The Agent Fleet Has Arrived. Now the Real Work Starts.

Enterprises stopped announcing agent pilots this summer. They started announcing agent fleets.

Salesforce released its 2026 Agentic Enterprise Index this week, and the numbers are hard to dismiss. Organizations running Agentforce increased their activated agent count by nearly 3x over the past fiscal year. Average time from agent creation to first use: two days, down 53% from the prior period. The average agent that started 2025 handling two discrete skills now handles six. During peak retail season, that same agent handled nine.

These are not pilot metrics. These are production metrics.

Cisco announced it would roll a personal AI agent to roughly 90,000 employees by the end of July. Siemens deployed a multi-agent workflow to process 2,800 unqualified inbound leads per week, routing across seven siloed business units with full cross-division context, running 24 hours a day. Pandora’s AI concierge, Gemma, handles 60% of routine support requests during peak shopping seasons while driving a 10% increase in Net Promoter Score.

That last number matters. It is not a lab finding. It is a live production result at scale.

Salesforce introduced a metric called the Agentic Work Unit, or AWU, to track this. An AWU is a single discrete task completed by an AI agent. As of April 2026, AWU output is growing at a 15% compound monthly growth rate across the Agentforce platform. Retail agents represent 22% of total monthly output, with 18x AWU growth since February 2025. Travel is at 7x over the same period.

The pattern in the data is instructive. Consumer-facing industries are generating the highest agent volumes but running relatively narrow agents, often one to two actions per agent most of the year. Manufacturing, financial services, and healthcare/life sciences are running fewer agents but deploying them across the full cognitive complexity spectrum: from record retrieval to database writes to analytical synthesis. Regulated industries are not moving slower because of compliance. They are using compliance requirements to enforce agent discipline from the start.

This is the distinction that matters for enterprise strategy right now.

High-volume, narrow agents are a volume play. They automate routine interactions at scale and deliver measurable ROI in customer support and service delivery. But the strategic ceiling is low. The real competitive surface is in multi-step, cross-system agents that can execute complex workflows with context across business units.

The Salesforce data shows agents are beginning to move in that direction. The share of agents executing secondary functions outside their primary task domain was under 1.5% at the start of 2025. By April 2026, that figure crossed 5.9%. That is not explosive growth. But the direction is clear, and the slope is steepening.

For enterprise buyers, a few things are worth watching closely.

First, the speed gap is narrowing. The 53% reduction in agent creation time means competitors can close deployment gaps faster than before. What took months is taking days. First-mover advantage in agent deployment is eroding at roughly 53% per year.

Second, headless architecture is becoming non-negotiable. The Salesforce index notes that agents taking on cross-system functions require decoupling agent logic from front-end interfaces. Enterprises that have built agent strategies on UI-dependent automation tools are going to run into a wall as agent complexity increases. If your current vendor stack cannot support headless deployment, that is a near-term architectural decision, not a future one.

Third, the sophistication index data shows that industries with strong governance frameworks are producing more sophisticated agents, not simpler ones. Compliance is not a ceiling. Organizations using compliance requirements as an agent design discipline are building agents that can operate across more complexity tiers.

The agent fleet has arrived. It is running in production, at scale, with measurable output.

The question for 2026 planning cycles is not whether to deploy agents. It is whether your current vendor architecture, workflow design, and governance model can support the next order of agent complexity. Because the companies that figure that out first are not going to announce it. They are just going to start handling things their competitors are still processing manually.

---

Robin Green is an enterprise technology executive and the author of The Intelligence Loop. He is Chief Revenue Officer at Occams Advisory.