May 7, 2026: 79% of Enterprises Have AI Agents. 11% Run in Production. The Gap Is a Management Problem.

Seventy-nine percent of enterprises have adopted AI agents in some form. Only 11% have them running in production. That spread tells you something important: most organizations have started, and most have not figured out what comes after starting.

At a glance

• OutSystems surveyed enterprise leaders across industries and found 97% expect a material security or fraud incident from AI agents within 12 months. Six percent of security budgets are currently allocated to agent risk.

• Multi-agent workflow usage grew 327% in under five months across 20,000-plus organizations, including 60% of the Fortune 500 on the Databricks platform. The design pattern reaching production uses a supervisor agent directing specialized subagents, each with bounded scope and access.

• GPU lead times — the wait between ordering AI processing chips and actually receiving them — stretched to 36-52 weeks in 2026. Enterprises, cloud providers, and AI vendors are all adjusting strategy as a result.

• NVIDIA's message at GTC 2026 moved from training scale to inference efficiency. That shift aligns more closely with what enterprises actually need to run agents in production.

• AI-attributed workforce restructuring is running at roughly 16,000 jobs per month in net drag across major technology and services companies. The organizations managing this responsibly are redesigning roles around oversight and judgment.

The deployment gap is a governance problem, not a technology problem

The OutSystems 2026 State of AI Development report surveyed leaders across industries. Ninety-seven percent of business leaders expect a material security or fraud incident tied to AI agents within the next 12 months. Six percent of security budgets are currently allocated to agent risk.

OutSystems responded to their findings by releasing an "Agentic Systems Engineering" framework, defining agent ownership, access boundaries, escalation paths, and audit requirements as core engineering concerns rather than compliance afterthoughts. Whether vendor-published frameworks drive real organizational change is a fair question. The more concrete signal is that major enterprise software companies are now shipping governance frameworks alongside their agent tooling, and that wasn't the case two years ago.

The underlying pattern isn't unusual in enterprise technology. New capability spreads quickly. Governance infrastructure lags until something breaks publicly. What's different with AI agents is the consequence window. An unsupervised agent with access to production systems, customer data, or financial workflows can cause significant damage.

What production-grade deployment actually looks like

Databricks analyzed behavior across more than 20,000 organizations, including 60% of the Fortune 500 on its platform, and found multi-agent workflow usage grew 327% between June and October 2025. Growth continued into 2026. The design pattern reaching production is what Databricks calls the Supervisor Agent structure: one orchestrating agent directing specialized subagents, each with a defined scope and access boundary.

For professionals working outside AI architecture, the useful analogy is a department. A manager holds accountability for outcomes. Each team member has a bounded function. When something fails, there is a clear place to look. The structure is auditable in ways a single large agent handling everything is not.

The enterprises closing the gap between pilot and production share a consistent trait. They define agent ownership, data access boundaries, human checkpoint requirements, and failure visibility before deployment. These are governance decisions. They don't require more tooling. They require clarity about who is accountable for what.

Infrastructure pressure is forcing useful discipline

GPU lead times — the wait between ordering AI processing chips and actually receiving them — stretched to 36-52 weeks in 2026. This affects the full chain: cloud providers like AWS, Google Cloud, and Azure building out AI capacity; AI vendors like NVIDIA competing for advanced packaging and memory; and enterprises that depend on cloud-based compute or run their own hardware. Supply cannot match demand, and the organizations feeling it most are the ones that planned to scale through raw acquisition.

The shift underway is toward efficiency: heterogeneous hardware configurations, algorithmic improvements, and multi-cloud architectures rather than simply adding more. When compute is scarce and expensive, teams direct it toward agents with measurable business outcomes and cut experiments that haven't produced results. That discipline tends to get deprioritized when capacity feels abundant. The shortage is making it mandatory.

NVIDIA and Google Cloud are building toward what comes after the current crunch. At Google Cloud Next 2026, the two companies announced expanded AI Hypercomputer capabilities, with new A5X bare-metal instances running on NVIDIA Vera Rubin NVL72 rack-scale systems. NVIDIA provides the silicon; Google Cloud provides the deployment platform. Together, they're targeting multi-agent and multimodal workloads at production scale, as well as robotics and digital twin applications, which signals that agent deployments are beginning to extend beyond software into systems that interact with physical environments.

