse: "in any given function, no more than 10% of respondents say their organization is scaling AI agents," McKinsey writes.
Company size matters: among companies with more than $5 billion in revenue, nearly half have reached the scaling stage for AI generally; among companies under $100 million, only 29% have. Budget and data infrastructure explain the gap better than any company's stated ambition.
Salesforce: The Biggest Public Test of the Agentic Promise
On July 9, KeyBanc Capital Markets and Bernstein each cut their rating on Salesforce stock, separately and on the same day. The reason for the downgrade is specific: Agentforce, the company's flagship AI agent product. KeyBanc's note is blunt: "Agentforce, as a product, just isn't there." Their diagnosis explains why: "customers' data is not in order to do meaningful AI work." An AI agent is only as good as the data it can access — and at enterprise scale, that data is scattered across legacy systems, inconsistent formats, and processes that were never designed for automation. Bernstein acted separately and pointed to its own market checks: conversations with customers and feedback on Agentforce had not been strong, and in its reading the company's disclosures don't support the growth management describes.
KeyBanc estimates that roughly 23,000 of Salesforce's approximately 150,000 customers have adopted Agentforce — a figure repeated across business media, but not confirmed by Salesforce. Also from KeyBanc's survey: more IT executives plan to cut their Salesforce budget next year than to increase it. Bhupendra Chopra, CRO at Kanerika, adds the commercial reason: three pricing-model changes for Agentforce in roughly 18 months — enough to make any procurement committee nervous.
Salesforce pushed back on the wave of criticism. A spokesperson told The Register that "Agentforce is the fastest-growing product in Salesforce history, with customers like Engine, Falabella, and AAA going live in weeks, not months." The company's own reported numbers point the same way: for the quarter ending in April 2026, Salesforce put Agentforce annual recurring revenue at $1.2 billion, up 205% year over year. The KeyBanc estimate and the Salesforce disclosure don't directly contradict each other — they measure different things: how many customers use the product, versus how fast revenue is growing from those who do. The pattern matters more than any single product: at the world's largest CRM vendor, with practically unlimited resources, the move from pilot to production still proved harder than the marketing message.
Four Blockers, One Root Cause
Salesforce isn't an isolated case — it's the most visible public example of a pattern McKinsey measures across the entire market. Four different blockers explain why so few companies clear the 10% threshold.
Data and integration — the Agentforce diagnosis above, generalized: at enterprise scale, the data an agent needs is almost always scattered across systems that were never designed for automation.
Governance and security. McKinsey's second survey, focused on AI trust, shows the average "responsible AI" maturity score climbing from 2.0 to 2.3 out of 4 — real progress, but only about a third of companies reach level 3 or higher on strategy, governance, and agent-specific controls. Nearly two-thirds cite security and risk as the top barrier to scaling, ahead of regulatory uncertainty or the models' technical limitations.
Product maturity — if KeyBanc's verdict lands on the world's largest CRM vendor, it matters even more for any younger, less battle-tested AI agent vendor.
Cost and commercial model — a good product with an unstable commercial model still struggles to reach scale, because the client's budget can't be planned for the long term. Agentforce's three pricing changes in 18 months are the example.
The four are different problems — data, governance, product, cost — that converge on a single structural conclusion. Companies have had access to capable AI models for years now; what's missing is the discipline to connect them safely to a real company: clean data, clear governance rules, a product tested in production, a commercial model a procurement committee can budget for with confidence.
Five Questions Before You Buy an AI Agent
The gap above changes the question you should be asking any AI agent vendor. Five questions separate a vendor that can prove real deployment from one selling access to a good model:
- How clean and structured is my data for the exact task I want to hand the agent?
- What's this product's real adoption rate within its existing customer base — not just new-customer counts or growth percentages?
- What governance and human control exist for when the agent gets something wrong or hits an unexpected case?
- Can I see real production proof, from a customer in my industry, not a demo or an isolated pilot?
- How stable is the commercial model? How many pricing or policy changes has the vendor made in the past two years?
Good answers to these five questions come with customer names, production numbers, and explicit human control — exactly the pattern Ode and Presence show at large scale. Here's what an operational AI agent actually looks like in practice: a system built on the operational foundation still missing at most of the companies stuck in the experiment-or-pilot stage, not just a model wired to an API.
McKinsey shows the gap at the market level; Salesforce shows it with names and numbers, at the world's largest CRM vendor. The central question for your company: has the vendor in front of you already solved the four barriers above, or are they passing them on to you, hidden inside a convincing demo?
MassAI editorial policy: our analyses combine primary sources and independent publications. When we cite estimates or commercial reports, we explicitly state the source and context.
Sources: ↗ McKinsey — The State of AI · ↗ McKinsey — State of AI Trust 2026 · ↗ CIO.com · ↗ The Register · ↗ Investing.com (Bernstein) · ↗ Salesforce — Q1 FY2027 results
See what MassAI agents can do →