62% are testing AI agents and only 23% are scaling: the bottleneck is context and control, not the model

By El Ecosistema Startup · October 21, 2026.

62% of organizations are already experimenting with AI agents, but only 23% have managed to scale an agentic system in any part of their business. The figure comes from McKinsey, as reported in an Analytics Insight market report cited by El Ecosistema Startup.

An agent is not a copilot. A copilot assists users within their own task: it drafts, summarizes and searches. An agent chooses actions, adapts and carries them out within a workflow, with access to data, applications and tools. According to Dell Technologies, in an analysis published by eWeek, when an agent can change permissions, approve expenses or issue refunds, permission and audit controls must come before the action, not after.

The Analytics Insight report puts the global agentic AI market at $8.19 billion in 2025 and projects $12.13 billion in 2026. By 2035 it would reach $290.62 billion, a compound annual growth rate of 42.88% between 2026 and 2035. These are analyst projections, not established facts.

Governance lags behind adoption. According to an April 2026 Gartner report cited by Analytics Insight, the average Fortune 500 company could be running more than 150,000 agents by 2028, up from fewer than 15 in 2025. Only 13% of the organizations surveyed believed their governance was adequate.

Standardization is moving forward at the same time. According to the same report, Anthropic's Model Context Protocol (MCP), a standard for connecting models to tools and data, passed 10,000 active public servers in December 2025. More than 150 organizations had adopted A2A, the agent-to-agent communication protocol Google introduced in April 2025.

There is also a technical limit. Citing Stanford's AI Index 2025 and the RE-Bench evaluation, Analytics Insight notes that advanced systems do well on short tasks, but human experts regain the edge when the time horizon stretches to 32 hours. Many business processes look more like those 32 hours than like a single isolated task.

The article proposes CRM (customer relationship management) as the first testing ground. CRM already brings together historical data, business rules and measurable metrics. It is a bounded environment for testing levels of autonomy and auditing before moving on to finance or the supply chain.

**Our take.** The gap between 62% and 23% is usually explained as a matter of technological maturity. We think it is mainly an organizational problem. Giving an agent access to a company's systems does not mean it understands how that company works: its rules, its priorities and how it wants to treat customers. That has to be written down, and almost nobody has written it down.

The competitive consequence is awkward for anyone selling "smarter" agents. As models, APIs and frameworks become more accessible, value generally shifts to workflow design, integration and domain knowledge. That favors small teams with industry experience and integration service providers. The report itself names services as the fastest-growing segment in the move from pilot to production. The losers are those who compete on the model alone.

The figures should be read with caution. A ten-year projection of $290 billion is a scenario exercise. What is solid is the adoption gap and the warning about controls before action. Only 13% reporting adequate governance, against more than 150,000 agents expected per company, is a real short-term risk. It means permission errors, unauthorized use of customer data and runaway token costs (tokens being the units of text a model charges for). That is what needs governing, well before any distant superintelligence.

There is also an effect on jobs that the article barely touches on. When an agent takes over ticket triage, regression testing or review, administrative and back-office work, the most exposed kind, gets redefined. What gains value is the judgment of the people who write the policies and step in when the agent fails. The transition will be uneven, and hard for some roles.

In the long run, things look different. If processes with clear rules are automated reliably and with audit trails, people are freed up for work that calls for judgment, relationships and creativity. It is a small step toward that abundance, but it only comes if the boring work gets done first: documenting context, setting autonomy levels and measuring the impact on customers. Whoever does that first won't be selling smarter agents. They will be selling processes that work in production.