NVIDIA's message at GTC 2026 reinforced the same direction. The emphasis shifted from training scale to inference efficiency as a first-class design principle. Training is how an AI model gets built, a process run primarily by AI companies and large cloud providers. Inference is how it runs in production every time a person or system uses it, and that's what enterprises manage. NVIDIA moving inference efficiency to the center of its product narrative means the hardware roadmap is converging on what enterprise deployments actually require.

Power is the constraint that's harder to engineer around quickly. Most enterprises access AI compute through cloud providers, so the energy problem is one step removed, but it shows up in data center availability, cost, and capacity constraints in major markets. Meta, which builds and operates its own AI infrastructure at scale, has reportedly been exploring space-based solar to address its energy requirements. Space solar is a decade-scale research initiative. The fact that a company with Meta's resources is evaluating it is a concrete indicator of how far outside normal grid planning the AI infrastructure problem has extended. For enterprises, the near-term reality is constrained capacity and elevated costs. New agent deployments should be costed at production volume, not at proof-of-concept scale.

The workforce transition and what responsible looks like

AI-attributed workforce restructuring is running at roughly 16,000 jobs per month in net drag across major technology and services companies. Agents and automation are being cited explicitly in earnings calls and restructuring communications across software, financial services, and professional services.

The honest version of that trend: contact center roles, routine data processing, and first-tier support functions are being reduced. Agents handle high-volume, pattern-matching work consistently and at lower cost. Companies deploying agents at scale will need fewer people for those functions. The data across OutSystems, Databricks, and current earnings commentary all point in the same direction.

The disruption isn't limited to frontline roles. On May 4, Anthropic announced a $1.5 billion joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to launch a new AI-native enterprise services firm. The company will deploy Claude into core business operations for midsize companies, offering the kind of strategic transformation work that management consultants have historically delivered. OpenAI is reportedly pursuing a near-identical structure with TPG and Bain Capital. Business Insider, citing a source with direct knowledge of the deal, described the Anthropic venture as the McKinsey of AI.

That framing is worth sitting with. The workforce disruption from AI isn't only happening inside companies. It's reorganizing the service industries that advise them. If AI-native firms can deliver transformation outcomes at lower cost and greater speed, the consulting model faces the same pressure that contact centers do.

The distinction that matters for organizations and their people is whether roles are being eliminated or redesigned. Contact centers are the clearest current example. The organizations handling this well are shifting workers from routine inquiry handling to reviewing edge cases, correcting model errors, managing escalations, and maintaining the oversight layer that agents cannot supply on their own. The ratio of humans to customer interactions changes. The nature of the human role changes from pattern-matching to judgment.

What those redesigned roles look like in practice:

• Reviewing cases flagged for escalation, rather than handling every incoming interaction

• Identifying patterns in agent errors that reveal process or data quality problems

• Managing the configuration and rules that govern agent behavior

• Handling exceptions where customer history, context, and judgment change the right answer

That redesign path is not only the more responsible workforce approach. It's the more durable business strategy. The governance gap in the OutSystems data, with 11% of agents in production and 6% of security budgets allocated to agent risk, is exactly the gap filled by people who understand both the operational work and the AI systems handling it. That gap doesn't close by eliminating the people closest to the operations. It closes by repositioning them as the oversight layer that enterprise agents actually require.

The companies treating AI adoption as primarily a headcount reduction exercise are optimizing for near-term margin. The ones building human oversight into their agent architecture from the start are building something that survives the first material security incident. Based on the 97% figure, that incident is coming for most of them within 12 months. Whether the oversight structure is already in place when it does is the question worth asking now.

If you want to stay ahead at the intersection of AI, automation, and human performance — where technology meets psychology, processes, and real workplace behavior — subscribe to Agenticism. We cut through the hype to deliver practical insights for leaders focused on making people, processes, and technology work better together.