Agentic AI Scaling Stuck Near 15% as Firms Rebuild Work and Roles

The pilot era is ending, and few companies are ready for the next one

Over the past several weeks, consultancies, hyperscalers and enterprise software vendors have all arrived at the same conclusion about AI agents. Better models will not get most companies to scale. Redesigning work around those models will. Deloitte research found that only 15% of US organizations have reached what it calls scaled, orchestrated, multi-agent adoption. Coverage of the findings points to organizational problems as the main bottleneck, not technical ones.15 McKinsey's own readiness research puts the share of organizations in the "reinvention" stage, where they rethink how work is done and how value is created, at just 11%.12

Salesforce released the second part of its agentic enterprise playbook on October 1. Its core argument is that stacking more agents on top of broken workflows will not scale a business. The workflows underneath have to be rebuilt.11 AWS now sells a formal methodology called Agentic Process Transformation. Its pitch is to reimagine judgment-heavy processes instead of digitizing broken ones, with agents designed in as "the primary workforce" rather than added on afterward.13

This piece argues that this consensus is correct but self-serving, and that the real constraints are governance and labor. The integration layer, increasingly built on the Model Context Protocol (MCP), is maturing faster than the organizations meant to use it.

Wide adoption, shallow deployment

The numbers vary by source, but they all show the same gap between trying agents and running them at scale. A March 2026 survey of 650 enterprise technology leaders found that 78% had at least one agent pilot running. Only 14% had scaled an agent to organization-wide use. The average pilot stalled after 4.7 months.1 McKinsey's State of AI 2026 survey, fielded among 1,719 respondents, found agents scaled in at least one function at 40% of large organizations and 22% of smaller ones.7 KPMG's Q2 2026 pulse survey found agents deployed at 53% of large US organizations, and the share orchestrating multiple agents doubled in one quarter, from 9% to 18%.5

Vendor-sponsored surveys look much rosier. In CrewAI's 2026 report, 81% of executives said agent adoption was fully scaled or actively expanding. Respondents said they had automated 31% of workflows on average.10 One analysis rightly warns that these measures are not comparable. Respondent surveys capture self-reported scaling, while analyst forecasts describe features shipped in applications, not features customers have switched on.7 The most reasonable reading is that most large enterprises now run agents somewhere, but only a minority run them as dependable, monitored services across the business.

The March survey is especially useful on causes. Integration complexity was cited by 63% of respondents, quality problems at volume by 58%, missing monitoring by 54%, and unclear ownership between IT and business units by 49%.1 Companies that scaled successfully did not spend more on AI overall. They spent a larger share on evaluation, monitoring and operations staff, and less on choosing models and tuning prompts.1 Organizations that tried to scale without a dedicated operations owner were six times more likely to have incidents that forced a rollback.1

What reinvention looks like in practice

The most specific evidence comes from companies describing their own mistakes. Fortune interviewed Adobe, Freeport-McMoRan, Kyndryl and Salesforce, and three lessons stood out. Deploy fewer, more targeted agents. Measure business outcomes rather than activity. Make business teams willing to redesign workflows from scratch.14 Salesforce SVP Andy White said the company scaled back from hundreds of agents to a handful it believed in, admitting it had "chased too much all at once."14 One of the company's squads rebuilt legal case routing. First-time correct routing rose from 12% to 34%, saving more than $350,000 a year.14

Those squads are central to how Salesforce works now. Its Rapid Innovation Squads are three-person units made up of a business leader, a product builder and an engineer, working alongside agents. The company has grown from a couple of squads to nearly fifty running in parallel.11 McKinsey describes a similar pattern at an unnamed client, where product development pods of eight to ten people became hybrid teams of four to six supported by agents.12 AWS reports that Amazon teams that redesigned workflows around agents finished a project scoped for 30 developers in 76 days with six engineers.18

The common thread is that team structure changes, not just speed. McKinsey asks a blunt test question: does the organization make decisions and move information with fewer layers and handoffs? It also notes that coordinator roles have grown roughly 1.5 to 2 times faster than line roles over five years.12 Read plainly, that is a warning that agents will absorb much of the coordination work that ballooned in middle layers.

MCP: the plumbing is moving faster than the org chart

If integration is the most common reason pilots stall, MCP is the industry's main answer to it. Anthropic open-sourced the protocol in November 2024. It has since been adopted by OpenAI, Google, Microsoft, IBM and Amazon, and Anthropic donated it to the Linux Foundation's Agentic AI Foundation in December 2025.22 Its TypeScript and Python SDKs hit 97 million monthly downloads in March 2026.23 BCG has framed the business case simply: without MCP, integration complexity grows quadratically as agents spread, and with it, complexity grows linearly.27

Adoption figures here are contested, though. Some guides repeat a claim that 78% of enterprise AI teams have MCP-backed agents in production.2426 One review traced that figure to no reliable source and replaced it with Stacklok's survey data, which shows 29% of software organizations in limited production and 12% in broad production, or 41% combined.21 Stacklok's own research finds security is the top barrier to adoption, cited by 64% of respondents.30 An enterprise guide reports that independent scans found exploitable risks on most public MCP servers. It also says only a small fraction use OAuth by default.25

The protocol has started closing those gaps. Its Enterprise-Managed Authorization extension was promoted to stable, so access to MCP servers can be governed through an organization's existing identity provider.29 The 2026 roadmap focuses on audit trails, OAuth 2.1, gateway behavior and agent-to-agent communication. Most enterprise features are planned as extensions rather than changes to the core spec.23 Shipping an MCP server is becoming a standard requirement in enterprise software deals, much like supporting single sign-on.29

The implication is that the technical connector problem is becoming a solved, standardized layer. That leaves the harder questions sitting with management: which agent may do what, on whose authority, and with what audit trail.

Governance and labor are the real constraints

On governance, the data is stark. In Deloitte's survey of 3,235 leaders across 24 countries, only 21% said their organizations had a mature governance model for agentic AI. Meanwhile, 74% expect to use agents at least moderately by 2027.8 Deloitte says the companies doing well start with lower-risk use cases and build cross-functional governance that brings together IT, legal, compliance and business units.8 Its audit practice adds a caution: agents can amplify broken processes. Letting agents start work faster does little if a manual approval step downstream is still the bottleneck.19

The workforce problem may be even harder. Salesforce's research found that 89% of organizations deploying agents are rethinking or reworking their workforce model to account for them.11 The company has committed to hiring 1,000 AI-native graduates and interns, and says developing talent internally costs one-third as much as hiring experienced workers.11 Deloitte names new roles such as agent supervisors and agent governance specialists. It also reports that employees' willingness to support organizational change fell from 74% in 2016 to 44% in 2024.19 Its scenario planning includes a "hollowed middle" future in which jobs become chains of agent tasks and the gains go to a few.16

The coverage gets one thing right. Vendors talk about "reinvention," but employees often hear "restructuring." McKinsey's point that trust is critical to whether employees are ready for AI-driven change is not a side issue.12 Deloitte's recommendation to leave slack in processes so people can handle unexpected agent failures runs against the efficiency logic behind most business cases.16

What the consensus leaves out

Nearly every voice arguing that businesses must reinvent themselves also sells the tools or services to do it, including Salesforce, AWS, McKinsey and Deloitte. That does not make them wrong. The independent survey data on stalled pilots supports their diagnosis.1 But the prescription is convenient. A messier truth sits in the details: success so far often comes from narrow, unglamorous workflows like invoice processing, help desks and case routing, not from sweeping transformation.142

The companies to watch over the next year are the ones running a small number of measured agents with clear owners, connected through governed MCP gateways, and reorganizing teams around outcomes. The rest will learn that buying agents was the easy part.